Mon Nov 23 16:03:13 EST 1992 From owner-mpi-ptop@CS.UTK.EDU Wed Nov 25 13:39:28 1992 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA21716; Wed, 25 Nov 92 13:39:28 -0500 Received: by CS.UTK.EDU (5.61++/2.8s-UTK) id AA09797; Wed, 25 Nov 92 13:14:37 -0500 Received: from relay2.UU.NET by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA09791; Wed, 25 Nov 92 13:14:32 -0500 Received: from uunet.uu.net (via LOCALHOST.UU.NET) by relay2.UU.NET with SMTP (5.61/UUNET-internet-primary) id AA22187; Wed, 25 Nov 92 13:14:33 -0500 Received: from kailand.UUCP by uunet.uu.net with UUCP/RMAIL (queueing-rmail) id 131337.3987; Wed, 25 Nov 1992 13:13:37 EST Received: from brisk.kai.com (brisk) by kailand.kai.com via SMTP (5.65d-92031301) id AA12688; Wed, 25 Nov 1992 12:06:04 -0600 Received: by brisk.kai.com (920330.SGI-92101201) id AA08958; Wed, 25 Nov 92 12:06:02 -0600 Date: Wed, 25 Nov 92 12:06:02 -0600 Message-Id: <9211251806.AA08958@brisk.kai.com> To: mpi-pt2pt@cs.utk.edu, mpi-collcomm@cs.utk.edu, mpi-formal@cs.utk.edu, mpi-ptop@cs.utk.edu Reply-To: William.Gropp's.message.of.Wed@kai.com, 25 Nov 92 09:28:43 CST <9211251528.AA12985@godzilla.mcs.anl.gov> Subject: Nonblocking functions and handlers. From: Steven Ericsson Zenith Sender: zenith@kai.com Organization: Kuck and Associates, Inc. 1906 Fox Drive, Champaign IL USA 61820-7334, voice 217-356-2288, fax 217-356-5199 Bill Gropp writes: (Warning: radical position that I'm not sure even I hold follows:) An interesting issue is whether we should defer all nonblocking communications to a thread-based execution model. I'm not so sure this is a radical position Bill since even nonsynchronized communication will need to be defined formally this way. Nonsynchronized communication is in effect creating a parallel process that has the job of passing the communication on. Al Geist earlier asked the question wheather buffers used by nonsynchronized communication should be accessible after the communication has started - the answer should be - no, unless by some explicit mechanism that formally amounts to a communication with the process mentioned above. Any nonexplicit interaction (e.g. a write to the buffer) would have to be specified as formally equivalent to an explicit interaction. Also, there is quite a range of terminology in use. One common error: "Asynchronous" and "synchronous" has quite a particular meaning in EE and when CS people use the terms in relation to message passing they usually mean NONSYNCHRONIZED and SYNCHRONIZED. Also BLOCKING = SYNCHRONIZED. Let us begin a glossary that defines the terms we use - if no-one else volunteers I'll take this to be the responsibility of the Formal Specification Subcommittee. So I'm looking for volunteers from that subcommittee. Steven From owner-mpi-ptop@CS.UTK.EDU Wed Dec 2 15:08:45 1992 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA12207; Wed, 2 Dec 92 15:08:45 -0500 Received: by CS.UTK.EDU (5.61++/2.8s-UTK) id AA24266; Wed, 2 Dec 92 14:56:52 -0500 Received: from super.super.org by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA24261; Wed, 2 Dec 92 14:56:46 -0500 Received: from b125 (b125.super.org) by super.super.org (4.1/SMI-4.1) id AA00585; Wed, 2 Dec 92 14:56:38 EST Received: by b125 (4.1/SMI-4.1) id AA02506; Wed, 2 Dec 92 14:56:37 EST Date: Wed, 2 Dec 92 14:56:37 EST From: lederman@b125.super.org (Steve Lederman) Message-Id: <9212021956.AA02506@b125> To: mpi-ptop@cs.utk.edu Subject: hello Most of the other subcommittees are having discussions and (hopefully :-) making progress. As a member of this group, I was wondering if there has been some discussion I missed? If not, I feel we have a number of potential issues to discuss before the next meeting in Dallas. Given the holidays coming up, I don't think that leaves much time. Does anyone have any topics or suggestion on how we should proceed? Steve Huss-Lederman From owner-mpi-ptop@CS.UTK.EDU Thu Dec 3 02:38:33 1992 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA25221; Thu, 3 Dec 92 02:38:33 -0500 Received: by CS.UTK.EDU (5.61++/2.8s-UTK) id AA05999; Thu, 3 Dec 92 02:10:42 -0500 Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA05995; Thu, 3 Dec 92 02:10:39 -0500 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA06896 (5.65c/IDA-1.4.4 for ); Thu, 3 Dec 1992 08:10:10 +0100 Received: by f1neuman.gmd.de id AA14748; Thu, 3 Dec 1992 08:09:55 +0100 Date: Thu, 3 Dec 1992 08:09:55 +0100 From: Rolf.Hempel@gmd.de Message-Id: <9212030709.AA14748@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: Getting started I agree with Steve Huss-Lederman that we have to get started immediately in order to get something done before our next meeting. I'm sorry that as the chairman of this group I didn't take the initiative earlier, but after my return from SUPERCOMPUTING 92 there were a lot of things to be done first. How should we get things going most effectively? At the moment I don't have any input from others. So, the only document I have is my own proposal which I distributed at the Minneapolis meeting. In that paper I tried to integrate the concept of virtual topologies (Tori and more general graphs) into the current MPI1 proposal. I suggest that we take this paper as a starting point of our discussions. >From the (hopefully starting) e-mail avalanche we can then distill the open questions and controversial points for discussion at the January meeting. I assume that there are some people in our group who were not present at the Minneapolis meeting, so they do not have the paper I mentioned. Therefore, with another mail I send you the text again as a PostScript file. I processed it first using "compress" and then "uuencode". Before printing, you thus have to do the reverse operations. If you have any problems, please let me know. Rolf Hempel From owner-mpi-ptop@CS.UTK.EDU Thu Dec 3 02:39:47 1992 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA25246; Thu, 3 Dec 92 02:39:47 -0500 Received: by CS.UTK.EDU (5.61++/2.8s-UTK) id AA06040; Thu, 3 Dec 92 02:14:55 -0500 Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA06036; Thu, 3 Dec 92 02:14:41 -0500 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA06999 (5.65c/IDA-1.4.4 for ); Thu, 3 Dec 1992 08:14:11 +0100 Received: by f1neuman.gmd.de id AA15778; Thu, 3 Dec 1992 08:13:54 +0100 Date: Thu, 3 Dec 1992 08:13:54 +0100 From: Rolf.Hempel@gmd.de Message-Id: <9212030713.AA15778@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: proposal Cc: gmap10@f1neuman.gmd.de begin 664 virt.ps.Z M'YV0)4)`F=(B")DW8LJTD.$"AH(2)8;(*1.&SALY.D"0L9,&SAP6(.R4D3,G MS1LW(&JXF)&C!@@40U*`B)$#AXT6.6"`D!*&3)HQ8=B`*)*G#(@I;\S0N1-F MXL,25-+085,F(QDY:43*<;$QS5,H8/&IYLS:O%DU*G3 M1@P9('#0D/&T2-LA;]JT*>.&SIRT9$&E M#!8B/^F`D`%#I\\QHLF4,0.B\^?0(!(J1OG""8@7JEFS>=-3X^K;0O:(64S& MMYD^ME],N6UY#!H%NWOGOHUESW(G?83!PYI1')A84 MF_39-\.9"O#91GAU!&5&&660L<<<=%14QQRGB?;"H6XDRL:BC8(`IE%P:(@F M$`W"D<983(Y0ZJGL88&F$4U4M(4"""IH8*T+)F>$$$(,:"M=721'AAE,SL%: M@(``:T51:F1&"Y&FP0::PP;HX6W&*H#%;)$09\('S;%'AY;":$5+%Z%#)P=5A+QA!H-+4' MS$T)T3(;JLF!W&U^->B6NDD4$2UAQ9%1!QQ,FO'&&RY*IG1A??`U\#Q!APK=GBT<8NM%G#3=#P-H1";.==" M&H>*M8?>9%04AD9OTT&P42]HO-@9AA?,XM_?@2!6V)&+/1,(/:$7X7%XOZ#W M'6F000<:?Z/10N"1$LXV56^-GI*+=8@AN61HZHU&&6F<@08=I9\NN.H@L,XX M&B"<-4?L"DQ>N^EX)$7L77H"_KO1JQ,VO,&PRSYY]LNWD(?SO:,^./706>\Z M7L=K+]E,,MS`_92`XQ'^]&\+[WJ!Z<]^]VV6X[<9S7*PFB&,9_1T%."C+$GNVMYS)Y\2TL98A0W\IP MJ#FL`003@=1%TC@E(C!I#&-8F0,;ASC%M>YP/^I/%S;7M@>";'MXM%KDP%;( M%YR,#BG;X/J68[GEE*M`CW.@`^U'O!@H('\.G&!;*MC!]95K6V`C)"-+.#;+ M%0B(4KI-$OZ6QU>.$01$R,YFQ+"V/4QJ)/EZ`2_A0,.EO7$-;GC#'=P@G+45 M\PUK(*1RDN`B`XI)/[E;C+G:\`:16`1-9>BE%ZG2E*-0,XX6F8B+T*!,.*`1 M!+^40S##24QV0A,$R%0F,^GYS&A*:%20:@KO%!#/8`94#G3H)SZ3N5O2!HHZF!#3`G!Y)%BT`I MK<$>4+H<_"V,8P[4R:`HV"Q&9@TEJ52:S@IU&Q+.@9UW"!YOT',;*2BLBN8R M$["2(P0BO$$T>O*CNN3#)#D(J0Q5NZA7J=*L;UE!`6BRPA[.,`=KMJ8[+0AC M7`O$'3A^M5EC!:N+8&2?R[D'0$9Q:D7V4-6KJA&-;5Q#\M`Y1S2%80_<].8; ML%4&-J3NA"^8UQX0UTQW0.N4XD#> M`K'W!4MBVWK1%""VF1=-O&/;6>Y[!^7"<;3?G)+\=+(4+PAF@,9!S;%-H\%C224<0$*TO;6G,8R)#!P6X M!C2HX==LW.31&?38HT`XJ*6,,+A!U0`OP^(!#'Q,YS:1B-9AC@^=+W M@?0,SJR`&YP9RE&&`1'4S&8WPSD)=B9TGAD-:2?'8`BN=O2K2P-I1$>:TFR6 M,I,?'6D8.%K/6RY-$?:L9V(7P=%#Z+.F)6T$8AN!U$8PM0Q\0$VA14QVD&;? MH)+LY#L;80C?OD&?Q\UG2UN:5H^&`6GZ#`-&%^$&0]"+J'U0@QS@4LE3#O2K M2>-J/%1#G;#J1P#*LN`R6_&P:&_ MG6ZSE'H&9O%!#NQ=9RE/N=>;[O67O'^/O2L#UUP01N=RIW&,Y45OG2E/QGF#I=YJ)TM[/M$>=\.Q[J3"\YO M&-0@V7G&`=@=KG!:'7G*;7XV#JB-;WC?O.OWV7,,9ETG?1^Z?7B&M1"B[/:H M^QWB8^:R)RV.<4TK7=^//GB;05X#MM.:X%:N4Z0#3NM`4_G03>G^CH'<]3WZ1Y=>]#E'O=`)/?F3PUP(MZZXE$5_Y2Z_O/8RKSFBH8[PG488$`'/.0"Y#7QPGQG0&=_$1KWI&QWSZEO_^C.G-9,OCF<:'-K[ MW3\T#LJ\Y2O3//JR3C]I"BUH0\O[3GBFB>Y+(X08&*'41;@/\[W^?)/C7/J/ M9W([AWT$6(`&.'V+YGPF!WY2IG&1UGP/:'3JMG(4B'NX9W.Y!W7S5WDOEW.S M]W\XEVBV9G(B*($M)WQ@-F@Z]VLV<&C!-@,V5RN1H.AIX:4%FIFP6]I!VF.5W3VIP".!H8C M*(%EIF@:>'[SMW.`F'YYB(<.D6A[&&8=!87"1AIZQHB*V&[L)H)!`(MW@Q(`3,!R4EEV\?N&L^5FNO.'UT1WI+ MEHO3IW"9"(RNAWJA5G3;&&C=&'4*%VI+!WJJF&?.V&HF=P-V.'L@IWC)%P/+ MAVK3*'TT8`.SV'WB-FOC%X:6AH3]:'IQAX!R9VP":7I;-G3Z%FATJ)`):728 M]V=.9W3$5G`[1R&&"(!&F'U2MF48V)$^^)$/IX6/)HS'1Q,@=P.8F'_$=G&S M6!H7-P3?5W`*!W[J&'X:5QJH&'@G&',>J9$;Z(&5!Y)!Z1!Y2&LJ-W?#^(Z- M-W=GD8QN!VGKEVZ,YF/!!XKN9F4)=V=+QG'V1VRT,G%+%VHV();$1@/LA@-1 M:7^(EY2,YWB^]H>"Z($[YX5$^&]0!I%D%GCFMX%^*)=Q^6H0N'YH*6B"J6QJ MZ'E2-FJP%FWU!P-SF&;(9H^DT7P'EVYW$F6"B9F"QFAH>(2>F8X;J8&BF8'^ M1AIV87#_-FM])G#F!H6!UGWUR&1?*&MOYHX@%P3,1QK])Y:/5FM.=H^E.6M[ MAHO^QH![YH`5Z8&Q5F4[QWHPEVB$5I=!UWI7-HJ5MI<-9QRH(F)Y(VIP`U:I?.!F4!=Z:@%I\@=W]M!H5K9P,ND8QF MUHE9>F;.1FA$0&@0MVJF].J,ZRJA6%GU?&J@B&:C(NIT31W,*<*6!RJ6= MZJ4)1W/0^JQ::*U..GO72JW;ZJ-(&JU%FJUAVF3<*J[A2J[:.JWIBJ[2VF0* MH*S#JFO+&J_L^F^_IJ/0&JW":I=0^FN]6J8R,(MP>JLSR':_NJ90FH0*RZ[2 M%Z@+.VGUNG3;292!RF2Y-I7HBK&=.J"Y9G-LIH.>=JDAZ[$JF8Y[MF>WMV<< MJ[+7R;)OZ:B61ZU*)[/@FJI`"*Y,A[)/-K-9RJ]4^JM75ZCX%[!N>:<3**]Z M.JE+UJ?RNI&26K$TAZK$VFO2IZ@3^:E)^*FXAH%;^Y\0NX/7=ZE*:WUBVZID M>[5J"K9;F;9LUH'!2K)-UJ,ZBJAS*[=32Z_VBK?)^I]7ZJMR=ZFO>F<8Z&F! M1FF;>'^V"H4YX`/">&\_"VH^&[E3]JYZ2W-[&[&6F[F5N[F82Z_EI[F=&[HX M:[G_:F6E*[K)FK*<.[JK>[FLB[J@ZQ"N.[NQJZ./^VF02Z4R\&8#J[@&"[23 M^[!ORZ[#JZ_">[S$B[S&BZZ?F[S.N[Q`^VQ%BG\AAVI0VFO7:[*;JK/:>X0X MRJG@^[WBN[WD>X24Z[WEV[WJB[=X:Z[NR["!UKS*B7[L:H@3J:[_IJY<^JMX MQK^/%@2T&J?VQ[@V\'QI2X)W";FQJI7[&K/IL&3!H.Z=J(MN)TB7,+9:<*+EK_W6:HI;*2DFL)/68/9J8R0)L(T+,,I M#,+-.,,Z7*0I[$D.ZL,0JYZL-L/YJV]&?&=(G,)3=GQ"#*:U&G]&#*1UR,14 M/,5N:,325Z*0JWJ@A@,*K)!?+`.CUKLS,&W66ZI1B[L0=V9F9G+GZ6F]=IXP MRL3M%JAT?*5Y;*24MF40>FLM[&AT'*V##*2#S*6'S'&#G+>+?+$O1VGSNF>(&FZA)G'O1K*.RF8B>\!Q>\S*[+'(;&;+ M[+;-[&G%_&X]VFP-W&R_#)-&JI:[K)6Y+(X+>Y!WEJ^U[*5&.IF(2[`HF<9M M^F2]BK]G6K<,>V?RRZ7XRG'W'+/CFL_1BJ3WS+-%"M!`:KOG&[*PZLQ/.[[H MN]#K6Z_LR[P,_=#@^M#E++F/%\`$2P-N.:4J!V5FVG"\;*3%^\T!;LW#;++-!LW+O,#,'-.F2]/)S,S3W&G$7,T$E](I!V_<[&V%VK_'9R?B M.,YG&M+8R\I36@0KRI.2]G`4R,K.?(1BZVD8.'PD.L%/1QHSV&F)VSZ_^\5C MVJOH^ZW/6[]VV[/TFG*JNY6]5K=Q?;5SS=9JW;W3?*XU[=%6#64QFFNE'*\I M&JT3;+M%Z&PX9[N)K;/`NJ?=*[*.3:\U:K%9*VERS6A4.J7>6*IN+,?JQKM# MN\Y*IM7]"V5WXL7BF;]AEI.7"80@#<:S!\8X8*F)?:P''+A2?,"U;-`A6:RT MG-@A.`3QBVC"K6O%3=RJ:KFR764UBGK4S'$R[(806JNW:GQL9]"@MI@>G=UN MEVOP)K$WU[6^)MY#\+#'QJXV(*[IC=[J_9_AF[X2_=X1[=#T#:YOW=`37=_R MO;Z!^\W97:2TY]0Y0`2[.]97E[:9?<`PR[[W:G,-[G\,[MX.'N$._>#45^$4 M/M$6CF;W_988_N$:GN$+#N*!NF4;[N%+1K->&MO:VJ4/!\C:6W0"2-9!:J8X M=[A#NW;LC+9@?-NPZJCV2X*F"YIVW6XO-Z3DR:4QRLTQ:J_..*EA/JQ4;F7O6N99>N5IWJ)EJ]=;_>4[&[8J MY[6X1^12)@,U0*O1)J=CS>/MO+9*IZ59:>MFGJZ60J"6]W1FEY MAW!X(H=F>6^\-HCA&>3NW@+H'=/H(A"VYJ6NZ>]VHW M(`0Y4']K)V?2_LMG&KA-YHT1673TVW"4!W/GV9W!QH-".8WIV-%82(*)3?"S MY[%-3K5._9RPB-`(6Z,&+7<@=Z/M+H>H9M";C>#MG+`'V(%AV_$BS[;_72>^ MYIYS6B!UULEYFI776H,3292#^YGN7-/F?-2:%Y&LF(J\!WKBB,-;^=$'OBO@ MK)YFX7&V"G)KEP/N4W(L7]6SK)4@K'!$MY,^^>]0/:*D.?-$3HLYI\-(?-`A M>^BT-HLT`-;ON/0DAV_Y.J3Y:H)P+X1!:*4$QW#9'O5XUJRDBI>0[GJ7AV?W M*W.'%OB$/X*!GWF9MW%Y3Y%:.:5N#&5[]V<]_6AY;O'NGH-R*ZNWO>\/R668 MAVAU2;_Y_M_]ZK,#1_-=YO5#1VAA+])N'*[!=A=(G_;P6:(&7=IW*7J'FG.' M3^Z^?W2__VH<>='$JOE&RA*J&O!-??-BS+92K&C)[:-W$OVV!G1\YZ=1[V@T M_JW%=W_1IO3[YWPE5X(:\S.FM2^H-#.?]`6?^7?O_R M;_\]*'$*G+VUC)3M:0CT'#ET'^H4OC%$/B^$L?Q:T,I$$7`OT0! M)Z`%]$L7D%940+\TE6`;PH%<6$W+-:(;4&K2C([C`7YK;/TD8)/U5F!'>E7" M9NSI&1CH;X#.OM,S`(C@_)IWTVN*0#4S4TO-7_4:']-V7<4Y!9[,MIAZW M!/V:'"L^B^GBGMUI-&7!K=<"-1#6PX)=D`N"P3'H`O=:_/)X M&N_15+Q2<_DR7GG+/:`F\B6S@S:K6!-`RC^R1B4%(N&W!R%3Y0%,Z0<0_D$_ M>',$82$DA("P#RK"]*,!*]*S>83N;._(04]3!VD-G:`!-8CV93R"LZ=^W'\# M<]H/\(&>\T609$XI/(6M*!6:0E6("E=A#?PVP^?)8)^&D>_P0]SI9Q5#WN"%`%92H'D-R/;\PM[4Q;L6.V!+& MXP$ZK90U&T\(`<&'H?["BD=G*1W#1^-J8I[$)$BIL'25_J` MHD2LB$'H)\J]HVAT(A$Y?(HRQQ!JMI"$!X=SCEMI:(L# M+S-B)3MG]:2:3[H!9#'8K#OWAW)ZE`XJ/M1,%_8WA%8006``NCS6">@0)=-H M2FIJ3KC(!1.`NS"$3]JHU7\CY1^X@W0N#B^("_ MN`L?D'O$4M/.!!H[,,82AJ`;DW$R[LS41W`#:;[-LCE'[W':'23]-/O>#$H: M%.-/^PT]'@7,K)F:BEX0$FB-K)&U:6P<1[-+Q*WCM`^>B)+4C+2+.5_.-GI& M@U=]/I7`\UH4J/XM'!+UO>:<#EJ181!&%KS0)",3GO61C1_H1K8H.& M"5U5I1`.N'F(H/%+'KDE`VDXU@+T?'/N[V6]$L3SLEZ=8T66ZIWE2`IT&$$2 M!JJ31GJ3+:=- M0C@PE2=)U#K*0$J/T22VFJ!VQA3(R3\"BC0,@2M7=,K?9YL3:*D2O9ICU'_D ME4?J*%A*!'4^*2<;+P]GU$FX4MB`/!%%FGR6S2F%J8?X-*2@E)Z&Y=?K-%:I-8^RNT'WP:N6FS]2@">1H`!\)U&"7 MB"6TR@5UM.@3@(*5_.D\03GXN&NN3+^[<(#HVTF@9J67[-C#U$"([L*AN(Q) MH#:FQN28;RG8^"=+QI]@):P\F.KFW>0``%8G-EKUF7/^*C(4 MYS.CFBD+-B[R+O1#&,`R<1"J$78P*6P6'7*#:2;0N@%0[B:*]9E+9):>S]@C M><1&(44Q`.:8,E[^\C8RIP9XG74G;RZ1+%M`-@I%S;D!67TVW:93ZHQDDCK>;4 MC)P:E M(IBW<+PF(`<6=Q#1$N*]_ MT1YV-SXWVI>Z2@J*XFB9[3."]D)3%$Q\!@+QF:[#9ZJ2T3F:`8_*72IQ6<^9(H21DK. MP]6:.P0.R]'GR3H#:?K0BAR*>GKHZ7F&".GK-:2G9.].U:V9`4!')OW(^?-R M`!XH\CT2B._HGO.4=_J3K(Q!;VK=#;B*0V]V58@RH2_*^B!.'/4B$1"6W$\F MD@:82`6@1E_FI[(Z),KOF-!/%85BSFH"D7+4W)A&T*-E.DJ/8D1?"K?U)QGI MQZC!Z(! M-'1=GB>]US&C$8L30;YI<9I&XAF@3J./S(U':"WJQLT82TLDV;JEMC27-B`* MY'0Z4$2B-46IHX@@1A3F_-/"_%H'ZK3%J`EJ!&;1#0@"E4@!S"(?H*NDG>,S M5,#2>)ZR1+/%[E!NR^%W09-0#Y21YG3^@#3.@X#6*1_UH;D1];P@ M:90A,=*?N5?KTO.)TQ/YG892$CVDEJLN#:#KB1%1U2.D;E`(C>W"T]5-3YD, MI3GL!@"Q&U(:0GUD)/HU%W60NIY+LR*+$$?%2F/N]X"ZUR=Q3*=(M9$'"O6D MU/@T(QG4KY%/+S6?9E0MPVYDXR%U1HF4EEZ9+%90>R&M2EP#3!Y=4UGEE;@I MHW.H#XCF6*E?LU1WJ8]LJ@KJ-RK2C[IPMHQ'!5!/YV0ZG#)'U4YJ`%JI!7/Z M@-5D>BE)9)`*D5\.1];(M7I!<6=9FV`[]*=6-[_Y13F.?]HZ,M31F,NH9@,W MXP`B.BBG*/D87=2E:J,9G4`Z:/V@U4V5_M1J6ZV1)567YJ[,*+=LXF[TC+6' MH%Y6!)1H=BC#I%\I<_Y8QB]D,$W=XW&FT+29>LV%>M!P(Q;:E66TC(K(?`I+ MS>HL9:N.->GHTC&%2Y>HC621*(J,*M5UV5&D)\'$C["4D<*XWQC'(LV]"I28 MU>YM/SJ!N!1`=2,"!,R:'C-")>I:J@J;/5>FXARE\(HCV\>=HXX&ZLJTT?3: MG^CHE8%M"^C*D,7X2FM>9&5"3QB(C5J<%XF*O$^,$D_^->:()XWS7\6/B(1` M]P%'!J;)%37O!+]IL&9(Y12F/"J8+*L)90E\],+FA#QZ]J`1!E(V.11PTAVD M.>U.H[Z1?&5M7[ZJEPK`TIE3FQ.8[ZA1PPQ[HF*F>*VHM`;HQ$R\$U5=DHQ3 M.=+TQQHH(=N?>J@=-;(+R.C9T:PC8!V.NN2@_`EKGIPG.V4%$'&SL5?6RDXC M+&MZ=NSCD;!?UD<5S`G&GR:8?_I4=W7)MAQ\.B?#ZD'C"T-6KB;4&-M:I0^U M:EI$I:[(PX[:W]IPZ-!VHJ=+9AT!(%^[&6VC,S0WQ[:# M4J0X!YCX%9I1ES2LDB)4:1H#?`!:TJ2YQC\!)OLDY?*$%8NAW/.8GJ.)""NY MZ9&+0&?RE59:$2CKCE(S@WJR#D4!0<9*;X-4^DLT_`G?3L1\:\<'Y5-1*%O@G/3":*)7&/EB8Z6J'&-(&>LQEU M.&[&-3?T*-**7$@[3POG8CVK]?:V8EK'VB,;CJ7JK2/WL$+:F9LEWZII)%4( MAV!56P4`$/WCC^NS4I663E5@RD?+[)T;D487C$J?+1-1(:I#;;H:U39"50`E M=?/I316Z6"GV8=VBBW2[+A\]LQKUR\0UHBOE#E2R+*V^)M-!*N/#;T#CG&5G M<:N(IK))V:-`*-%%3&2WX21=@]EPSNR$@I%Q33]MQ@-U<3=CG[D]W2E2M MVGN.D!@ZJS-S[!BW.5=P*.\5`GDSX`H%*1PU6]4MW-N<6M7TM%V"I5"?C"?Q M=D,0R_35/^F0%)+O]'L:-/8^(=##>6JO-,TZ8.?J@)W>V)F^CJ&Q-%6([&@? MCR,`3Y+(:7I*ANJ)M/?)93>.19<\JS70IZN0\H2MLW#*,"$?J M7X0W(AGMI_I4`I3+R*>*:GC,'/%A2DRR+?W-"SJD*I`%O4!AT"/%/Z%D!^]$ MZOTYK8?U$AW7ZRR9CNSUP+/WD\U=J;.(<.\=?36=J>#T7K#C8\0."*4X9X=) M/IO;2<,>KSI*-W*'[M"ANZ-P]`[?29Y\%/#DI9'D8[M/`7Y*B22"O\I/X=($8UB%Z;"7D8T/N`(3(8EIQ;.F4K/ MR@@!%64$J)P-(&7_QQZY2X#IE@I.OS.OA5`FVM!4.&)]HC3JPT%T`[?>02R( M8:/@\49_L#K)@%_;D4">$V6?RBFU25&(I:RH%GMR3^]W%]J<"5I6]5\$#8H_ M$U=:((S4@FH.>'HR!^GA'493JH&A:RMVEI\/%@.^!%N"J8P-=@A:!\VL3313 MN%*?`:5N,]@MY<%3)BZ-Y..$.G;"QR*HF(2@J$Q.BI^@E7ZR'_@3=OP,F40T M3$8#2IGZ$S[)(EWMF2(R',M.8VR`=B5MK4R._D89A4F=1F+.G`X7TM1U<9ZH#D7$N0!I!DY,J&41!Y MO6`9;H\EL1Q%>`GS"%.\M(0*$5X"E0Y)TN7?+O.&I5SO=$D(N/CB.:6LPF2` M4^8SX@84'1Q)YQ"RRJPP,9O2Q=R0:!$#@C]53SR=HK,+I[/\QNIY@PI/`GJ]EU%NP=`0DOBTXXZ MS6:^G8!I)=FC6A.555/O,3C^J#*A'A5D9PM2<>ZA0'0=+22!Y_4Z7RPNA_/( MYW"?GQQP<24\=L1;N*P1'Y+D#[GHUTPSNKDM#^;O@W!F$@8&S)*&",'C+GB& MZ=^EJ\>AN.503`&4E]\1#F)*'Y+C^!B->Q\4TLMZ2YK2K6;$576'S@]YTC.A MYDOZ&F(C;]9O$2.0.]$^YZ`'%$KM9T\>,PKL!O7#)-Q4SD'_*BF9@2<S7F%(6&U1,\PT M`VZ)S2L9EY^(H(_`S-FDT>GY+J&A/:FFM MG'J(I9?..<043^%BU!2C@U(/!,C$Q^PUZ4PL<_Q@,IZ!G8A&@S`9#9Y#42HN M3LN)T=SHQ#N=Z)-S8C=_.N'*:=+/6),M[2Y%7% M$U6,B.\[^L9?R+DF&V73;F:;W)DR*A54LU0G@YK6\XAJSYYX"HF??^."LN^^ MU#-YHJ/HA8)["6W$K3;$IW0;&RE0*-%N;%%^KK7S-K_>UO[[7_7H6 M'F2`_:\WL<$>V`=;7U\J@:VPLU:;PF[/34U5G/:Q[H*`9(HS:\^=!2%65A35 MD_.9=W:OWL6M_G6U-EN6,ME;:4UY*95=I``5L1*U*9M&;IJL%NJT6JQ*DR:R M;(6ZMX?E-F_I9+G!U>1JH+VTH>LG-_6-0ZE7"=M>U?S^&W9[@'+2?_FO8D2Q MB8#%)EKP=X+1N]OS"?,7R^-BVTB29BFZJ.4\&XOR;/;,2!VUP?GZ0E*)O),[ MFZLU,+NT3-OVN(IHHM%$>C8$YB=KMMB:V2<2<-:O^_7C1F"#"U(-,T?LJ-1=;1P_ M'FNBN>SL$W_!%>8)N`ZKU/8E?"6N@AR8`9:VBW\!K1XG)_6U,'O;P!O4:2KA M';SGH.NC<=A-*Q(X$E@$KG;I=8NXLV_%NJ+HLY9:VXYU_.K7>!ZQA:-ZI??N MWN#;^M!"FXN]^E=+%'PTYPZ92)[]Z=S%>?E]/F]/56 MZY'F#/N\JG&\P:_/!C?'4\9F;RH1?BXK>*'JWCC0TZ%L>SAK[#9KYFI'!X8S MG51W>6CXR:X^5\O1_B&QM<7L]D!]HPX7@?7ND4CC2FU<`U9\FS.2\"ST*V<< MJ8VWX_*C)60,R:^*WX,\8-<4'O;"-B4#FBG3:]X:1Z/1*0,&M"*>QXM>)?M] M;R7W=\-QHP@O438[LI+P&I[8VGB7T6JVB$";93S.@?FX0_HVQO6&A^,'&(*H M]]I:M(=\1FDU1:Z?0Y"->Z5^/.API8<7R?U48Q9C/8B$#YWW1XU(>-'QY'J\ MV,ZH@V@6U[B\"N4<+7I11PCIL\Q:^Y@3"QPM8==SYO+*?R=Q'#E\WEY_4XH& MG(AO[+TWS;=Y6]5J0@B:GVQBG<3#.9PCX=& ML'X?`UU$%%:8)ZRC=;/.MU951%+J97N`CR`D5=EN%EV7ZU$4#SJP<[OH'*%$ M;#9^?2(6NT$GV-LP81]$KT\=A75.A]>YZ=-:[$\/KBNI(4<1M5#Y:6-LJO5! M[7@F#3=[.KVF"FD2Y?,@4(L8EQ)6;36MK^V MV')C3Z_FM7/3%MH;Z=SX7[#B;LI;ZTNT2X^NC5%>>NW5F*:SKVY M,_?EKMQ!=W5WE=<]N2/WH*?5F_M8_VOX+-2Q/W1NML5[X[OM68Q[Z78JGM`; M&%H;4F[/MHW4A%[';UZZB>,XX![ MMR-NWR["Q_LOU\\/WI)C=Q$.N(?/D9ML$7ZW5_@,3^$AO/O#Y`R^EV^C=Y[@ M^]M!,_'*:IVK>'7.XA^G/-_M21R>KW@9W^)7NSU$YNFT4 MY;5\AZ?RI#''UZ<-9[?MO`';[$_WL>EW;=1\ M=][.Q_@RC^?W_$S7?%"]G4\I/W\]Y_R/8_).I\&S0[.6LJ7O<,QB@U&5P<#] M'.0!O+7%"X`10H[UZ+7I<[EG:U8>/EJ9>6D>YTF]S2Z59!X1MW%M1.8IUY6' M.:_>W\1Z9&C':?UNI]E#$M0/R1>/XV-\C=?Q-7Y7\GH8_^9_?;$']A-\V"M[ M=M[KQ3RR+_;"GIT_[<%UNO;:K$N/J#L(W``?8VW/'J\J9X*4T1]RW/?Z!&G, M]JZYW?ZYL88XWYDB>S^3Q+W.Q?M--U(A^!_&PQ^LE%X MGO+X+)$`347LL[W',54L0-N=WAN@N=VB\FO+-T`/'^3`YZ M']Q]YI7\&;$28]%[^<+YH=^`X7P1-^"$JT(R'#_EUK47O<)@5*3<(]O+>_>^!U-A2WU?I#;O@5_^YX; M[G?NAZ/#W7[0]-K^]+54%_%;4*/V7?U_OJL>/KO?W8 MI=-C7_J^'?&K;=(%^4&E6--G4;8O'@YP:%8B<[;O[_W`O]V7-95O MU7:E<7=1=&>;:;\*':R:E@[JV@X=@(W]-K/Z@^H7[>!J_MZF_*$_MMF<[&S` M%_[[PW#N3_Z+?O>'Y^)?EK;L!AP'8SN@DVR"OW'+[P!^_EW4OG)KT7Z]UKSE[[D^`F!7I_\E@"X*\G?\"'Y925PSW+16\XE. MUUJ!&MC?&M?!S2(R@+;7O,D)X%IR)^NX?\A*_'*BG'>VG[6CMHUUWM.^TMW, M*XD=R]:I''E)$L<3_`U_5@WQ%ZN@?[I'OH+8!"7RBQ>"U+Q560A(4W],?RL3 M:5"TW'_=$#;SMRQU/\WO01TQ*E1@090E-3!88.:S`QY^71TNY_5]0HK=%HC7 M98%B8%=G!I*!7"`:Z*YH@6D@&S@&NH&W#T(3Z,`J4(D'"$W!9-B0Q@;'G3O> M#*GR;G!SX<8+DV]\(47/52>L'((XCR%XU.1M`QPBR-;<,R);2=/+<7-1C\Q" M`=Y;ZM^ZM1!IFR(/8#%G#`6I[5AL(Z'CL5$@, MRK+A^2E_RP^8(I%C(]6:8A#F:VR*LW&^)(0(863#$*HM#V%$Z!!.A%5ZS=>A?2I2TKECMCAQ""JI?G-Y4H,5U@<6-_F$M]BQ,S MZ(E`N)TZV!0659/?W4;[_7S_'U7(%%J%4Z%4&!6*,^I;Y?>M9(5Z#5CX%4:% M82%9.!8Z?EBA6?C->(5K8;XV%K*%76$55X?X-;A-FZ+2U8426W:#V\P`><[8 M9ZL$`3-`9P8!73,^EX@':Z!N8M:W@?T\'$[,+],8"GY1G]O&;JF#Y2`[&&Q9 MAN<@]O(.>H.;X62X&?)RFB%G*!IZAJ,A`5@/8COS()+$V?0TL,HL4HQ0-Y-( M"'CN,3]-B[&"Q*2`>4M6Z*R@+"Z@%>C`](;L76#$VR5VA='(1AS61<;AQ`$$ M'8?`(7"H&T:%T(H6(@O^AB`-]%:'?]]Q8;ZO;R2?OH%#[ MFF57'AI^.F'F=_G]*9'?X[<>5GYF!EK8'E)^[^%ZR/FQA^DAYQ=]1'E-8>A' M#2(X!EO(TM$4*LL4,6@@)H._X(&(#"Z(Q1L1V/LQ;`[B?G(7,&])(+BV$])M MA(M?(P\9`<<'H\(A+G5=D6#T>ZALVAC!9B(F;/W:B8BP*2WNX(I84+V(DDL- MHQVN2OC9^(/-O(A>S,1TK74;60:KLGR-\ M0<1-KZ6@"1K'#QCC&Q%'`4>YDW;A=G;!C,C/X1M<"-%D0?U?OE6*@JC0+=7A MU_?U"3UO8!FH!IZ!;B";J";*+8$>HY+6Q7!TXIQH)R9$1$=`)*V<)XL6EZ)F M^&6UTZ]3C/PN?@K_5RC:/8'"*'QLWQ3G&HY]T<8DPA&B=,_-9J/%=8MHRL/V!>9O]4H2 M@EY1,6Y>S^8(D6WXT22E+$%]RA)=A\^0/3X0-1)>`7;UHAP%V-TH&9.N#+ MU$5+V49=W&N8+9XQ8\W&@Q&)C-G;]C9Q?","H(*%S\PY]5OO5T85.R%2RN<` MA3PTXXV$P1EW[DRXA-^)'@L;<@MH_,1%]C-G9214O8$Y$# M+;$<9*/F83;V-L@;VBB&24MK(]PH-Y:-2C835V;ARI(U\(]KH-X(9 MIR&M44DA'\^&,3)WM$J(RIXAB-!0QE/CZ"AU1I`CK05J%#0G$6QF+5F.Y4CF M6#E>CIRCYN@0H$28X^:HBB1]@!V%8&ZH2K',`22IT'JYF:W1L'Q!GB/HV#F* MCC"'QS,[MDFRHX:6.]*.N&/HV#O>CKMC*H+U8#58S<1V.CX;J:-)0_"T4=J6 MY=&+_4Z6%9.1:(T9Z%6M*.]%2FDCW1@W;H\&7F+W-I899&.T!#X.8>&C]B@^ MRHT9W_E(-XZ/?$CY:#?:C>QC_'A[F(_K8SW(&IY*QZ/K9S=Y&[(-_\@_FF0` MI.KE!^?2)]Q M-L%9&J2YH2\E4#B)K#9Q=9`.QS"C*HDV".3&!VN0&V33+I)N<%QIDP#$PF!7 M+,&'Q)_X)^A6P92M$2@`4N4$FGD9ELS#]`4]7QQ:T49$&I%!9.#Q0T8US@BM M`&ZA..=)R92.A5V,2'KT;EAM$P\I;@5(2Z0EPU2173>D%[EE6)$+7!8YH<%'8^0,I6FH'6W8 MR(&R`5H8TGX$2,I'`A(9&2#5&IG&3F0[%2V(4R0I1DV2853I)$E6DDD*)ODY M99*J1_NCE!ANC!M, M5[E1-JU?[$8,*6]@5VF,$E0:F$=RS"M&?%&G\%+ABB>A(NS M:R21'1/'U$=:*<[DOV7*@$R7TC3I0[HXT&0;*4,MDM3D-GDUR9$?4X!".V&1 MWMEHTZ/H=`DNIT3]A;CR3M>:JA3OM4T7YH5F47AU%>5%^ M:&Y'QT-[B$_I)#6("$$=^9?(U%%93R*;U%-$<2,"7/X%W,Q2.!+6XH,P36C1 M*@DU>4[6$8%BF(10L-*JP@U)4AY-_'&1.90A):(R4LH<@)3UU$%M4-!0332F M\%^S$7O&A5&#ZM:34DO14B*;2ZG"=)2@(E>IC`A/-16?]G%%16%3TU+V8#0/ M98AB-&4]391WM8@9*#8'#6!S@&0X";62>[QC1Y?3-5SME7T42>E7RAR`&H$B MK[5MIIH5)@;Q/W+'FY$:I96U!I.Q_(1\LXS\9E=J63<3V45V:5WT&4ASFC1K M1%<`1LTU9.%8<*-/34_KR#<2^#A%](G%\;4D(HW&%<50HI4AI?G#:S!2!XGN MD6A(&05U56&5;M7]Y3M4E=9-68CDO1Z`A:WPMLHG219BYDU@*0OE< M0D/(QWF`_7[G(_D'$TT/=E$R"4]Z6(&E"+;TL&!Q3`P@!`` MB]0=H,_UQ9`Q1;92S@/%U%`$"I&6;F16:1I:22N@D'>3;&/PJ3\II))I3%IJ M/L8RV63JB00(",?<))D*!T>U+V4:9A-]DN<$0,G'O-%SC3QKWC%&$)Z9*)_# M->WE'+"&K>(0J$=V@285>CEP_]Z_%\'=F5_.?&5RA$8QAYX)3+U1>10<95Z6 M5VW5@P4XB4B&!LYQ>(E(V,@N]?PA,[`22#,B19I'%XRD7(%1OU&'M2[]6D^, M0`3A`7;-E)EQ24T;F=2WU\P`0WY5K%6BE!]VUV\DTC`1M9/[&F7S5Q M-)K0T/G&VQ1;SY49]6JYFH06H:5JF3^"EJI5?@R;?4>K!6S"2+?F$Z-2]E2S M1XLUM-1-(4JI0J9A&8#&4))Z)1I?AY,='M=4/E7 MKQ%U:E2R45A)=8);LM&PZ8RH;>3>1Q-_<&/4RS")TJDR:IO.R7)N75)5SUGM M2%D7E-IY4C*=<.?;63PAG0//F%.38%4;5=.9=\:7EE;*Q7<^5JRFWSG918B6 MT,PI>A0M-8IR52;-'W)E'O)6$E;M5/:A@[!DEY.(-)7<./>FH7ETO9LW'N>I M%LF;GV?H6>X$4LV3C=0*PFP'"GZX6\AE\EC/#9_(Y"KF:OV?T MZ7M.G]`G]=E\&I^HF/3Y70&5O%L-4L8,+=RAUX=L!ILPDK)9?)J?Y.=OA'Z6 MG^GG^JE^KIUYRW\CM.PY^H<\TK65/Z&EI)5_3EHB$D5%6^V?_Z=HR7\*H`&H MI$51*58[2WY$8"6@R5.B@89$&IW)[J%P,I242/XAVI`^CUPN?-J*]!,U7@-/T_0U8B6%3?*Q*8$L`(S"(T>%(^E(75(K+9RY M9JJI@UJ?[:>P.7CF&D*/-N,?"2UQ"BS)+7I/\>5\F83"EQC4"K&&_I2]5T0%1W:7O4G[U4P56ZX MGCPE.[5O^ATM6M11HMR&:A^I)WO7[*I`$D@-4#TP8;:H@B* M&X*'"EZ`&2\:8,%*O^C+H5@23"?8('5U#%(H:+KYJN&;\E M2J.IY89%GV1>49?F19_46*X5V(5CJ5)I66@;T M1Q,8GF)G.KH!,\QZJ6L85I5 M,V)OE3EWPI4ATP%Q3&94*]4EQA:N.9%:(\TF+/C\ M0%*%J:H9:\DX@><-.H,B=[]99N2Q6%@@F`85>%U0&Q3B)5LR';TF!\.9*J.= MR@.:):TFT@COYHV$7B",Z.&2%:'5ECR2:SA7-('@E4&9F`#FM..L[5)\I8(" M)KHMC4Y-9\Y\?LJ*:538!"$$J(UFIC5\K\EQRF9B*:%4S!-U4!IO4+^TRP0L M&2E-<-V4+N%(Z@FF\";LQL?5G9I00%@+!GK3J>`BBZ:/]U*<)YIRT7>YIA47#RZ7MJG\:GNU5[^O9,///8.=J:HG1"J/D6"?Y3.V>!&G09 MJ(F49HI(=:`,*MF54R&H#>J"ZJ#F/1)JA3IT6:AHYX,J@>1;5>A04C)=2C$3 M@W:DS#W0%JHB;16AX6?8-M3@3CU*&V*JX!I;I(PT4U%=<^1318S]5B93>KIQ M>2UC99S#L92@KD?,5'00IWCGD(J%:%6ZSU8S6($Z)2JTA6=N*B$2)E@XI3\G MD_WADNTYN2`(E/LPC0S'+`I\6FMHY\2Q7Y:?P\VE:?Z8J7U'OJ57E26PU'SE MG\@X;NI6*F@2'Y2+H)F/VJDMCJ%)\#Q86I5FDHP^'8DFX_5^Y"+J$L*1:!:J M]@D,$GD!(]AEY=5EG56,ZE;C5FTD)Z"+!'&X2`_I9R2^Y4HHSJ;#178@,H`- MAJ6*G0(BK"@0)IO)U6^E;25HO!7>B7"@&^E(,VKF"7^NBD^G=V M7I#*EE&+5A\>3&PSP2188%;Y%M#%G(LU!V-3405[1 MVX4"^4#G38>IK;XEVVJWZA!XJ^`JMRJNAJODZKBZ@T@\M)-_=Q]@5W1"_V&$ MX203UP@Y@720B5>]6F[`'E^'!6E!PALPB#)50FI1P9CU@JT:.59)P6J>&*P( MZ\&JL":L#.O"ZK`VK%:)'[,-`2JT$\VWKF(B51Y?=8J='-N3&&2%/2V#1DZ\UJ6')" M311,`D)!4:Y0*10[KDCI5-%*M!ZMKH=7,K26K,2&O81LP"0-2"GS)XJ:R.,< MP@ZE7>#9:5/WG#Q91F3U=XQ(0^D>(G0('8RHXHCE-)+NR;9($2E]3!%\YEZ^ MK6ZK%[*UMAO%B/T%F#J7\Z2:I&+NK=63<>FWQI-HDD-RMWJD-E#=RJVE'%-K MJ5&UQF3KI`34^W1@96.*V;=N'$B2;Y5'@6$22/SQ=R$>5<9Y`L0\(`B81Z2V MFC\%5W\$)6(9#A0?HKJ6&1D:];BZ,E\^V>LZA,FN8\_Z13@"&TS81$)PO&@% M!XTF[*!#P1-*&;P.KS0H\"H5.35>2&"C+LU,EM/:8L5)5_[9]!''=*K4ZW"3 M9T!OA0SU,9;->:!-M&$@L6AM7V$I\Y1A.VL$%MC4JKM&/[2L0E&KV'`3BW@@ M[VO[&I;-K\](_1J_JF+XZR?C(*49]@C5:IU^9]M4]`%:H9GS1^RF*Q61:20" MRT8NL.J4%M)B(1^*HP_`K&P&RJ!LRQ+1E,QTG-<3)UC3$0 MC8*]W`E-E%>6`Q&1>\GJE8%1F*05=C&2>+0A5'?%%(*0_M(#_ASPJ$M4*U%^980^VR)*O1 MFK2N(I%5'QMY-B=0:Z(2/^FLR"P,&R=1*U]F+)L\@C"QA[DT:WA;I1+>2O7< M.RQF-IO,;1R3JS8K4.*MGJPU$F@LC#B0W!JW4JCQ&=P*BFUPQ4=[0HR<0(M* MO6:?`!RI">4Q<&"Q56P^2\7V);H-)BOS8:PYP6!UBUROZ0;B@7@\I.5=FI)E M0"2:J^9Z8IU8JL85HJ<9*:/&V>.>E%[F4Z5)*5U,Z!M&!`[ILPKL?\+`BD(D MK;]1TK9EE1D"HIAX1?5'Z55Z@)KY*PO4R$ZR,2TD:]/2M(\L_&I<':MY1K)8 M!_&O_NS_&J[Q*,03(>RRMZB6/A6!''Q8F] MG$S]ZZ35SL:R)U#@,O@(950M&6*&_"$ZRTY6?N)@O4D.II#P8(^*L&$7)F&B MJX'C'Q&VP%)+[60D0;UY16E&8T'@/=\($=25(7[0LZK`$I& M!F_<)$E9M//M'80UIKIQ:>0$=!6(FU-J9D0(H#&'C!E`"JP!PLJN/.X'ZZ[L MN#ZN7)0_B1KJC@!4?S0>\$Z=(>,B3A/<];'D+J+G$9WPQ>8)#.YRQ.0>(#)N MOR'BZB9G$P3*F3%)_9G_1%6I+#UD;.IR:B"LUUQ'?$`D4HK'&4"I'+^4;EJ? MXJ>6JWL:Z+6GO%1;Y>::&R4(=$5TI2<;""/5G[2Y/5;0\>;,'G2""0F>45/( M5SV78:&,I*AG!'G6BGX'F+HN@31`F`?B3OD=H&CD>MZ(D>6K*6ET&]56J6CQ+8SSU`^^1;0HG>NH^,WZ&(3B":21UE=/VZ M4@V?6`512II)5R/5E%GX5[HA:>I=NFZS.\,\NX]:DRB&AB%LT*H$=S6B[FBV M^[5JNT-IMPNV?KO9KB?J[6:[88RZPYBL:-M>+>)LJ MFF[MCNM'`@UX96?)6KS:DRU<`+(,F:(!A\@Z\[HARF5KELH@ MKO_5\_I6JLP8:OFL2I^6J0DF/G]5E6=D&^&\GE$DY5#-O.,IJSM0(5HGIR?A M]X&)(M(-E<9&5V"OQ1GV@B/8BVA4E3H&0D>?'.I_!8DIEA/9:PJ:>DCRBGC MNYS&1XT6.709@5'%3;^$_`V5;8:Y-F8.AEG5\\BJ?J?YTB@TE751 M!J?OSZOZXKET[IABCK6^K.^<*_OFIU$E[6O-?4*P+DM[>,HCMV27R4+:DA-N M\(MGK&&%`I:YCUQFHL;)`TUQN"('IL<#O'@17*I@>89<:?M1'!Q@J3%W-!XT*Z\RK:V_V5E"1OV"('V'\6$3B*R. MY8E%ENTP!G22:.`KQL<*UZQH/"+;4I5&P$ M+*6A)XG7?*L-38C11A:I`+@EC8]](HR0NP4] M&T[,$+1,3D0J&+)Q)R0D0*SF^MBBL?ZJH+&314PD)L&UI=A&;>GHU&[%/@J, MX_-GC"`(98DIKE8:&9.`J1`%)?3AAVGNR@EJQA)LO;"_[B\B[)XIPC='_)N+ MH"+&AZ2!EGZR')CJ4PE?PD:')9P)8\)_$CFG"5.N7]'1TYX8P.GD8PMT93H$ M+*G3H6U`T1Q[Y$5W$7J"!-,)!,">\RE+"FW`] M_`[+PW9)*/QFM`])\)%IMQQ,J&PY5&@)O3>L3:,-V!C3B6$FWE0<\LPNTO M2IR0U5FWB5.X%?FH(-<:)\!H0;;I1$DN)E]%=0`O_Q(Z'?Z(5AUW2 M)-"5MN3$O/#.X0LW6V2Q_%46=Q3-Z",BXZ09VEZT(>4:+-S<)SP/#QT"%3XL M%]/#][`];'\Y5[]1_D'.]JO4T='C$Z>379&HQ7HM:@032V7#9AEI5F-X!>]RPI[$. M"Y5$'E3M3]8:3QFBTIMJ?XAZ#K&/^Z;JH;8&I+Q1E89DZ^C"7:)79*74HVI0HVR:G`,'":` MM,1)\'O2<^FX6U+[$:(Y'+L8<1*X4:@1:8`\H^F7XHG0CJV#]##^!`W<%RAPJW5 MR7CR-F(GISAPG"`X)\LN.Q\24J<).VX(2;+A#HJ,JSCUK'4C<7*<+'6<19:* M946-I__(BL%X"B)1@P!1Q)C8+%;+WE>BQW(C*C4\_]/=1'&/RF:PF MI\IH\E=2)J/)9.?KLRD'QQI-C?C\?HC@[6,$:<1O\5L;$^B<9DSC))4`)20) MR>C1*Z.X6DB#G`&W3#%9-5DYA1GT*=OKUV0W%FP%RU,),=4RMES"JBDR?NVA1GK!1.C]Q21O*]=W!<9C#O3LG:G48HW;PR;\,\?3#,#[/# MS.I%S!0SJZ>=0LR9B+)FJ.66WK'!O*-1'W%,]LA*]V\X4PTDI&$(JY/^!R#294R=&55&1?=,3P$' M-E(JJS"A6LVE*-5&Q2EO>:K=E"Y.E@,U6Y;J1K#!B+`;X25B!\[$W/CI#9?!"<:E#`PF:8S&O'FA%!ZLOJ&#G& MA?B/"B2I(0#YRW_355/_=*A,)!>J48%9>.?MU5$@84(:DS&8F,Z61G"9+XV^ MAJJ`7&B@P6A):N-@R5$*CVO6D=9Z'J1@!/I-']_71P+'%_@^G5+V-:GYCK> M)%!'*2:7'$KJ$.UUI*1NEP[%%(Y!Q M":;^)LNI=0Z;)*O8:YEN1LVJ,B(;B;49T@=-ZEV!51$#=H^2T%OI!PVHF"I, M:0I=0HO0#-@(W=U3L1&(*8VF'GU?C2`8LRT?C M8FH*H8*1^H9R_M&!)V^D<@[2.30A+4@7THCT(>T3;:",M/'62$M5CW0B%4D3 MTMHI)/U'6]*8M"1]26O2F#1)L]WPT5X-$+V7]9(U'2,W^-9<>+,Q26RDTEY' M6=*CS!FN])X92TL9;FHB18_8TF4)+KV,.1H7V2Z-EY4EC&DOG4"5)?1(,0;&KC%7(YT\TT$9)IL*EM;\A53;.IMZH(]:B&7-ETR(4TLZG? M-(+2<6K3@H?0RGH25/-)VK)^@!HSB.+5S-"'4(5'0EUTKQO)M2.&GUR9US- M9B]$W?9>HX](TP51W\R\[4;]3.%8I!S>\* MM!4)65=/4=9KOI7DUXE6&6-789UEB"EB2OKF=^B7FNNWZD#),+1)`YUP;8G[ M6@$F\/191Q4(K)/P-?4./+IB!\L2)%*4VGI$:;\6;(AJ(),"ASV=8X*WV]PX%D-UCC^9:\ ME7T8CM9782>B;.T88/L;1E"!K7L<:#60L;:&(""=""NBH*DGI2/VW+Q=UY@/ M/E6=U,/PSX[F/`W8$<\-)/2&UP]8%:IO^<*!)5PK0?ID9$V0:V,#N7K)_-C!SH'[TG1FE/6D8\:L8BX]P7TO#L2/(,P" MCU$JS8FU8NW&;)W(>BN()21I4I8[TYCS6)J,"A*P^'2$D24UMP45QS9K+I?- M9>-H?=JY9W8X4]%&G@//SM;V]64\!N&SN74\%*#I&>`00450!8]X=JG%.^;9 MP*/ON#OR@6I'=3U9[\R`B.Y*C`U)&&PJDUQ/L5BP`PS23L$-K*.Q'=9.<0>+ M]KN945GPHZUI2[%1K*.=:4NQ7*R@_8[9A#6K/Y-H[=9&]NJA=$PFX>G#XV*Z M0F3VMU0*`;&J4*R=A\S:>Y*LC8SEVK?VKBW'VMHYVL@*:_/:.5IJF!=!(<+( MFN&]NJ+RTO[BU`1HHK++>UGQV4MKM;UGOXW4-NT(:-5&NMAS9V&!DB%(JLDN=TV[!ULWUJ\];_ MY9U]#C77VEAR+7`3UP0WX%"MVB0IC#B M_7C;.S.:&B]65_(:[3!5C`.M&>XULZQD9+28YOEDRO0[(FO- M1=9\VIXVAQ8@1Z*/Z'`R_(QHU5B),G]6-ZL5S[,,>KK9]@H\OO;98+>U_3O> M4+P)">Q@63^+B1,);7C;OXL?*'$*/G]E`LUE$-#7Z$>UJAD=%%HS^0?-'))M MWYV.Q$R!V8[9A9E,R;;)+1@^>A;HM"-V$$`7QWJ[H0@%3]+<<36&*`V90?H` M3GY`A__&FU*AG/<2VGE/H4OHYNUYC]Z@=^E-5(?>G_?I;7I+H:IW6W6M3(Z$ MT1?S)]X?04`0`(546TGN%\4Z-:_2,\%G\OG4V`>S-WBV.MB'V%Q\;X+WL)`M#$43UU@8)\'RW,]]&2HI1/ M:'+8!=8L,U=:7/G3&WW@TR+0!()OI23X3\V!C^`FN`<>@I?@P18*?H*OX"FX M"ZZ"L^`S>(X)@[_@-+@,KH.+@?X/67,XGB1$`%N*DJ2[IJ;-F!?U2`*B&5A' MPW!"`,W1Q0%9](<3SH0[X:92G4"%=SG*ZR9X<3CA6KCRVJSH=\KK8.-]..%B M>!B.9BFON:K&X82GX5_XY@GGHN&MH-+,;T#'<#@=OH:CX7BXH[5^K.'<*!K> MA_/AFN@:_EVA6&"4D2+YD-^#N!PNL@E3@C@2IHC3,#^U?*(8H^$I3K+[B(,P MDGBEJ9Y$(H+X)AZ)OSA#,Z9Y9X3BEGC*_8DW&=HH)=[$E.",5"HNB"O-JC@( M8X!KM($-*=AF!3;*'V7G@V?91_C?=;WPA.);S5N@D_CO]4,WHR;X,VJ M,EZ"+Z7B7O?-@M_@[/@Z?N10BQX.%DSR>#'[?ZK@)WH[KX^MX/UY@ MFF/18I)7O>TG4Y(PKFY@5TGO<(S&L4;8:3P:K8:<>E9F=4Z%,G-,:?6E<%SR MLZB4^+X@5%,@A7<:*A\YU1*2!RB2?#M MDL_D!,A+OD6SY*633=Y%9]$X^4KND\?DYJ`/RG$RF$N$5HQ!>7`L[+J1^6/?WN5&7NUMPKN`;>H'D#]S\/?@Z MJ7IX/9Y>_M2XZI6B,T+/KI6S*'CM2HT/\3UG*B-B-60>CZ-=6RFN.D!!:M>' M>C*.:>:9>584?%NC9`MHGKX`Y"`?IR)I"6W%]^^-"H, M_&VOE*N-2-U,7@C0);7^68NVU#I6QEY&UB=5J!&4GE/3U<`+M-#6/F MCSND9ZE1S_DV,YP+_="L=BL#<@:_)*\@-)'+6-,7T?$A%GZUG M,%,L^=2BO<4,L)"N..*O8)QN;CARF18K2Z/_$V_ZY)XG3[YRG5Q,T_]]6.&* ME":FZ8%Z($V.Z836[*2]B_&Q!&<:W?,'XM32VQ8P#:.5-F?%*0#*$`V@^F>% M3H!Z@;F,ADXBL7L8>HA>@(KHD)R'#J*/Z"AZB9ZAJR^[YJX)CYPQ5%#]4;WH MWLMGU0V9$">#SY'=92S<$Z>,4[4XXX**3@W,0>-#NI!.I/?FS/G3*R/QC:'/ M7X^4T;,6QDA>E*8#B1-/_:6GF2$ M&"$E2*Z5G'@9T[`(*'J5*HIK8BODEC31806=:!MC@2=JV'N^G^QGG\ZG_^GN MI_GY-N\/1O31@%D]Y8VD/JL2:@#ZI-ZH`Y_6NJ%>J6N MJ6?JY>>@SJE3ZLSFQF/NIW*$?N?ED;%@P**3U46=8 M'(SLEYO$Y@_KJQGJR MSJQK/M"K'9R]2>O4UCFZE"`C6>,RU8092WM(8AI/"4LGDG8*!Y_09`?!$U/F MJOSC4[D*GD7K>J2\KC^5YWHKB#S92..Z&45L7;J.%$C:II?KO*+,V8K^IQRK M=+CL%Y2K3+%?[&LHQCZQ M9^PO=,>>&W70,S0'_9Y/&0H[.**P-RQ(87CR<0?LU'-+"I6_I#1P^ZV?R#"DK`$;C(KIJNMT6OY&CHRU,KX)C56>W73F>VV:)=[ MK-M^JWB4&4:/'>VFE>/38:?HX5@G9J$_[AXZB0ZY`TV<*QU23UWNS!GF;KEG M[F#:5Z5[@*!+#JUQ]H2JM5/]D3P.ZW1@X'?>53/RFV[$JR_KIBRRWJS/[L,Z M[Y"UV+7Q9E5Z]50@D5CF\H=';S9 M2%QS1ROZ?FDH+P9$X!>0`3+&;G`&(`:S`8UQ-#P%WI&D,!%D!"I`1J$C&`Q: MQHT!&\@&V@00H";`&&2`]PX'?`5H1%F`%@P3!H-ZBT2Y#X.#$<`&H`!!0`H0 ML,0&*(",`0>D`'(`"M!A!`4I@!F``@@5*<`9@`)8`6D`0J$II`!H``I`!:0` M+4`!(P8H\`=\`H\AO`&*P5@@P!<('_QB$,&C`$T`%)`$Q``I@,91(-`I>D+5 M8`2L`2B`%.`"I``>T@>/!+`1!WP<@`)4%@+\&">,K!1Y``IP!#0!1``+@,$K M\2,!IM#`HP!.0$7`(3#P&3R)0!'(`6-`"O`&H`!HP`H_!!`&7CP*H"1,\1!\ M&@`3!`9@PTB0`I0!"@`*,`6,`:>"&[!$I`!VP:!``IT%2?R6<"A8"SJ``*\3 M?/!2`(9@!K@`/WP;$,0/\6P`#S\H?/!`0`J@!J``9T`;0`9P!65`"B!N\`7C M!Q)/Q;\!9;P=4,:S$66\&)`""/%N/!F``MP'3CP.3Q/("3%\#7`6!-$Z@0T_ MR0X!(D\1`\#K_(OP$J_!L0Q_,.%D':<,HK\$6]&3_%-PB" M01G/%OP$X[QX,,5[#/#!13_)D_,2?'@PQ4\!4``3_\1G!B0\>"`8I`%>?!PO MSZORW,12OUKL\"`'`2]CM`#^_$QPSR)(!;>#'+#" MF_*H_!-O&60&ZWP_G\$G$^J]*O_>[P9203^O`&3PC7T8[\#3]AG\&-#93P0W M/62OQK?P+_Q*;]V[]'W]2O'!'P2]?1W0T)?QD/UN[Q@$\7%\AV'"\_0S@3V" MU]_SF;UD\,0G!1K]&-',GP28/0N?TL_S9@!;@!IP"#,]"G`'8/==_!CP$H`'L_Q+,(7$#B M`=E3]K"]\P&/W`"S?1E_!Y3Q20&&0`8$\A(\-E_0J_)BP,9P!DSQ%T'.,,]7 M]J#\-`_@0_>.06^_SEOSQ/T8$4?$\V?^-*_"A_G%O`0?3XSUCL$98-;O`."] M&5_,J_%S0!X`*9#R//V1V>LD\>(]#Z\G?/#W/5%?X1OZF3RB#R9P]L(]9>\& M=/=5/4R_W5?SDSQM+\$?"CO^XE#.D_AJ/&\OQ/OV)OPUW^9C^D]\S"!/;/E? M/&#OSQ^9X@8(D,2[^GG].G\1?`0I`),_.DCUBST*,!$8#72\//_?%_,9?!C` M3;`%%'ZA3^)G\#Y!2)_6._4!?F#0!DCU;H!KC]9_\2?!#I]\@$=T@OM@PV?P M,#S;0$_0[V)!1H`7\`7X>[TQ$[0$F(.S(,'#\+1<:%_-NP%F?)R?P#/[W[V] M#\OY^LY"!D_,9_+:_0>/X,/PTWQ'_]%G\"%]X##22_`0O#W_Q,<3>P!5`,TS M\O/\8+#7HQ%I`SROZNOY#OU1G]/O\IF\!+\%S`!=P`I_W_\$4@$;4-!/\@%% M94#1V_M\X:"0Q(\.%0$//\C3\Y`]1P_*H_(4_7Q_59ST*$!!'\T7_7`^*/_% MK^ML[='_;+Q!$ M?*L?Z)OU;<#BL,/;^WV&O9'$B_:+?D\@P%OWICX*(/%#]EB_@T#&/_3@@W^';\:[_"S_@_01D?QG?UPKT#']\7^,Y\(Z_&U_."O63/]$\$ M'P.*+P90!2D`W?^&^?HYOC=O\5_R*4"Z[/#/$8(\`5\BP`'8/:3@VO?S:ORB M(+Z;\&9`'`\Z+!"U_CT_]8_RRCTH[\`[^=F\8#]&R/D,2W]HO^1:\AI_#5?(C_XI/S#GP-L,*3^'S^ M]/_WK_`5/SA_Y`]6:X;=CP(\`2S"/$_B/_O3_V-?\:OZ5+[C@0EOF9?LH^TY M\&(&;8'6GCQ/C4?B@^,U_TAXSKUY'K#/MU?\8^Y!`'E_*``N@`)/O2?!^_'1 M`2YX]KX6Q*_CY[5_\QX)4`A'I$O`BC!B^Y9]F)\UK\[8+2O\F?.JPS4 M]`1[,;Z2@"*/]*<&(.Y1]%1[&#TM'Y=O#[#J4^]=\]``$X$M@4^@H5<20/%I M_A![H8/YGTX`SS`"1`$8`5(`_[N+``\/A4?K4^/]^1QX,SY07B/P[2?M2_9Y M];IZ.SXX8*B/`5C-F^8=#ZY_-S[BWA1O`CCGH_AM^CZ`E,"^'S0)BX'R0(Q/A;AR#8D\:H`QP/<7D%/-_'!N_H]!"F`V+X]H!\PF7#IBQF$ M\HA_^T#C'PH@GR`*;/^M\%QX1(`G7C5/`?#%D_"5\I9Z(X%DW_[/66#/:^0E M\!IYZH*IT4>P"?C$8PC2^AQX$D$R@!-/"LCTX.,U^7""0(%?PA,/JO?MJPF& M^^*`Z#]JGLU/\1<&6`.,\CZ!T#WRP!V/OX?>6^&Y^@J"1$$ZH,9/@K=$0"B\ M"RZ!YSW=P4U0$2@'8"<,^=QZ<\#.WT5/"G@X0_WY`^N"&H1=(`&/"YC<<^`M M"K9\V#[U@#>OVJ<'G`'R]8Y\<5C.;-\PR`0SP\`&N/YC?EB^.!_S8$\CQ['WE*W4?]4^/A M]B2``@6:7W0/[3<4O`>N]\R!;CYH7QR/\7<$?.(A^%(`;(,""(&@'V'7LS!\ M".`1X`440!+`"7`+_.(]`8X`3[VU`,X@+XC-4P?.$9QXTSSV7APO^2?\0P%\ M_(AZJ@&EX`I/0D`1S)!Y'U!_W<'I'SEP)##]V^B9]B@"5KTPGS`PQX?A&^+1 M\#$ZW5\\&A[TSR.GVH@RF<68!$<]WJ#FOY5<0].IE M\)!_4[Q*7W[OHJ?&,^6A]LIY?;]$WZ+OVV?8DPWB.0H8/[]B'I//Y'<&<`-, M^!YZZ[]S7F/0'OC;2PE*],!YY$!=WY(0+;C7&_:M],X"/CRIGXR/A^?@,_2] M]-1XJ#Y3@;./GI?T\_N=!#)Y#CRCH&I@7D`1_.75"5!_+ST)GO\O=)#\0PV^ M"!*!0[PDH7*O]C?/B_\A!,5Y!<*H7W]P@<<&<.))\.)Y8T&B7PZP`+@'W`]6 M!5$`F[T[H`-/->`/'"44^(![Y#R^X+K&_D?;4^6Y$Q`*/P$A05,@^'?K>^=Y M]HIX!+R%GZ2/@*<:>!%@!>``WSTYH9*P')CVV^9Y`"]X&3QJ(`U0/V@!/`/D M`40`X$$PGI9P/OCQ&PGD!4\"6T&^X%[`DT$T[\P(0`PZH<9;`/(`F<"4Q+[W\)/N*?*PQ=N M":-Y"X0]`$O0J"?CJ^;M`XV`DK]P'X30A)?!@^!-\R2&\KP,7E(PLW`1F!24 M\X*`G4&*GCMPFF?K(P/@^HA[G3Y$X%"/KB<;K#?@]4Z`%#V<7Z0006C/2_:% M\5![?T$!'_)/9[@!'.[!`J\$*[VAG\T`LM?#&^(5"R^!8C_3'NR/+Y/\)]PB`,D'CG^_OG6<.-`SJ`KF%GAJ/10#9^^*1 M`R\":P!\H%ZP3[@;$`5."2&"PS[.835/-CA^H`V*!5@$:\!Y7AQA+<#%PP!& M^B1X,,(,WB_AHN?`&^BQ"(=XV$,ZWN#/C+<5K!$"!U>"H[PR@)@0?G@4)`)& M]:R"U4*3X0.PMQ<46.EY$CYX2('!@-SPW(?Z6Q.^"P`$7,*_GI/OL5?.@^]5 M"OIYDSPK03EO,/`Q6!&>"/N#.;V+`*;@CK?,4^3%^<"'0[S>(>/00!CY0Q"J M\12'VT/#H&/@VX<4A!L*\2YZ"@#9X,BAEH"$U3Q57A6AC`)N$.6(3CQ$@@C3`=*$>,(2\2S(3`1P%,(M!$V M_6I[Y4"DH-H0#0#7&_JI"I%\*(!D@EO/#UCQ8_J%#=]\\41B7U047`BS/@] M$U%YX;X%8#D0ZR?Y"P^P"%Z!,S\ZWE2@H/>_4Q>6]C:&1@_=."E M"Z@"J`'FWY[OJ.<0O!46#X6&)D"EXFB@%(/Z0QYN"*N*`#PM(*TPM[@_#!/T M#_."7CT%@#U/0Q`8L.)U\8J(J[UV("+0!'A2M!8T\B9YDD55WG<@0)C1:QJ6 M!(J$7<#I'V]/`<`D]`-F]`Q^$"[:8%SP:1A+W`"*%S&))($)8?O@/9C$4Q4* M%!N+*[P18>(AW/?=R_W!"Y^'#KSC(-L@=C8:$#A M_HJ)[KZU`;RO*H`]<#[4^PJ(^+[^G5#1')@^_.#-!VT(%+Z)(5*Q[T<5+"EB M!4V&<;SS8(9`C-C5VPV,"=]\/@$2X0R`/G@IM/R1`3"*T$49@V-PP^=C_`-: M_M![;CS1HO.,#)J M$'-\?`-KP:@PS(!U>#Q,/3\:3(I=/2$!$["]&#&%_D\)G(SQ">VC@&P[. M&?M^K,/Q8H@/ZG!J-]U,(GHQ.0I8<"0`(,"T=YQ4(YP%QQZ&<4/.W] M$]-^;\`L8ID0->AQ].^E_M")G<$$H8]OE$?0F^)-!,X`HD*=XSQ/CMC[H_5Q M_X2)?C_)P!FQGY=M^.`1$7P)![S/X#JQ[U=T?/RI&]&&UL9XHQV/G`=()`R\ M&I&$LT3U'KS1SPV.8#P;`@%A%4@8:`5>]LZ+5T.PH.SQJ$<&//(Y'A=^R4;8X]%QLYD(N]1/*!OQ!9> M]"*.$KSN8U4PC@=^+!,J`3N*F[]+($4`&I@=/`BZ!3#T" M`CU_'#V\8;Z1G#?)0_[Y%.>#9#PWWP.2*3@G0/W)$2=Y24&=X*SPS==6H#)B M('N/,#X%7ZJPF`>/*.T5(#.$;4`9G]IP#]`4H"56&_V*P$/AWC1/LBA$S.JM M![6(M#VF()Z@"AA.%#MJ]*2*6`&T(]41TQ<@A/#Y%AN(X$2IX)#Q1EAU).M5 M#,N$N(.6X@F/@#<%"#22!]1[T,'-X(.Q[^A\1!/F%BE\;T%!X)SA/K`7P"7. M\XH`6``H@!2@")!B6.%M`6@`F4::'MK1G?>_:S0V`1Z-,($IP,7P/A`N2^() M`2`(Y\?$WV5QJ%<2"="X6!F^']4"(X#POTN?`:R$J M`KV"?KX#WF41)M@-E.H]"MN/;<+S%UZ/Q(=0Q!RZ\?*%8\9DG_9Q*HC;FQA* M&#L-]8@,60W`)8!AA`YN&*>#40"8X`[/@5>"["A.`CV1S4:HH+:1!H#[>PHR M(HF+`\$I'NT0'%A`=`VA_A2*C\3`GOIQ+CC$@S-*\#2`1<<48"F0>>@]].9! M'2^1@+X](`60;&A_3`@:_AQX&L!TPJ0@!:C@>ROR%R<%<\4N81DP?M@('/*= M_3"#W$.YH5^F$+@8@!#J'R6/WT;A8`$O"1`33#]Z(RF2.4$](1IP8+A]U/"M M\#2`AT&-I.@Q&@GKHQH2!GU_\\(;WWUPY(=1+.>9!-N.!TBYX1G#*0@CM/SA M`:"!NT.`8SBR'J'NT_>A`&8`QL0[X[0GK=&T!8>_M1XYX%^7IJPP+*`/P#Y<(UT):'WU2X[!FQ!U:$MT+`,;#WB8Q)"B$3#\B M(K^2YD'BGWZ0IVAR-!/F_UZ$.$%O88CQZ$]C@"G\GUX)1O@6CY\TF*\MH`@CW60;7/':C*,RR2 M!):4-TIYWC"R'EBC'`N8&\5_L3TWPVSO=H!-).!M$MM_/$+)).IPFG>8]`J8 M(*.)SL6V(D"1F(?LFRLJ$-]\S$@XI:]PBK>"9#!J((>#FTBK'H>/F))]A`I* M_TAZ9<.MW_317.@L1!?"&I6`IK[18:ZO=(B,1!'*$W5]6#_0W]DP0_E1Q/ZE M"#E<7_7E=01:G\>U)V M_D"1D4?+XZ#RC'%OQ.9]^U1Y$#S48=6Q3]F"I/ZQ]<*+8D4M88#+7MROB^<@9".@+,]\B#^C@9E/YH<'P$]^]S:6"`?X8.(Q!#D:.#C< M&U6%B#X/Y$$0ZI?QBS)2*U^!>TK#8*`2@RB9I$Q2 M'B%[^,7Q7!*/"RBR#`W"\#!^TTL6GR[04VBLI`&Z*:>*![P#H)5``7@&6.D= M'U^2KT?'HU=O0$GAJTR&^NB`#SV5($PQ[1=JK##*\AZ21D"R7]-13DE&I";2 M!N&2.<;Y'40`+!#O*QX4#PYZ:X.4`)#1)="_D^#!(OM]]S]1'L!O)("DW/*% M1V*/YSS\86"/K!?"4^#9(7>`HD&@HXG/\6?PTUY:$P^*V<1LI3-2<4EQ=#@V M+M>3H<'#WX2Q"GGP`XG@!9Z#*X4-XPZR!_G%^T$>(:N-,LBYPWCNX%>/5$=L M&(\`'<8/(R#P[`=?3.X9#YN1>\H5I<:/YS)`O,*F4'1H!#Y*KP@[=)G-\)!:^$[$K3GR?OAAE3S&'Z M%ZJ0]8B50CTRB"D==`)$`6Z1[CS0(3J2/$I^'_F5#\,E(410GL@01$A. M\S]^ M,?V`1D'4X&'RBJFG9"O&\>R,H;USWB$S/@@(?#[R&RV55$/.']F1D\GG.P;& M\?*9Q;P0(%\(6XFXW!`&"W.$BK[*0)I2-4#':Q7.\SJ51L?Y8GTQQN M#.,$:\8`WB1O1#E/9%$Z)C&![T7$HA93&HB.9!(^,X\&@$"HGU?O1ACF&UJV M_H9^20"NH>,1MYC)3/39\;9]J$"X):#P>6CO(P3Z!8=\14H)HMPQ2CFDO&=J M(<6):;\YIM4O6]@3?`5^"C<$"4I4XHPR'2D_%&<^]5Z.`D(S168//AD.]#J\ M*MF--\N"(ROS8'GLZS?V'16#*<=U?&"EV9=KVY0UX@SC3+C`[2 M(@%X/`;<%J@181/@"OU!-&'!F`^X&$8"./*=CK(P%Z M%?IZUKVVX:D0H0BA=`%V!.6/>`!IG]8O0PG/D_`%]@B7*[SCH`(@.=A`66WV M^CJ;0DP>I`]2TFC8C/&9*FF$8T6GG@EQ%#A^]#TR)-$-/@;47U(0_A=R#.R9 M&N.;BD4U7AASE(D"N`*4!2L4=<5^(UJ3'XC91$H4"/J"34S/)J,QM*G!V^"= M`E]Z`L799FFSIH>0U"):+9N-K3WL90&1.9C$PVDB%-E\"4;UP!BP2'DHC!W@ M-CV*M\N_Y3UQ+)!R=.#]$65\PD=3(!%PRI@?I/X)%OV+LP$&WODR65E03&46 M$,T"K$=@9@6OH$>=O&'&^&20>I,"!E-B4(#>K&5&`7:01\SPI@L/.5B#?'`B M.44$';[H(#Z2&TDV)#\N^@I\!,,XGCKPG-G.-%PB`I^;-\/]HU7/1^A12.*Q M!SN/O$NT8T:/:XGCNVT2.?N/OHZ`WI93OS#`"Z+),L.<&\;/9J-Q0SC:A![. M'066S,.YIHSO=DDV_$9>$/&`GLC%9K0RF\E.G%8"^=J`Q,>*IFP0Y,!ZY%[Z M&L<`FLKGX<,067K%Z.%W%X?DZ&I?+0Q-C)["1R`ZN1J$?[8>SP()#_ZPGJ M)@][!D&ZX/VQ,G!Z9!&4"3E_F$#;`1&86,R^1?KO.SA)!V0DT6IG].RZC=DW"Y:!+Q[D3_IY1+2XOAPW!/B M%AF9A3\.P830+&`O!$Q^,#V=Q,,T9#+3N>B'C/;E.QV#-TUM9$LS[:?'1$L2 M]JJ-,S["@-^S.ZY*U>,0[P87\`R#*!EW"AB.BN96,S8(F@PV:G8^^R5 M]JJ;U$GMI'Q2;F@TM$)N`K^%&KV]GG=3#G.#BDRQHAN3#!CDA%K*,/>44,7\Y<73IHGC:Q#P M.)V`LD'588&13)FA9`\6+Y,`7,XS1@%1=5CEA&)Z&$&,7;VRIQ\0Q!?'*SOR M,\F=84EXHQJOZQG'9'66&S67\#V.P%H@(LC)Y`O:)YV"7\V*(UC0]@?\LRCB M`12#X,*7)2%RBD?\?`6B*&^3Q\S$XA>3[\CC1'KV,Y^+@$X#XC33(4BR-$:N M$$>66D=K04&R58CJ[`@X_BR/$<=BI;[0#MG56V'J&YV$\,:JH[#R7:E<;#70 M!B^;1`22GHLP[3A+9`M4&XV"J[U&)<.`>1@$(%IZ-:J)1@`)GOH/@QG`:[QY M'7.5I#_[7G;SZ@E,O`MX_UZ89,@O8FI0NTGZ"RS&*LT"4$=*X/+W!I4I0CUDB8`,4"T^$64'[8$=R=QAU3#!V M%Q<#_\,/GA&@@L<-5#Y"%*V::#XYP!D@/-`CJ!/>#MH`\`$SZ%>0MNE`1'N2 M#-T+TT0Z01E1[6G??`5*-Q6/.+PB@$H1^Z=\5%AB\SBQY&O:!1^)RTXOH!G16D"])/&% M%/D,MLM"(O9O6!EL#!TX*H^1.<+8P2DQ;H#.,P*H\B8"Y`$\`)?SE`E0)#@R M^>:/=\":93C4A6DG%8MHOZ\?/(R`6+0.;+((M'PZO3#NJ M\4J+-KZTY(F0"^A;/(5>,^V6_D6Q8D31U4FGK`D2^T2+S4U,GW`/5^E^9$J. M!D623E`KXJ-PJO=D-/?1&`V8J#]89`)SQY@1<`G\.O!W%X>5`DN`@FE-]/J) M!3"1[D+.XKBJT#'*&5/>:@&V/NM M%'8SJ+\%H0H`!\#T:S6B`&``.H!,`PL`.9II*.,Y1D&6&D#FZ*-A.9H5I M)!UX/8`5'CS"-]IIP#KF116C&DD)7GQ4@D<<#0?&(Y-XW=%CHD3`G`G3^^NU M/6V&($>DXZ\2QV<4I'U^/661Y4I-*"`0<]G'?(V*'FV2!DBX:$!+OQDZF.?% M1QUX#E*R:$O`_G<:7>'M1S&;9@D/A;VOB:D`2&\2,5$`1LR+)"7O+MH-7`7B M1=FCV5$W`7000!K#.R8F`02:SD0B9*_3L1F2=`P&\FB.]"W:2/./@%`ZYFN4O"F,H!5X26FCT4';J(X@1RIYP^L% M2`EX3(!&H5?PD(G3?'1R(^=[VS[OHR@TQO=0Y#[:1/N+5L))Y`?O";!,*)"* M1T62U4`BX"DOPX<0]`9B`#V&&L`T``M`#<"2U$F6(GFDWE$JWJ/TS=C@,T-2 M`D.E>X`8P*C4,:H630.L`&2EI-+-Z%.2GC\Y)21A&=T#T,Z9[4/9H"H(_6"?Z72;S>J'YT_;]T;(J,$@VH$N%8UJ&+..9@`-H*P4.UHMA8^N M^6*EZ%&S0`0MB:6\^*C:KS>Z)+TR.`4A(PV M*]BE#-.`*6)T-2`O90&X1:NETSR$*3WO5IHO];8`,)&,_H6VB9KATW(FA9]'I%T^A(=-97@M"'5$R'9C*2F<`*5.-*5IO21HG8#WV M2]FE/5-W:8O49$HP90'@3#&F*M-Y7G4!AK M3-=\A[\EJ:3C8RH7C)D:1ML-4].!Z6+4:IHQ59IR3,^E-@&?Z5'RE,?]$P:3D/%KDDI1-\2:%^8M)!)QL%WK&3LC>43,><1,DG'D:21@H*K0E^ M^R:,I:/"*G=M.IJ:F`#,`G97K4 M`O^D_KYF7E#@/$J==&/N-P6>QY#RRZ/IAS5C2*U>6 M`66D=0"S90J`#I`!-)>FRU!?LXU2Y7[P/J"D[`',2TL#/TU6I5D`KU<[E7&& M0^V*D$(IZ.ZT=]HF#%C0(%&EQT0\H\^J)U`#F3K>*(%,000MB-(`^%9IN&#>G.=)-A%.0>$H`(^!9 M`9.,W<=,GNP4B7)OK)V&1Q5X&@*P`?2TU>@[72GL+M1;?;QL89T`/-*"N#?. M04.@R;\80"!ODD=C`#M:^S".S4QII*9/04H)]"H:0A5XFP5&Z!7/J[GVR^T] M\=QYY+XL:+208=@+93HZ1=%]'%$4@/IO*KK`K-^!`'ZC$$RV@221*YH`S>&M M!KB,>CU@H@M3*%A%B/DM$V$"V+TDP9Q1%YFA3.B)3VF4?K]R9UY0AS?%4P*X M``)Y<3U0I@.O,9`"V&"64:V'+SYYWL/SS5=87(*>47EY8\LXWIDR\0=O#/"= M4;U]3-$S:JC/Y&$PZ&+^^LZH>4310GB@BV=&M8J!,)]X($F`7^3QU+5 MRXB6>A/6XQ;`EYB7).#!#*0"Z02A'R72!:`\M*)"2MUZ6M3L'Q=UFN=%14Y. M^GB"!;Q`ZAX1D$K&]!H&!1(%YD12J,B/=$!(I>;E_P!^P<4RJ@KRI*A')"5N M"=1[CU1^WAE5_)??5"D\324#=P`48M*PX.GI4Q)D!NZH;U2!:M"S1`!QS.O= M404#>52`WS)0,EE$;*3V4R&I?P`ZA0T@6#@9W/;9 M`79X:KR@*M%R6H+Z>_DY4=>%5H6WP-1@KFC=\Z(2^M1XB3Q_8*'PB:?#(_0Y M\'YX_DP.X1.O`$KHF^9=`3:$08'4H]0Q@T<$^*+6]N"I`3SRY1WU@/=(;2L< M#$.6M,(?8-VQXY=$J"P$"U.+?@-`IWTRIZI'S2N>^>*/2@*='CT5N5<.=%-F M^Q8&^SQN)\"OVC@8&.@%"`D#/[X[JI&0,$`'V.$5!X2"3X")918O=!!YS.!5 M\="0*SPF`#&3OK@!#`/85&>!C3?G9D&/R6^4Q!-J$-L8DWBORJBK7,Q6&_CYZRD?``$GT(%C50TZ" M]N"I>,:+Z@:3!JC*BPE4'>,))X`R(=213I@7M`*D`*"4.4$8Y%NR(9`7#*J. M0S-X1,M,@R=!NEH#H*[Z\(!X.,&QJO)1A^=.I56&,FFAUKT@0'#U!;#$ZTZ& M"Q&J)DS!Z:*Q#1!-=326\2*-&LE\(,'S%@BEO`A\!_2H6L2P(5]U`Q@I,%A& M2T6J$%8&WD^0B->(A+"^`6ZK;H)J2A)/T*='!4^>^29YDP+>X)L/L>J)'+&: M_,:$XM4,WQF5QBI2M;'B&D%YFU6\Y3(5(S@22*]*%5L&_,[OI[(/D+H.!*0> M-L14Z`IX%N^,#>$A48PWE50R*GLD`P<(&*[@0AFK"K!O` M0P]_^S[E`[C0CG!V4/]E\*8`X8$UP))2FDI-A0T^\62E4[PC0!1OI_E/G&SV M'^VDGY8D7E.U[1C/G*G*1Z^I.LF@Z!;`!N!?K3D"6%%Y`M:K:`[/!=!6K3FN M#?R!>4$D@`N@!3`$X+1^]8X&00&595ZP`!K(>R,<6/.J<+Z]ZAW5KUJ8//Q9 M/"5X[CPHI2T2%QEI+/2U]LI\>M09J"$PLEI!W&#N(I,QOCXSA<(/S`H-'+,V M\A*'9E82*[A0S3K%8[-N$R6K",(X*SQ">4AG'54R_?"L;@`]ZSR/SSHI\+/N M^Y2'@E9!JN?/T"K!6[2F**-YC=:*Y*%5\D`;W*3F5W$A5=4;0*5U"2`'X`+4 M)SEX#E#!WF``$]DE?`*04\.;2SU_*PD$X.JOA!2B`:24XE/<`4:4@"=J;;7" M4U,&4\0XH]?QPRI2U:S.2#&`W1M5MV'C53QJ@(OHSI>Y:B" M]!1],%8U`Q*ED*G1=`(:!82";-86*MX2,*#(.Q&,":6,?$L40`Z@&[@56.$M M1J>!Q=3LWRQB!K`0P',<5U$`2@!":Z)`V!J;R@'<4&N,.51,(P^U!,#`Y#&Z M#_J,$;P6_$%7&^!'3P"7O@2.KD9/!TB M$S^=A;X-7MPU]EC"E%K:0P]^?TF"XT"33X@,/!<.0),'ML`-H8=U? M0&&5SLE):NE5!MF,*![8(T2CEE&^)Z-S@Y<#N%7H.OW%?3*PVR@HI\%7/:,J&8K]>KJ/^5.-#*'!8> M`B&264>98"@4_)@#U"*6'=FO5@X:Q/LU^'ISE;^"/OT+$U@5+/[5RHG+3`'` M.\RO4-0PZZ!4T\K2]"D2#'"",#V7@+B!C^+"DLSIF'/XND`))+*2)."&LESXF:/:.@X&S7^ M_'Z5=,T;G[-`E5?^LR;6-V6A4L`IQ*N2R3<%R`(X`9X``X$D@*3Q16=\_;WB M+W%Z=5'PY<#3.*@WS1,,"NH1[ZT;Y'/P7"P3@`JP"H6&;MR>VLZ M8^T$=THRI+LP<(HH-9QZ8:65?;_18^N3DXGOC(BV]T:HS$<[XI%O?JH7\)FN M`A>/^=AV86EO(AEQ#3_."C^QG,B+H^35/>G87,@Z'W`=]K^\Z$^3%TC;ZQI2 M9$6A([Y984>Q)NEH[?8E8--\)DGWG_[Q"$BT'#FT)4NO?8HIP#$O"0`%H`(D M`9X`3H!G;`&$""L45!4R^^S)X6]B$X@@6FVFF4.4% M8R-XUL3>9B`0[0@`+?ZQ0)>/0D#(8&E/",#DG)\B)2B)BTU5GY8T?`HIW5GF M"J6E^D;Z9471-0@!K!W6-*61*U";H12RY(EDO*+N+-U^F-(J(9SQ16D%U!%J-/6?ET#"8:"Q/A!?/(HF8&$$94!@*%+1@Y@V&(]6 M.I>B8];7GD>60IA;!4Q6,_VQ%=F4+)]2A9J195+R!P62AD7FWSI2N[B)K<8> M3JV)KL[)(+?3#EI;/$*N_V"%9P2)9_7UC`GW//,19B=$]C_<8JI10UAE+!%$ M%1N)=\%'(B"R0PK7+,XJ#Q6-'\6-8X"O#^EK/$.&4M5XJ\MX87#/#2F4M1,` M"7V-%\5+I1R2SKBD-&E*2A.OMU>0*"B3$_N0W5V4]CZDJSQ+J3413SFEDFJ(8-S0MEZ#$\^0D.0%@H:HSDVAU<$H`)4`:0`0=E)GGMU[_K&K`+$ M85FQ9@%AHC7Q1TJ(??.5]":&2%D*X*=RBAG3'`)F,L>EY#P/:-_/L0D>-%,H M89.#+('(J9R!0.!`Q4&R7C5XZDU?K#7/FG@:[>SU.$NRR%@U@^"1!!@I>.Z= M9$FPJTZ$9KGRZ)F7=0+B\#BT9@?F967S0[N@'(^.:`^14+]^H1N317OT/,Y. M)8*CD@$?[=(12GE/3?[%"7-\[L3QH@3O30!0#.`Q^3RBF;Y#;#)5EZGF#-/* M`1"+E+UUXORR(ND8F&J*)`5\DM0"IAB6.%J&]:%Z*(*H$LPU[-BU#?N&O=#. M3^FPJ#\[;(@`4JMZM=$F7R>C:2W47[&6#E"?985^",F0UKULYP9PVYG\\PG$ M0KEZ!LE_(B&1_HCC0PTZ8K^`Z4N$H4(T[6CC;!1L]M22X4"VY8(V)[N3[@X$3(7CSN*FDY73*-G^=&<2GK'30P7JDY5<2 M#[RASEC2*S4TN+<^-.[1-@FSI@8K)/<0SID$G!4*8'>I"459J1M@,>H":-J^ M18^)ZTB$HGTP]VD2L!E6/@V4XT6]Y0N6?<"V#-I6.7NQ,%)=Y#SO(JH\?$?" M(^),7EMHZ:^VPOF,M?4H$8WZ%S[$^L4'DCS"/@ M;/>U#UFV`3F4.ODCG2N:)\.&E@$.*-;S9JO*@]U:8P.N\=$,'H[1FFB[-==R M.'&V2UNLK%`0;_L]E=D*8&6D;@"E[0R`!>`&H`&P`)JVLDMDK>KV#5/(U/5) M-5.V'EFN+>O1#NMY+:XN3M>VC4X<'H93!RK?-=-^\E!2[U] M\%"@Q$*FX]/5S9:N23[P;YY;O]HBHK'[B M^%Z<"DX/)Q`60\L2H)W";5E_LL$9[2B60>N@A=`"]R:TSL2NIX56*"O+N\EF M\*H`FT7Y;`<2C[#.:_$=#/V-D@%5)[96!DEA%-)&3FNTS\%+;2VR[QFLY=': M:7^T7D!%[,R6>5BS=;3:<'V-Z8,#+I7D2SMCY632'#-Z\D)IJ%:2MH=XI=)2 M$16+J$I,JVZR0;GR!.#-0N&TY\!RWIR6YEFGQ08R-CV?N4X^K2X00UN7:2[V M]@"U3D!!+6DOB3=T+=0V,*$$7RDU+$V`#4O%:]0^8R&UI\GYZQV64HO&?<(^ M,:.8($9?'@XV$-NS9?B19+V`*UE8E.Q\UCG`R[6'LNL!2C>:C&;MMA6;'-P]?HN';^" M&(6UPT:UGB?12WF'5`W@`3BCNTG%X#'6R*",[>9.\YRUR-CRWEP4%DN-!<%Z M-%FQV-A]+C.6&LM3T`#"1H&OB`9Q[#SWAZJ"U<;21?NY\5Q(@4%7=@N.16`J M6\>Q]H9Z[((6':N.!>XM.K]XWMHB0#PVFYL7(#!:$\FU#DH-Y3XV72N51BUT\RUA-@ZKKRU:-MBQ*)Z M-%6T#L&B8_BV)>#)XRGP/(>:HP1Q:*^Q6*A]5+XF/3SRJ;C@U_1C175)J M`.&=B$W&K8RO5(JA_3`5,M\"KT&>9Y<0@M3+'CLD^!%9>NXXT&[8VWV*JNS_0-2_GJY] M!A.S4,+%K#@1M4O/S/%E-!E]CD=8HF0V[D>9M8+.#I&B(4YSH$;6MVN][,R" M\FBK%$#P+$D1BRAS+'*R/*F30ETW'E'W.&O"Y9?V=`M]Z]N`)"J/CA>;%4[^ M+FNSR02/IBD3OUJ>I0+>_=P`"M05I=:6R^E]^.=1`4NFYERE8\MPSNC6W3B. M$->Y.4#&83OQTH?7794>!(E]F[^@J__PHNF>S']:'MNU'44,+9ET>WG@=#R" M-)FTL]V(9)'4IEEU/-/6&G6L*L$F(W>W/+#2:^16(ON'LU""9CZ0:JA0G.:% M$A^E8DT,[3&IN7C.X\0B>+6[ND5X9>H2ZA?WDPO"'=..?4XP)$[22:C*4^_Z M!P&"^C^CZ+//*9E^=/#2!?.V=I>)5O1=P=>MB_PZ\"T7A:PK/!NNOM";* M2P5X*\\B;*QR=X'!4FOVE="[\WV@Q??->');K5YU5+$[/0OK4O4K3KV M_5Z[!T&M8D0RT9CLDR,L^\IY03].+I;/-&GH32B>^7Z"2SW%H$+7&RN[_>#U M1MD#0D&-9*L7!4#HBW:0_]RY!]TUZ3ATS_OHI9XZ0`6]43Y";WO+".#JW=$B M=`F]-]=;KV4@UYL!9/66>M%ZL$=KX@H`N& M\$I[;D[RP+-3DQD45/:*0WN]$]UGKWQ6VOMGG><>F1RPT5)%8&5OZ+@?[/M9 M*<>]DT?`[EM6$5B,Y"U*)CN`V[Z-8AT3K:L8/%Y^\G:SV<3>K`KT"[@PJ.%6 M)XV[@@')`%9@#!"3M/*59^"$^\@AWL'5UJGI#<$&!=.$PM?XX:V5<>IFS/1R M.A6*M%5^)SRU-]KK38;N:?NZA=X#J[:T*VH,3%[^(>=[5EY.:285YGNS!?GV M>KV(8\CLGEE6+P#\O/"E&'VPHKX/'BQ1CSE$+(C.&2&W(]$8KAW5QJ>#53+V M>'N!#TW\;N:2\@=U%"1B/H&,IU)6[2JRB1NAG>`U$\^'3(`I[G%VY%##V]`* M-!.^GDA#K@302KM`3'V6:$^$4]N7;+\780O1#.P5'0V%K-HD)KOAWGF`=Z@H(S3DIMH9!RR8D>5YCT$+\TQO%MT!3F4]CBT-,^N^E5P?K5KRC'$^;%<2\""YTUT40#(U?OCX14J` M:2VY_3Q(13TVY/$;-YN%(E1MD9-6Z M8>&PKUQ792QW:SO+=54:?FVY5DXL9TO@#YO!J\)>(PFQ.,V]+8(4IUL?/#=Z M8,V"<%G^HQ105;O$M>:^:F.UR5K0:U=46EL0#5328E^PD41=;>I5%YN#S+]B M.1^QA-,O(![@37HH[(/F!8F1[MYMKD.7GSO5Q.?2A?``]T88*`@W8O0#<>B>[6W@X*`+@;X/QL#1`"S>LT4%5V%;K64'+NJ+;UR M=+>U'UTEWCM6I`NN9<5J="^Z*-V(+"&6']O27=>^=-VUP%QX;3&WZHA2/'L. M9P^%_ZBUVW[5]2`#N6?/V. M\JJZ]=0>[9UV8F@G/(]6>GV-1%+$JU'0SFMT%-C"9#VE^<>#[">&_Z208"](9V7[Q<$ENG*$SO!7T\LL#-V>R<4?'Y2 M+)%_047R+]Z7I'D@C?%!)>V_J%WI(683S^$2N`MPS[ZV.=JP[:^TBLC*1&'" M1FL1[(-G,+04'_G_Q>DM$&)\3,9V9)56IHO5!#129A^&B%<@[3\O/.+O=@O;HMUG7RHPA1JXC>&" M">.:?T\6Y7@P&!QT;!1D*"^AX+W:[//6C)>[U M0Q=[\-UWKAL@#T"]U0.(8P&:M-!&KA48XEB;51M@$(6"_].'WEMPZ.@0K`># M@FV:@;^2I*>TP6?ETYO8_6JSP5(6`#;2FL@%C*HB88^]L,AR+SU/^)KHA45Z M!8B]"<5S+_%6/LH4'O0FA4N]O=$+&5(X(@P5'@DC\*;"JE-V'W@/!6P5-O9^ M\(;"6V'2*Z/7)#PV((>N.P.:!$[4X"O M^0Z!*]K^+,.PO_D0'@,W"CNE"L%.)$@O30O(10'4!Z*18X$6[^57H.EX5`=W M^`9X'=$[<#I!"R@1]@-G92F*+>$O,"]X9YO14P0R/]D)"(6\H`S27'(9Q@M4 M.;G!^U<9YP>2I(D)'0GT:?.KIJ,E[MCW01N4Q>%!<=&^:E^S[/S6'HO%'?)A M$P63FSW?GA<747GUC!4&?N>O,KP,HWTR#VNIA<+^?R^ZEF!P<&63\BNMA.%2 M#2&59,@@K<:!]?BP1?7-`>&[L5O<;)NVT`?)Y71.7> M:/&1&-B<`/<7!>#]5>!=83]XX5]SO/&#N,$RW%HSQC-<6 M2IVRO$WUZ]HP[7>@;0_2\):X:F`J`!L8AQ<##NG"@6W`'I*_9%_7IUOW@@)W M:*\*QT7N,%%W#GG>C3\>/0^,6^+!L&F7:%F#\/FQ:E/!J^"/K4[7+VE-C`6? M\9ZH,=U0L&_P#_RRU49N=CNDBE\PGFWR%QSX)'@^0L^/GD".(UQW2]SL966R M8D%`D5I9KN<55%SZQ.4Z1M.?E]?B(*YV:ZL7)*'Q&WSTQ/J) M@V^LFLQ&;&/V/)JDM0R[2_2;I-UW<+:0LL5Y$*:=140I^M))^ M\FB[A%B/(>5W"#AY;!\"5<.2O&`?KT+3YV#WR(E$T0TW:'G#4[S?\(,R[>NHI9*8)$B]0L'BL")PM!?=3`YW M<84$T\A$98'P.,+.87]G[P]FB M]2+%'MPQ\*U8*+N[L/_A#N>*CD=.ZE9XT;L41O26A4G&C5Z*WL\7T$O`X]"6 M3C7&)N)QL*)X(1A97/JY%W^;@DADL8B7#KBFW!2C`2J^?TU/WF-XUH=;M`2K M)0C0A>+,XJ M_LR`_D!IY=26F#O@7/_J3=J_KMH6QJV+8:#AW5WLC?I>V MC3V,T--@;(Y42+S$?0%[=-W`W]IL;B01^SBN[9"6:]W$7<,H,=U6WCHEEB_" M:P>RT$PO,:GR3WS8505[;!6[:.*?'@;4$RGI?$<>&82P'I+7L"VS!CLK1B/V MC.?"BL5'IP\6CRBM9&E^>-N;SU@2R+T1&%S->Z(VA%6V",ABGF>1;TS3-77V M_0*G1DF7+I*V*NN,;?DN,[6\],?+HMR7%4MA_/6-1Z>O1$WW(2FO(EN3773!EF@HF)S+PPR=B`K M0HV_@3PYPZBOE0D42#;><8W`7]GWH=Q3_*>"94H4`F4`(61WI0KOD%E"]D2" M!@F;I-U3W\TR?.L7H20&\`"^9-I^J&K6@9=,%027"8W&_,*ILX.Q?M$)H/+0$"Q/4!ZN!B6Z(&#?JT#7<-QQ[)`^ M:^NY]-_TXP1PL:O/'0$W8QG)G%-G;P<8@>GG!0$W8)NDF>3;X6N0`YP"_@"W M!_.U&]TB<4?W2`S231)KCC6Z)UW/\1.8UO<6GN-FCWW`)-@Y,>C6,XLZ_G5B MB37%!-CQJ<_8F>L^6`(G\;K$(=N1PT-RD_S3=1-/?MO(?EIX(AYTP`MQQ$H* M#,,)YI#J"J:NO%92.\BYZJCQC\E_7$XG\$^O> M`-`,>RXA+ORPA]G M!!F$`6%-)HX/ML@_B"!>MM>)SB4&`HR#!#.`0R'!T/N,.(5J6B.+`$[ M!O2X=CT%;5*VX*AHS&1>(A>I0F4]<4+V53HD']YR2GE#L\3AY+IC9?27N M:2'"0$0@;!V26ES'U`:2_?P&.D4^%>E#O9D-TW8N\)>MU M\;*]*T1=<(,@S7<2GE.>$/N5I;W!@,=3/0`(G3I2`B?%9%I!(V;P2*K>JD^N M&55ZD0'-H_2QH#=8\3KZE@F2,V6K9B#O;SR?32@ZEW^:L=>Z+YH6ZM<2!?"I M];#)FULEK5^&2I(J0?TQ31]YP-[/ANVP*WQ1501V\5;&`>4/GGOU?R?PI?-J M!FNST^617F@G?TR>R&SV)0H)ZKT/PG3DQ;LD.;ENVN$XFXT5@(D#7&S#? M^]X,A4PG[TD`+QGQF\GF4U.2_=EQ)PHR!%`J)1-+%_>?_V"JP"P5IKGUG(AN MD_^`1N977B$P3(@3W"V[+._&O]0,'T$8#,GV;-(J#^6(VTC-6;;[U\2MYS_U&*J M>S$AG@374-N7R)"?U9L4"%Q#"-`/'J'/\=@;E>6]FB5OK\HP0*O9X(>4P!?+ MFM>#%$$'*J[9AJ=K+B;:!5P"L.97Y:]9@M<;I5,@FX?-SH)B,Q[/EY=LMN$M MFY?"5%55HAI.==?;6!<4W&WK;OL#DI/!,^\5"*2O&[ZCT',W.Y*2PXE3*(8U5Z`CP%XH1Y$KS)[3.VA-.Z"N2G,0-9MFQ8GC.[ M;(.7!]9[;]&O-LL'@#H_!O@`REOU9Y'Y49MQJ`)2EGO)!\5(\$I8^"IQIBY' MBK^$[-DI7M-W$;L\+&W"!W2-I3W=K(M92EOK7![F_GR5ZUF@IPDOSQR5X>E" M_;`$'=!V)C0OLQL@Q.[)!2?+NT*E\Z(4E>GOT[CF^*#*S,.@H\Q1>0C!Q3E* M<-^!2\I(7YZY(%>H;.\-%NV0^,ZLYJ.9UK@;\!BD(=.Z#T/5'O`OLUR4%#V# M(^F7(DLY:I;``/J&H>;R5M_%\6.S+X56BLM%?N7XB?/%65R=F=?17[S76P[_ MCY$$O$4!,N+TU6R1H`ZG<P\K M"-^X\^'%+L7&_N?5L^\QD5VC3N3\\.LU:UQOE8JZ&+K&4`'P@/A3#H`X;`L0 8%.J.\X2V@$(AG]!0(!I$%"0$?8$G0*D" ` end From owner-mpi-ptop@CS.UTK.EDU Mon Dec 7 15:09:32 1992 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA05769; Mon, 7 Dec 92 15:09:32 -0500 Received: by CS.UTK.EDU (5.61++/2.8s-UTK) id AA21151; Mon, 7 Dec 92 14:43:08 -0500 Received: from super.super.org by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA21147; Mon, 7 Dec 92 14:43:05 -0500 Received: from b125 (b125.super.org) by super.super.org (4.1/SMI-4.1) id AA12334; Mon, 7 Dec 92 14:43:03 EST Received: by b125 (4.1/SMI-4.1) id AA03830; Mon, 7 Dec 92 14:43:02 EST Date: Mon, 7 Dec 92 14:43:02 EST From: lederman@b125.super.org (Steve Lederman) Message-Id: <9212071943.AA03830@b125> To: mpi-ptop@cs.utk.edu Subject: thoghts on Rolf proposal I am sorry to have started up this discussion and then not followed up quickly but I have been fighting a cold since late last week. (I think I may actually be winning :-) I have read Rolf's Virtual Topologies proposal and think it is a good place to start our discussions. Before I go into the details, I thought I would bring up a few "global" issues that came to my mind during my reading: 1) I think this subgroup is likely to get some resistance to inclusion of topologies into MPI1. I think there are some good points made for their inclusion but anything to strengthen that position would be good. 2) The document has two classes of topologies: general graphs and tori (including meshes and rings). My first question is how useful are the general graphs in practicality. I must show my ignorance on this and say I have never used such a capability. It appears that it could be very useful but it would also appear that it could significantly increase the complexity of implementing the package. Can anyone give some information here? The second question regards what topologies should be given special treatment. Tori are clearly important. What about hypercubes (Intel i860, NCUBE), trees (CM5) and even less common ones like a ring of rings on the KSR. I don't think (though I don't have the proof) that all these structures can, in general, be embedded contention free in a torus. And even if they could, if they are common enough, they would be a nice convenience. The general question is what special topologies should this standard support and what is the basis for the decision. More is nice but also more complex. 3) An argument is made for including topologies at the start since the onion-skin model does not hold because of the calls needed. I think this group should define what information and calls are needed in MPI1 to support the type of topology system being proposed. I think this will help define the complexity of the system. Also, even if the full topology system is not adopted for MPI1, I think a good case can be made for including the necessary calls as is being done for collective communications. 4) A small point: I have found all the references except the two to PARMACS. Are these available electronically? These are just some starting thoughts. What do you all think? Steve From owner-mpi-ptop@CS.UTK.EDU Tue Dec 8 15:42:15 1992 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA27195; Tue, 8 Dec 92 15:42:15 -0500 Received: by CS.UTK.EDU (5.61++/2.8s-UTK) id AA23925; Tue, 8 Dec 92 15:38:55 -0500 Received: from sampson.ccsf.caltech.edu by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA23921; Tue, 8 Dec 92 15:38:53 -0500 Received: from elephant by sampson.ccsf.caltech.edu with SMTP id AA27482 (5.65c/IDA-1.4.4 for mpi-ptop@cs.utk.edu); Tue, 8 Dec 1992 12:38:41 -0800 Received: from lion.parasoft by elephant (4.1/SMI-4.1) id AA16279; Tue, 8 Dec 92 12:38:12 PST Received: by lion.parasoft (4.1/SMI-4.1) id AA00574; Tue, 8 Dec 92 12:38:16 PST Date: Tue, 8 Dec 92 12:38:16 PST From: jwf@lion.Parasoft.COM (Jon Flower) Message-Id: <9212082038.AA00574@lion.parasoft> To: mpi-ptop@cs.utk.edu Subject: Thoughts..... Just a few random thoughts inspired by Steve Lederman's mail... I agree that a general version of a "graph" topology might pose too high constraints on an MPI1 implementation. HOWEVER, this appears to be a problem that research groups have spent and are spending quite a lot of time on. Maybe we could get them to volunteer to contribute some code......???? I have also never used a general graph in my code, but quite a few Express users have asked us for n-ary trees. I think this would suit both the CM5 and KSR structures quite reasonably. I believe that a hypercube is a torus with only two nodes in each dimension, so it IS a subset of the tori proposed by Rolf. I would like to see something to deal with the periodicity (or otherwise) of boundary conditions. In particular I think you want to distinguih between a mesh and a periodic two dimensional grid. Many applications need to distinguish these cases. I would suggest that the function that returns a processor number be allowed to return a "magic" processor number that means "there's noone there". If at all possible, however, I would like this to be a different number from the error condition because I like to see us able to pass it to the MPI1 communication primitives and have them know how to deal with it. If this is possible a lot of messy "if(this), else(that)" logic can be left out of user code. Jon Flower From owner-mpi-ptop@CS.UTK.EDU Wed Dec 9 08:39:32 1992 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA10629; Wed, 9 Dec 92 08:39:32 -0500 Received: by CS.UTK.EDU (5.61++/2.8s-UTK) id AA10654; Wed, 9 Dec 92 08:30:57 -0500 Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA10650; Wed, 9 Dec 92 08:30:53 -0500 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA28376 (5.65c/IDA-1.4.4 for ); Wed, 9 Dec 1992 14:30:25 +0100 Received: by f1neuman.gmd.de id AA15353; Wed, 9 Dec 1992 14:30:01 +0100 Date: Wed, 9 Dec 1992 14:30:01 +0100 From: Rolf.Hempel@gmd.de Message-Id: <9212091330.AA15353@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: Thoughts on Thoughts... Cc: gmap10@f1neuman.gmd.de I would like to comment on the notes of Steve Lederman and Jon Flower. 1. There was the question of what the graph mapping is useful for. I have two answers to this: First, in the PARMACS project we had requests from users (mainly in the CFD field) who had very complicated geometrical domains on which they solve their partial differential equations. The parallelization of such (large) codes usually leads to so-called block-structured domains, in which each block is logically rectangular, but the blocks may be connected in any way. This cannot be embedded into grid or torus structures. On the other hand, the communication is mainly local (between neighboring blocks) also in this case, so a good mapping of the process graph on the hardware helps a lot in minimizing the overall communication. We have implemented the graph mapping on a number of architectures. I did it myself for hypercubes, and the result is near-optimal in most cases. I agree with Jon that there is a lot of theoretical material on mapping strategies. But even if at the moment it seems to be too difficult for some weird architecture, it is still an advantage if the user code contains the logical topology information. If a good mapping cannot be constructed, an implementation may opt to use just random mapping instead. After all, that is what would happen if the user did not specify a topology. Second, there are many different topologies that users might come up with, and you cannot have special functions for all of them. The choice I made in my proposal was to treat the case of n-dim tori separately, but to subsume all the rest in the general graph structure. This makes it possible to specify any topology, and it does not impose all the overhead of the graph definition in the special case of simple tori. I would recommend that if anybody finds it useful to have special functions for other topologies, that he then make a proposal how this should look like, and how it could be integrated into MPI. 2. Jon raised the issue of handling periodicity in a grid/torus topology. I think this is a good idea. The MPI_TORUS function should in this case be renamed MPI_GRID, and an additional argument (vector) could be used to specify periodicity. Do other people agree? Any proposal for concrete syntax? Rolf From owner-mpi-ptop@CS.UTK.EDU Wed Dec 9 14:10:01 1992 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA18505; Wed, 9 Dec 92 14:10:01 -0500 Received: by CS.UTK.EDU (5.61++/2.8s-UTK) id AA17251; Wed, 9 Dec 92 13:49:23 -0500 Received: from fslg8.fsl.noaa.gov by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA17247; Wed, 9 Dec 92 13:49:18 -0500 Received: by fslg8.fsl.noaa.gov (5.57/Ultrix3.0-C) id AA27235; Wed, 9 Dec 92 18:49:06 GMT Received: by macaw.fsl.noaa.gov (4.1/SMI-4.1) id AA01554; Wed, 9 Dec 92 11:48:10 MST Date: Wed, 9 Dec 92 11:48:10 MST From: hender@macaw.fsl.noaa.gov (Tom Henderson) Message-Id: <9212091848.AA01554@macaw.fsl.noaa.gov> To: mpi-ptop@cs.utk.edu Subject: ...and yet more thoughts... I would like to comment on Rolf's comments and generally toss in my two cents worth... 1. I agree that grids and tori should be handled differently. As a user, I would like to see some sort of error message (or at least some condition I can detect and print my own message) any time I try to communicate with a non-existent logical process. Any time an MPI routine can inform me that I'm doing something stupid, I'll avoid hours of debugging. 2. It is possible that a user may want periodicity in some dimensions and not in others. An example might be a global weather model. Often the polar regions are excluded to avoid singularity. This type of model is periodic in the east-west direction and non-periodic in the north-south and vertical directions. Maybe it would be best to have a single general-purpose Cartesian mapping routine and specify periodicity in each dimension instead of having two functions (MPI_TORUS and MPI_GRID). Here's a shot at it: NAME MPI_CART Arrange the processes of a process group in a Cartesian grid structure with n dimensions. SYNOPSIS integer function MPI_CART (gid, ndim, dims, boundaries) integer gid integer ndim integer dims(*) integer boundaries(*) INPUT ARGUMENTS gid the ID number of the process group to be structured. ndim the number of dimensions of the grid. dims the number of process per coordinate direction. boundaries the boundary characteristics in each coordinate direction. DESCRIPTION The processes of the group with ID number gid are arranged in a grid structure. By assigning neighboring positions in the logical process grid to neighboring processors in the hardware the system can optimize the locality of inter-process communications. The number of dimensions is specified by ndim. The number of logical processes in each coordinate direction is specified by array dims. Array boundaries specifies the periodicity in each coordinate dimension. If boundaries(i) is set to MPI_PER then dimension i is treated as periodic. If boundaries(i) is set to MPI_NPER then dimension i is treated as non-periodic. If all bounaries(i) are set to MPI_PER, then the topology is a torus. RETURN VALUE If the calling process is assigned a position in the grid, MPI_CART returns the total number of processes in the grid. If the calling process is not in the grid, the return value is set to 0. A value of -1 is returned if an error occurs. I'm not sure that the constants MPI_PER and MPI_NPER (or MPI_CART) have the best names. Any ideas? 3. In Rolf's parallel Poisson solver example (p. 5), array dims(*) is set up explicitly. It would be nice to have an automatic way of distributing processes among dimensions so recompilation would not be required to run on different numbers of processors. An example of this kind of utility is Parasoft's exgridsplit() routine (KXGDSP()). I'd like to propose a similar routine for MPI-- basically exgridsplit() with a bit more flexibility. This routine should be fairly easy to port since there is no machine-dependent stuff inside. Here's a proposal: NAME MPI_MAKDIM Divide the processes of a process group among n dimensions of a Cartesian grid. SYNOPSIS integer function MPI_MAKDIM (gid, ndim, dims) integer gid integer ndim integer dims(*) INPUT ARGUMENTS gid the ID number of the process group to be structured. ndim the number of dimensions of the grid. dims the number of process desired per coordinate direction. OUTPUT ARGUMENTS dims the number of process per coordinate direction. DESCRIPTION The processes of the group with ID number gid are divided as evenly as possible among ndim dimensions, subject to the constraint that all processes in the group are used. The caller may further constrain the operation of this routine by specifying elements of array dims. If dims(i) is set to a positive number, the routine will fix the number of process in coordinate direction i to dims(i). Processes will then be distributed among dimensions where dims(i) is set to 0, subject to the above constraints. Note that an error will occur if dims(i) is positive and does not evenly divide the number of processes in the group. Negative input values of dims(i) will also cause an error. All elements of dims(*) must be set to 0 for completely automatic process distribution. When the function returns, dims(i) will be set to the number of processes in coordinate direction i. For dims(i) not set by the user, dims(i) will be ordered (MAXIMUM, ..., MINIMUM). Array dim(*) is suitable for use as input to routine MPI_CART(). RETURN VALUE If the routine completes without error, the return value is set to 0. A value of -1 is returned if an error occurs. EXAMPLES integer ndim parameter (ndim=3) integer gid, dims(ndim), ret, dummy, boundaries(ndim) C Set process grid size do 10 i=1,ndim dims(i) = 0 10 continue gid = ALL ret = MPI_MAKDIM(gid, ndim, dims) C Should really do error control here... if (ret .ne. 0) stop C Set up Cartesian process grid for weather model example above. C Dimensions 1, 2, and 3 are East-West, North-South, and vertical. boundaries(1) = MPI_PER boundaries(2) = MPI_NPER boundaries(3) = MPI_NPER ret = MPI_CART(gid, ndim, dims, boundaries) C (etc.) Suppose numprocs is the number of processes in group gid. Here are some examples of how MPI_MAKDIM() would behave in the above example: numprocs = 27 dims(1) = 3 dims(2) = 3 dims(3) = 3 numprocs = 33 dims(1) = 11 dims(2) = 3 dims(3) = 1 numprocs = 512 dims(1) = 8 dims(2) = 8 dims(3) = 8 numprocs = 1024 dims(1) = 16 dims(2) = 8 dims(3) = 8 If dims(*) were set as follows before calling MPI_MAKDIM(): dims(1) = 0 dims(2) = 0 C Restrict number of processes in "vertical" dimension to 3 dims(3) = 3 The same values of numprocs would produce: numprocs = 27 dims(1) = 3 dims(2) = 3 dims(3) = 3 numprocs = 33 dims(1) = 11 dims(2) = 1 dims(3) = 3 numprocs = 512 ERROR (512 not evenly divisible by dims(3)) numprocs = 1024 ERROR (512 not evenly divisible by dims(3)) Any comments? Is the extra flexibility of allowing user-specified constraints on dims(*) worth including? Is MPI_MAKDIM worth having at all? Tom Henderson From owner-mpi-ptop@CS.UTK.EDU Wed Dec 9 15:39:49 1992 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA20328; Wed, 9 Dec 92 15:39:49 -0500 Received: by CS.UTK.EDU (5.61++/2.8s-UTK) id AA20006; Wed, 9 Dec 92 15:32:23 -0500 Received: from super.super.org by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA20000; Wed, 9 Dec 92 15:32:19 -0500 Received: from b125 (b125.super.org) by super.super.org (4.1/SMI-4.1) id AA07951; Wed, 9 Dec 92 15:32:17 EST Received: by b125 (4.1/SMI-4.1) id AA04444; Wed, 9 Dec 92 15:32:16 EST Date: Wed, 9 Dec 92 15:32:16 EST From: lederman@b125.super.org (Steve Lederman) Message-Id: <9212092032.AA04444@b125> To: mpi-ptop@cs.utk.edu Subject: additional thoughts A few thoughts: I have thought about the definition of the torus given by Rolf and I now agree that a hypercube can be embedded into it. I was thinking of a torus in the less general and more geometric sense. Given the fact that you can embed regular graphs into hypercubes this seems like a general enough topology. Some of the latest discussion seems centered on whether or not it is "too general." As Rolf points out, a grid is a subset of a torus where you ignore the extra link. If you had all the hypercube connections, then this would not be a major issue. However, on many real systems you cannot embed the general torus or hypercube so that each logical link gets a contention free physical link. In this case the algorithm for placing the logical processes on the physical nodes needs to decide which links will get one hop contention free physical links and which will get physical links that would require multiple hops and potential contention. Thus, being able to specify important topologies that are simplier than the tori would allow (but not require) the mapping algorithm not to worry about the links that will not be used. Given this idea, I think that grids are a useful concept in addition to tori since they are used with some frequency. Do people agree with this concept? Are there other geometries that should be given special recognition? I still have some reservations about general graphs because they may scare some people from the complexity point of view. However, given Rolf's argument that a random mapping can always be used then I think it is good to include this functionality. It allows for any easy way out for those vendors that wish to do that and still allows vendors to write more intelligent mappers if they can. It also allows for future research to come up with better heuristics that could be incorporated. Along this line of thinking, would it be useful to include weights in MPI_GRAPH? They can always be ignored but might be useful for some mapping algorithms and a new algorithm might be able to effectively use them. A value of zero would be given the default weight of one. The negative I see to this is that it takes up storage and adds some complexity to the user. I think it might be worth it if others feel that it is a feature that users would like and might be used in some mapping algorithms. What do you think? I think it is worth while to support periodic boundary conditions. I don't have a strong opinion yet about the MPI_MAKDIM routine. It seems simple and can be provided to all vendors for inclusion so my inclination is to include it if enough users feel it would help. Steve From owner-mpi-ptop@CS.UTK.EDU Mon Dec 21 03:27:57 1992 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA07780; Mon, 21 Dec 92 03:27:57 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA21765; Mon, 21 Dec 92 03:27:35 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Mon, 21 Dec 1992 08:27:34 GMT Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA21757; Mon, 21 Dec 92 03:27:30 -0500 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA11139 (5.65c/IDA-1.4.4 for ); Mon, 21 Dec 1992 09:26:53 +0100 Received: by f1neuman.gmd.de id AA11153; Mon, 21 Dec 1992 09:26:23 +0100 Date: Mon, 21 Dec 1992 09:26:23 +0100 From: Rolf.Hempel@gmd.de Message-Id: <9212210826.AA11153@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: draft meeting agenda Cc: gmap10@f1neuman.gmd.de Based on the email I saw so far on MPI_PTOP, I gathered the following points for discussion at our subcommittee meeting in Dallas. I tried to order the material in some logical way, but for the agenda of our meeting we can of course rearrange the points as you like. 1. Do we want to add functions for grids as opposed to tori? 2. How to handle periodicity information? Is the specification of periodicity per coordinate direction a good idea? (suggestions for syntax?) Is Tom's proposal for a function MPI_CART okay with everybody? 3. What about including functions for handling general graphs? - is the inclusion of a more general topology than tori or grids a good idea? - is the difficulty of implementation a problem, or does it help to allow a random mapping as a last resort? - is the proposed functionality sufficient? Does anybody feel we should include more user-supplied info (computational weights of nodes, volume of data traffic along edges)? 4. Should we introduce special functions for hypercubes? As Jon pointed out, they could be expressed as tori. Is this sufficient? 5. What about trees? 6. What about more esoteric topologies, like rings of rings? Are there any other structures people would like to see included? 7. What about Tom's function MPI_MAKDIM which automatically distributes the processes along the coordinate directions? 8. Should the pid inquiry functions return some magic number, if there is not such a process (as Jon suggested)? 9. What can we do to convince the general MPI committee that virtual topologies are to be included in the standard? 10. Do we propose to include the PTOP functions into MPI-1? If not, how can they be based on MPI-1? Please check whether all the points you would like to discuss at our meeting are somehow covered by this list. If not, send a note. I will send out an updated list on January 4th. I wish everybody on the committee a Merry Christmas and a Happy New Year! Rolf From owner-mpi-ptop@CS.UTK.EDU Thu Dec 31 22:14:03 1992 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA14662; Thu, 31 Dec 92 22:14:03 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA12437; Thu, 31 Dec 92 22:13:48 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Fri, 01 Jan 1993 03:13:47 GMT Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from pnlg.pnl.gov by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA12429; Thu, 31 Dec 92 22:13:45 -0500 Received: from carbon.pnl.gov (130.20.65.121) by pnlg.pnl.gov; Thu, 31 Dec 92 19:09 PST Received: from fermi.pnl.gov by carbon.pnl.gov (4.1/SMI-4.1) id AA21172; Thu, 31 Dec 92 19:08:30 PST Received: by fermi.pnl.gov (4.1/SMI-4.1) id AA11537; Thu, 31 Dec 92 19:08:29 PST Date: Thu, 31 Dec 92 19:08:29 PST From: d3g681@fermi.pnl.gov To: littlefield@fermi.pnl.gov, mpi-collcomm@cs.utk.edu, mpi-ptop@cs.utk.edu Message-Id: <9301010308.AA11537@fermi.pnl.gov> X-Envelope-To: mpi-ptop@cs.utk.edu, mpi-collcomm@cs.utk.edu Posted to mpi-collcomm and mpi-ptop. I have just taken the archived discussion from netlib@ornl and not found anything more recent than december 15 (collcomm) and 21 (ptop). Since I asked for my name to be on the mailing lists and have seen nothing I assume that things have been quiet since then. Al Geist's proposal (Dec. 5) for collective communication and the reasoning behind it seems to provide a resonable starting point for the discussion of interface and functionality. I have only a few minor comments in this regard, but given that the efficiency of collective communications is critically sensitive to hardware topology it *must* be essential to more closely integrate the definition of process groups with topology. I restrict my comments here to this subject. For example, on the Touchstone Delta efficient sub-group global-ops would suggest that process groups map as best possible to square sub-meshes, on the iPSC as sub-cubes, on the KSR as sub-rings. Currently, if one's interest is in performing efficient collective communication in subgroups, there is no way of performing this mapping in a portable way. In this instance one might want something that functions along these lines Create NG process groups with P(0), P(1), ..., P(NG-1) processes in each group and assign each process to one of these groups so that collective communication within each (and perhaps also between all) subgroup is optimized. Such a mapping might also be readily accomodated as a sub-partitioning of an existing process group, with the default being ALL. I could envisage writing, for instance, a fast-multipole integration using this functionality. Comments? Robert J. Harrison Mail Stop K1-90 tel: 509-375-2037 Battelle Pacific Northwest Laboratory fax: 509-375-6631 P.O. Box 999, Richland WA 99352 E-mail: rj_harrison@pnl.gov From owner-mpi-ptop@CS.UTK.EDU Fri Jan 1 11:53:58 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA16856; Fri, 1 Jan 93 11:53:58 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA13825; Fri, 1 Jan 93 11:53:48 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Fri, 01 Jan 1993 16:53:47 GMT Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from msr.EPM.ORNL.GOV by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA13817; Fri, 1 Jan 93 11:53:46 -0500 Received: by msr.EPM.ORNL.GOV (5.61/1.34) id AA04566; Fri, 1 Jan 93 11:53:35 -0500 Date: Fri, 1 Jan 93 11:53:35 -0500 From: geist@msr.EPM.ORNL.GOV (Al Geist) Message-Id: <9301011653.AA04566@msr.EPM.ORNL.GOV> To: d3g681@fermi.pnl.gov, littlefield@fermi.pnl.gov, mpi-collcomm@cs.utk.edu, mpi-ptop@cs.utk.edu Subject: Re: groups and topology. >I have only a few >minor comments in this regard, but given that the efficiency of >collective communications is critically sensitive to hardware topology >it *must* be essential to more closely integrate the definition of >process groups with topology. >Currently, if one's interest is in performing efficient collective >communication in subgroups, there is no way of performing this mapping >in a portable way. It is critical that MPI be portable even if efficiency suffers. Portability is primary reason for having a standard. Efficiency is important and tightly coupled to the implementation on a given vendor's machine. My feeling is that our MPI work should specify the functionality at the user level and not dictate how MPI is implemented underneath. Mapping is the key word in integrating topology and groups, and mapping is not defined (so far) in MPI. It is related to the spawning and placement of tasks. I can envision some implementations allowing tasks to migrate to improve load balance and fault tolerance. This greatly compounds the mapping problem, but I don't think MPI should exclude such implementations. The hope would be that vendors would supply MPI implementations that map process number to node number in a way that their collective routines would be efficient with default ALL group AND that the vendor's mapping would be documented so that a user could specify subgroups that could exploit this same efficiency. Al Geist From owner-mpi-ptop@CS.UTK.EDU Mon Jan 4 06:22:14 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA06430; Mon, 4 Jan 93 06:22:14 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA03279; Mon, 4 Jan 93 06:21:55 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Mon, 04 Jan 1993 11:21:50 GMT Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA03271; Mon, 4 Jan 93 06:21:47 -0500 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA17137 (5.65c/IDA-1.4.4 for ); Mon, 4 Jan 1993 12:21:14 +0100 Received: by f1neuman.gmd.de id AA14884; Mon, 4 Jan 1993 12:20:29 +0100 Date: Mon, 4 Jan 1993 12:20:29 +0100 From: Rolf.Hempel@gmd.de Message-Id: <9301041120.AA14884@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: Agenda for subcommittee meeting Cc: gmap10@f1neuman.gmd.de Dear All, in my message of Dec. 21st I asked you for comments on the agenda I proposed for our working group meeting in Dallas. Since I did not get any response, I guess that you don't disagree too much. I am looking forward too seeing you all in Dallas. Rolf From owner-mpi-ptop@CS.UTK.EDU Tue Jan 19 06:28:43 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA01032; Tue, 19 Jan 93 06:28:43 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA08439; Tue, 19 Jan 93 06:28:14 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Tue, 19 Jan 1993 06:28:13 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA08431; Tue, 19 Jan 93 06:28:08 -0500 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA20278 (5.65c/IDA-1.4.4 for ); Tue, 19 Jan 1993 12:27:30 +0100 Received: by f1neuman.gmd.de id AA13250; Tue, 19 Jan 1993 12:26:35 +0100 Date: Tue, 19 Jan 1993 12:26:35 +0100 From: Rolf.Hempel@gmd.de Message-Id: <9301191126.AA13250@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: next activities Cc: gmap10@f1neuman.gmd.de Dear All, this time it's my turn to fight against a bad cold, but I think I'll make it. We should plan the work our group has to do before the next Dallas meeting. At the last meeting some points were brought up in discussions, which we should invest some thoughts into: 1. If more than one topology is assigned to a set of processes (i.e., there are two groups with identical membership, and a topology is assigned to both groups), then it might be useful to do the placement of processes (mapping) only according to one topology, and then state how the logical process coordinates of the second topology should correspond to the logical coordinates of the first one. This was the idea of Marc Snir, whereas I favoured the following point of view: for both topologies the assignment of processes to positions within the topology should be based on that topology alone. As a consequence the mapping can be optimized for both topologies. On the other hand, the roles a process plays in both topologies are not coordinated. A third way could be to specify both topologies together, including their alignment between each other. It would then be the (nontrivial) task of the system to assign the processes in a clever way. Also, we would have to think about how the specification in this case could look like. Any ideas? 2. As it stands now, processes in a group get a number 0 - n-1 following an arbitrary scheme. Once a topology is in place, it would be a good idea to base the numbering on the topology (e.g. run through a grid row/column - wise). I see two alternatives how this can be done: a) The numbering is defined only after the topology definition. This means that there is no group rank without a topology, which contradicts our goal to have a topology as an optional attribute to a group. b) The group rank of a process changes with the topology definition. This might cause some confusion among applications programmers. For example, global gather operations before and after the topology definition would lead to different ordering of data in the buffer. On the other hand, there seems to be no conflict with other MPI features. What do you think? 3. Should there be any special topology definition / look up functions for trees? How should they look like? Since Jon brought this up at our meeting, I'm most confident to get some input from him in this matter. Does anybody have brilliant suggestions? 4. Bill Gropp claimed that topology mapping could be done in a portable way if there are some inquiry functions about the hardware topology. Are we able to define a set of such inquiry functions which meet the following requirements: a) they must provide all information necessary for a portable mapping tool for arbitrary application topologies. b) they must be applicable to any weird hardware interconnect topology. Frankly, I have difficulties with this, but perhaps somebody has a good idea. 5. - n. Did I forget anything? I will work the results of the last meeting into the document on virtual topologies until end of next week and send it out to everybody. Rolf From owner-mpi-ptop@CS.UTK.EDU Tue Jan 19 13:53:39 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA22518; Tue, 19 Jan 93 13:53:39 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA27886; Tue, 19 Jan 93 13:53:14 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Tue, 19 Jan 1993 13:53:13 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from fslg8.fsl.noaa.gov by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA27878; Tue, 19 Jan 93 13:53:11 -0500 Received: by fslg8.fsl.noaa.gov (5.57/Ultrix3.0-C) id AA26554; Tue, 19 Jan 93 18:53:07 GMT Received: by macaw.fsl.noaa.gov (4.1/SMI-4.1) id AA19368; Tue, 19 Jan 93 11:52:05 MST Date: Tue, 19 Jan 93 11:52:05 MST From: hender@macaw.fsl.noaa.gov (Tom Henderson) Message-Id: <9301191852.AA19368@macaw.fsl.noaa.gov> To: mpi-ptop@cs.utk.edu Subject: Re: next activities > 1. If more than one topology is assigned to a set of processes (i.e., > there are two groups with identical membership, and a topology is > assigned to both groups), then it might be useful to do the placement > of processes (mapping) only according to one topology, and then state > how the logical process coordinates of the second topology should > correspond to the logical coordinates of the first one. > > This was the idea of Marc Snir, whereas I favoured the following > point of view: for both topologies the assignment of processes > to positions within the topology should be based on that topology > alone. As a consequence the mapping can be optimized for both > topologies. On the other hand, the roles a process plays in both > topologies are not coordinated. > > A third way could be to specify both topologies together, including > their alignment between each other. It would then be the (nontrivial) > task of the system to assign the processes in a clever way. Also, > we would have to think about how the specification in this case > could look like. > > Any ideas? I'm kind of leaning towards the second option, though I might be swayed if I could see an example of code that could take advantage of Marc's idea. Any good examples out there? At the end of the last meeting, a suggestion was made that group creation and topology creation be married. A group would always have a topology associated with it at creation. A default topology of "linear" or "ring" might be supported. If this idea is adopted, then it seems to me that each group would have one and only one topology. There is no reason that multiple groups could not have the same membership. Each of these groups could have different topologies. I think this ends up being the second idea. > 2. As it stands now, processes in a group get a number 0 - n-1 following > an arbitrary scheme. Once a topology is in place, it would be a good > idea to base the numbering on the topology (e.g. run through a grid > row/column - wise). I see two alternatives how this can be done: > a) The numbering is defined only after the topology definition. > This means that there is no group rank without a topology, which > contradicts our goal to have a topology as an optional attribute > to a group. > b) The group rank of a process changes with the topology definition. > This might cause some confusion among applications programmers. > For example, global gather operations before and after the > topology definition would lead to different ordering of data in > the buffer. On the other hand, there seems to be no conflict with > other MPI features. > What do you think? I would prefer to see topology inquiry routines that return the "pid" of a process specified by logical coordinates. Inverse routines would also have to be provided. I think that this is safer than either "a" or "b". If topology and group creation are married, then option "a" is automatic. (Actually, this is a good argument for combining group and topology.) I don't have any brilliant (or even reasonably sensible) ideas about items 3 or 4 yet. > ... > > Rolf > Hope you get over your cold soon! Tom Henderson hender@fsl.noaa.gov From owner-mpi-ptop@CS.UTK.EDU Wed Jan 20 05:13:45 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA11399; Wed, 20 Jan 93 05:13:45 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA02944; Wed, 20 Jan 93 05:13:23 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Wed, 20 Jan 1993 05:13:22 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA02935; Wed, 20 Jan 93 05:13:17 -0500 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA16345 (5.65c/IDA-1.4.4 for ); Wed, 20 Jan 1993 11:12:40 +0100 Received: by f1neuman.gmd.de id AA16629; Wed, 20 Jan 1993 11:11:44 +0100 Date: Wed, 20 Jan 1993 11:11:44 +0100 From: Rolf.Hempel@gmd.de Message-Id: <9301201011.AA16629@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: topologies and groups Cc: gmap10@f1neuman.gmd.de I'm not sure whether we really want to have a default topology. I see some problems arising from that. 1. In some applications it might be completely unimportant how the processes are mapped. A default topology then adds an unnecessary overhead in group creation. An arbitrary numbering would be much better in this case. 2. MPI functions for group creation and topology assignment cannot be separated. Otherwise the system does not know whether it should apply the default topology (another topology could be defined later by the user). I think that especially those people who are not completely convinced of the value of the topology stuff would not like to see those two concepts integrated into the same functions. I favour the following strategy: A group is created without topology (how this is done is a separate issue which we have to discuss with the CollComm people). An arbitrary numbering is chosen by the system. Collective communications at this point follows the arbitrary numbering. A topology definition call is a synchronization point for all processes of the group. The numbering of processes is changed according to the topology. Consistency of numbering is guaranteed because of the synchronization. If applications people want to avoid the different numbering schemes to be used at different phases of their application, they just have to apply the topology before they do any collective communication. Any opinions? There will be some functions which combine the creation of groups and topologies, but they can be optional. An example for this is the subgroup creation along coordinate lines in a cartesian topology. We have to talk to the CollComm people about those functions. Does anybody have a good idea for another topology based process group creation function? Rolf From owner-mpi-ptop@CS.UTK.EDU Wed Jan 20 09:56:37 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA10131; Wed, 20 Jan 93 09:56:37 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA19145; Wed, 20 Jan 93 09:55:55 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Wed, 20 Jan 1993 09:55:54 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from fslg8.fsl.noaa.gov by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA19137; Wed, 20 Jan 93 09:55:52 -0500 Received: by fslg8.fsl.noaa.gov (5.57/Ultrix3.0-C) id AA28601; Wed, 20 Jan 93 14:55:47 GMT Received: by macaw.fsl.noaa.gov (4.1/SMI-4.1) id AA21143; Wed, 20 Jan 93 07:54:46 MST Date: Wed, 20 Jan 93 07:54:46 MST From: hender@macaw.fsl.noaa.gov (Tom Henderson) Message-Id: <9301201454.AA21143@macaw.fsl.noaa.gov> To: mpi-ptop@cs.utk.edu Subject: Re: topologies and groups > ... > 2. MPI functions for group creation and topology assignment cannot be > separated. Otherwise the system does not know whether it should > apply the default topology (another topology could be defined later > by the user). I think that especially those people who are not > completely convinced of the value of the topology stuff would not > like to see those two concepts integrated into the same functions. > ... I'll buy the second point. People who don't need topologies (or think they don't) shouldn't be forced to learn about them. (Just like people who don't ever do collective communication should never have to learn about groups...) > ... > A topology definition call is a synchronization point for all processes > of the group. The numbering of processes is changed according to the > topology. Consistency of numbering is guaranteed because of the > synchronization. If applications people want to avoid the different > numbering schemes to be used at different phases of their application, > they just have to apply the topology before they do any collective > communication. Any opinions? > ... > > Rolf > I'm still uneasy about topology creation routines changing pid's. It seems to me that this would be a likely place for users to make errors. The problem is that I can imagine situations where unsafe code might work on the machine it was developed on (due to "good" luck) and fail when ported to another machine (where the pid renumbering works differently). I know it is not possible to make everything idiot-proof but we may hear this objection when the idea is proposed to the whole committee. Tom From owner-mpi-ptop@CS.UTK.EDU Tue Jan 26 06:23:47 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA27461; Tue, 26 Jan 93 06:23:47 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA26404; Tue, 26 Jan 93 06:23:22 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Tue, 26 Jan 1993 06:23:20 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA26396; Tue, 26 Jan 93 06:23:01 -0500 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA17536 (5.65c/IDA-1.4.4 for ); Tue, 26 Jan 1993 12:22:13 +0100 Received: by f1neuman.gmd.de id AA13867; Tue, 26 Jan 1993 12:21:18 +0100 Date: Tue, 26 Jan 1993 12:21:18 +0100 From: Rolf.Hempel@gmd.de Message-Id: <9301261121.AA13867@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: Updated working document Cc: gmap10@f1neuman.gmd.de As I promised last week, I have updated the document on process topologies in MPI. With this mail I send you the PostScript file which I compressed and then uuencoded. Any corrections/additions are welcome. Rolf ------------------------ Cut here -------------------------------------- begin 664 PTOP.ps.Z M'YV0)4)`F=(B")DW8LJTD.$"AH(2)8;(*1.&SALY.D"0L9,&SAP6(.R4D3,G MS1LW(&JXF)&C!@@40U*`B)$#AXT6.6"`D!*&3)HQ8=B`*)*G#(@I;\S0N1-F MXL,25-+085,F(QDY:43*<;$QS5,H8/&IYLS:O%DU*G3 M1@P9('#0D/&T2-LA;]JT*>.&SIRT9$&E M#!8B/^F`D`%#I\\QHLF4,0.B\^?0(!(J1OG""8@7JEFS>=-3X^K;0O:(64S& MMYD^ME],N6UY#!H%NWOGOHUESW(G?83!PYI1')A84 MF_39-\.9"O#91GAU!&5&&660L<<<=%14QQRGB?;"H6XDRL:BC8(`IE%P:(@F M$`W"D<983(Y0ZJGL88&F$4U4M(4"""IH8*T+)F>$$$(,:"M=721'AAE,SL%: M@(``:T51:F1&"Y&FP0::PP;HX6W&*H#%;)$09\('S;%'AY;":$5+%Z%#)P=5A+QA!H-+4' MS$T)T3(;JLF!W&U^->B6NDD4$2UAQ9%1!QQ,FO'&&RY*IG1A??`U\#Q!APK=GBT<8NM%G#3=#P-H1";.==" M&H>*M8?>9%04AD9OTT&P42]HO-@9AA?,XM_?@2!6V)&+/1,(/:$7X7%XOZ#W M'6F000<:?Z/10N"1$LXV56^-GI*+=8@AN61HZHU&&6F<@08=I9\NN.H@L,XX M&B#0`+L8"DQ>N^EX)$7L77H"_KO1JQ,VO,''SR[:'+$OWT(>SO>.^N#40V>] MZWAQ+_OD,\EPP_'>DX&'^-._+;SK!:JO?>V1X[<9S7*PFB&,9[1T%."C+$GO8MYS)Y MFF<&UN@-?"VTFNENE[O=96^%SI-A"T`GNN=D3H`%^80BY@C`QFZT#N^A:4,$>I;&0XU MAS6`8"*0N@@:IT0$)HUA#"MS8.,0I[C6'>Y'_>G"YMKV0)"Q[XY6BQS8"/F" MD]$A91N4S`4MMYQR%>AQ#G3@_8@7`P7HSX$3;$L%.RC)DG=XLQ5QO>(!*+H*D,O.PB59IRE&G"T2(3<5$Q[P"',X+` MEW(`)CB'6(!H]]5%=\F&2'(14AJI9=*M4:=:WK*``-%EA#V>80S5;TYT6@-&M!>+. M&[G:++!VU44PLL_EW`,@HRRU(GN0*E73>$8VKB%YYY0CFL*PAVUV\PW8*@,; M4H"FN&O:P M6S2M80^W11.]V&9;$-P!36U(FX$&]=TI,9.[>"GO;=X@7C/=H:Q3BH-X"Z3> M%RR);>E%4X#81EXT\8YMQJOO'9#[1M!Z`QC#.L4UAP6-!YQQ!0+2^M*4QCXD,'13@&M"@ MAE^S<5-'9[#CC@*AH;>*4UI6C4:!J39,PP4780;#$$OH/9!#7)P2R1'^<^M)@VK M\0R#9N>YV7\NP@S^3(,;/-K@2W[TI_L-Y84_^=<+EW(,I"P#);<9!X7N]KEC M4/`USX#C/L@!O><,Y2CO.M.[[C*75V[E7P.[Y/PNN;"A//-&]]KF#8\UH'/` M:"8CG-^%CG6A!P[HHDMYTW:6,L*5GO0FOYSA,?\TLX%]GR?GF^%79_+`]0V# M&AS[SCCX.L,13JLB1WG-S<:!M.WM;IMS_3YYCD&LZX3O0KO/SJX6PI/;#O6^ M.SS,6NXDQ2V.Z:3CN]$=S\'':[!V60NYZ'NW5W.X]W3Q1WDNAW.RUW\W M=VBT5G(@"($L%WQ>%F@YUVLV4&B_-@,U5R"(`[ MB&LI-X)`J``I1X$PQVK"QF1V\7+A1W505P2D(75[IFKO)GM$`(`OJ'&TMFD^ M>&AWPH4S.`1L)H.@AX:2]FD'YW:(<.<6AY^&4&?,N&HE=P-T*'L?EW@?9P-Q>&;" M1@,V$(O;!VZQ%GY?2&E&J(^E!W<&&'?$]H^EEV5"AV]_)H<':9!%=WE]UG1% M)VP#IW,40HC^1X37!V599H$:R8,AX[?AW&E88J`5X(PMY$7F8$<2'D=Z9,.<8>REG)R%XP>AV;%:#S' MV':.EG[GIF@\!GR>R&Y4=G!UEF0:1W_"1BL1IW2?9@-?&8_JA@-.27^'9Y3( MQWBF-FE/]X=`J7-<*(3]YF0-*6:`1WX9R(<5"6B*9H9%N)GGB)$8^)D7R&^D81<$ MUV^QMF<`1VY.^&?;)X]*UH6PUF;L&`-!H'RDL7]?V6BSQF3T*)JQEF>VR&\* MF&<,*)$<^&I3IG.K]W*')FAR"72L5V6A.&EXN7!V,7,NJ),\*9$S>'`,.&5, M]FN*)FGH%H#T!X!M1@3"F'QK>7J*F&GC1F9A9V;T]VCO!W.*"&7JAF7D)GLP M1YYL*6OX"7-C69/G=J"O^6229A=8B)KVV65U0BNSJ99#$61X1!H@D!/28F^@ MN9,P=Y[4!Z+3)Z(EEV4H1WTG.GU@>7.)]I$RMVS4YVH6:&:W]J$8MW!!J6OP M"7,\EJ*XUFA7V9EF^9]-QJ)KR*(]9X&>!J,N1W]+"H"/MJ-#:6@XFFOU^:.R M%J2>F*4J6'(?67,*`*-SR6Q.]F]BZFEYEHGUMV9.J'8VX!+'2&:;2*5EQFR" M1@2"YG!Y5J.#^&_#=W]:&GI2"FR#BF><]I]\&FF)NFYBZJ4S*J8;V*@-]ZB4 M.GV+>G:(BJ/K5J@KJ*EG9@3/.6D?N9O8V6<1-ZD]F6ET>8Y82'_T=P-$D`-. MZ(XV@(YSMJ22BJLN6J.'2F7/IZ5\^I%\.JS8&7$SIP!2RJ=7BJE9>G`SMZS* MBH71FJ2R)ZW/:JTY.J3,"J34RJ5+=JW=RJW?6JW.2J[CVJQ+I@#%ZJNX9JSL M>J[]UFLUNJS,VJMSN:2]AJM@*@.QN*:R&H-KIZMFNJ1'6+#G"GU\:K"1!J]* MAYU!R:=*=FM0.:X3BZD`>FLUIV8XR&F2RK$9>Y+GF&=Y9GMY=K$E2YTGRVO; MJK+*FG3/ZK*R![,A.*PP"W$R"Y^Z^F0YVX_^ZH31MI9+IG'YB95UZJA)AJ?M MBI&-"K$S-ZJ_NFO05Z@0J:E'J*FV9H%7RY\+FX/5)ZE&2WU>BZI@.[5ERK5$ M"Z-JMH&\^K%!&[)4ZK98";?T&JSR.G/$RI]2FJMQ)ZFJ6F<6R&E_)FEJ"JO_ MF@,^`(SU]J2*ZVGW"H[J^J[;>K<,:[>4"[F2&[F5>Z[CE[F8:[F<.ZSZ2F6A M.[F>2[*>>[JDF[J=J[J7NVF?^[JHN[B=QKA/*@-MUK-S$K`[BXX*N[;GZKOU MVKO"^[O#&[SCNKG$F[S&F[/-!J3V!W*F]J1CVKA-)J;*6@2F.Z7PFFO;:K)` M>:XC6*,`RJPL>FN.MH&8RKWCJJEG1Z_?ZJ/?NJ)%6&R+RFGAIZF'BJSXN[=# MR&PW5Z,=J[#@BZ/S"JQ-*Z^/6Z2H&[F?]J1]2[T/;+NORJ:Y>P,UT'Q+NFL9 M#+<<#+D>#*\?O++'.[^6.J.7>L(FG,(EO,)%F,`JNZWAVJT2:7X'VZV$")'E MVF_E>J5ABJMVIJOO%P03+*OTI[L6J*HXP+BL>I7VVK:1IJ6@]FE9!I8I1P36 M"<56G&A\UVB3J,7#U\61YH*X)J(KB)UD#*)3?,;62<8Z3)_6V<:\QL9,.8-O M/,=RC&]TS)]0B6IL+,;+^,;GJ6J`K*!`FFA--I]LK,/X]L911@1RF+`.AUK,S\+,\4+;5J[-EMLI.O+7R M:"\M)-G,`ZG\`FI&N+,M* MO+42C+N&"[1.FG+2"Z1TG,,`7(RZH&W<0#+=`")]`M>94$V6W\?)58N*/`V\]:6L]`&J20 M67^QZH0E&;T$N\J@NJK,:KU3>[!,_+3,.J0%7'F_RM/>NM-1EL!76L`W"XYV MJM`RS;%+J\(DW-2>B<(LW,$@/-4B+,(F+;V_%GI#[(0TT'@G3<_M3-+X[+]O M^]'WW*R#&J:(Z:@6W6X?"]%+'-%P+;IS_6L33=<17=$?V]8`O='A]ZF?FL2` M^L/&9R>.6V?>;,_T:M).BFK7I[Y[!W/MK-!%Z+6BJLDK*W*^[`U_:Z#6J,^RMH&3*6[YMJQG=H#G*/&RZZF_="M3->:?6MB M[+8D.M/]VV2A.M/_2];ONJM'G=QL"\XS[=BW-GZ^-\`V0;U.RHU*;+_H=KOV MYSYQJ'$_[&1WTLG?J<-5AG3!E\0^N'".QMY)3,__*ZRJW+>2K,K2J]0>&::Q M#<6[^H%#T&L#!^#_C6M#4*IV^]Y3!J.GUVY.QII.FM4,^MG%MW9*[6F(B:NN MYM/NUK`VE[6\YN$%?J[%-L(VT*TE?JXG+LPA#-5.O>(F[,(N'M4QWN)4S>)/ M77-]&]+X"J2SMV9$8+NA;75E^Z2-#<)URW]4!N-'+GU&[L%+'J")^N0OO+*, MAKQ,3N5.GN5-ON58;N39"^4U)^5BCFLO6ZTQ:^:ZYZ6/I@#CZ]LLZJ103*4W M-[@4K'8N+:9(C*9T*K,A:.5I3F:=^;:W[*/7NFO#]Y$WUZKP/91J"Y%-K;[$ MG-GDC)5V9H%7^N8>S:+QVDDIUV3WU^EGI[[OBZ)&,+Y4:NI8.:/B.J)JW;6@ M?LC5=WN!#F4R4`.O^FQM&MIDF\3S'=-)5Z56.>L^2IZF?>B`*+;,Z,20SM&W M=Z+C.Z2:OFP!.*G)+KJB3K0A&K:-UNI@J^D*W.U=Z^W+#.[D7J+?SMGBKMD- MI^EP;NEZOI*V7G^X#K`X$&=$4*$14QPR.R$2/`&2'0F!W'?K/'N-HEPV)X]^+T]>69G]L.M M]HT\Q\1*?-W\V<3VJO2ZW61'3_2NAG>M5HX_WY,]690%%ZO`J'8*8&JMYH(% M;029=IY,=I+N5F>2AG<&AR=P2`-*B60`C>?9S8T.270TC/$OMW`I&YK:B?73 M;O,FWW!6*(+_N^:?5_!+ENA0&Y`VOZ=Z_M)*'7=WVA^YKK(YZ8%,F>)3;16*:TSZ.BA:9'52]=9VLJ'>G1+ MMWN^?XJ?!XYY?+9"OBM&?9X<%XRQ^G%JEP.VBF2OS[&Y?Y5BC'!#AY,>ZO<*8LXY\>+G-2[K:UWH?RK'W(C9V_TZJ,\[(,B^(/R'_^&)G`*MROU M3?V!771,;/*65^E]#H?%G$(C``O@S/)_)B_C])G]!^<:F/'K,QFMT=@Z&:4` M+E_/@WX$K%7-M\EC=.Z>!XPYO\DG[;@1Z&G2T=_Z?@?I_1PD03/^@I3T\U:- M)A9MO?37G@`54OMAH&SV"2J<(P#_GY:Y/`?0Z/BDYP.R3IK%"U(LH51MH&E7 MY$Z:QNE;>TJ2(1H#YX.H(*WY.7LG3P&I'L9H1-NJ(S[R+OTMGWI#@I`1E?E? M'^C5"2#T5=GRE7R[>DP0#LK!G[3MYF`'HH,!2'IM,.E5E-33$.`Y<.@^7+!C M1(AZCB$42J*OP2F]!O>6CAU?^D-9S1%*PD9("??2)+2$4$F]&1S&==FPTG-B M::-FZ`6LFJ.WN!]/TG[9+Z-)G1>(O535R_DY'!#/^!^!TVO:S:XY:*--@WVD M9Y.O4A[?ZS*8K3E%H[F$F'A=,6QP]HOX("8+:.=X0">L7G;+:_T]W/-KIJ$U M-(6@Z1JF0E-8#;$A*CR%T9`12CX+A_J*WIK!?!0NQ#6MF*9W8I],8$))$8$A/82#R)(1$E M`B[A4Z8NVTKT:,!H(L(`&H@!1]NOZF'UKXT-'8=$BX9@[N%#0/$GYIZ`%W@J MSP*,1N,'`K8Q..>1KI4Z0DL6T=2HF=^V;#0B.=QTS,8C@J.@I(H&H$-(Z&9W[;-\R&WO#O`:ZF*+J&CZTK`L?'(GX]9UAG M;HV3F41[Q_A1O[YHJB"-6QR&!([^P;_%B`A9$&,L@T!HOYVI[*;*B$_!>386 M4>[0@/W#:2P/9=1]/^SW=<6NN,7(S+7Z2)GF"FTMF>056V,V^F&AD7K=*U?4 M]#0CX8J*T&_^/4;FHQMUF^?3.#L.,O9&X:@8&^,/*F[=)RPJQ['H#W.?&SPY MS^?C.+]GTYTH'+<)C(#1(=@OW7>"L).LHD*'CIX-(6$(%Q&X5+T[.I20CQMZRTA-HDD]>6EF8)LI2SMNQ0=''*WEEA0-I&*[11**"\P?!D`U2`/@1 MWQD!UE`#!L64W%`BR/=8H)9D$VC/J@D[P`Y#>AL1E&0ZN/-P2*$1EL5QN)H\@:N02"&!*6DI"69];)S%D9FCI4S01E MDFHDHI[AHY!\DGGJF>'OY-2E$-0RZU(+?RT/5C`_]FEY#.:]LY^J67Y2F2HSRSP@&=!N364N(`)%I$K.T$*!1GLCVM\YN% M)B2%!WDG9Q*>QRI`Z1H`Y#M+SOCJ/S\H>!HH`'07\"(,.)TV".RUI%LT<+IG M:=HSEB8"RK4$%*-&5*P$==?NVJ&Z=*8IM5?NY):?DOJ@NESI/QNHUH*=6$IZ%E#E:4!U M)P)MGL%S`S%0&^5`+Z@'%9X4-(0RHR4(0%&4`NV?T\?;O3K:1,HXS>EL:1P* M5)V9KJDL^4/AY(X_(>X[0[)LR@Q(Q8KP@&D2I412<1M:(&DTG;10G M7^,.-8N9)Q5Y/[@$@*7A/'9+F;?F!)&`M$M< MIN'%O$H(EQPG!)IY1')5^1O\F0.V:.K44E3)0$F<-/>>ML]?TC-_M,T4G;V@ M9>11RRDZXTL'>5!))2O;7!$:0I`4TBR=I&-GD$[K:6#3\],@FWC$@=C/X"QT M6/15_4$[:NOV#XQ<,LGNSQC.%P86M^6*B2F M9/=$5:V9`3_G)=5*N!@QI9#3N1,X#R^1)[RCGUKF"_HX]4<(U%%+F6;ZYN!4 M42=T^F"V#S6C4&7++-Y/#/K#XXI^L9*1&(!;=@"`PB6*1 M#ZA59+`!-CX,A+"RE0`T0'!)YM5"6@E&=\\MG)`^-/[\''76=]*B[-D\!HB] MS=(XZ7_D$':"1@E07L5)=RGXJBIW`DK!U#M=2'XICCP2(\Q2KVJE\2L%<(-& MER8C>3MUYJ@;__.(.">M+*PZ%3'&)TM::4Q>)(6DU6ZZC39-)5D)'_=R3X4O M/Q&D-%7XXJG@>T^]!K0F5L&G;B2D7%V>^K2?IBO"!R/]JD*4<$:@KP;'J-6` MGD]M79VETJ(&)=W:GP0?;SVMJO6QMLH9J3H9#F2+:>3QLG+6K_HW_\_0PJR; MTO"1S>G:PD`=J32I%*\\2C9VR;,D7"4"6@IL/VF=%P802P/81)RPD.(%H)V( MI5KFRE$R.$ZD0J::64Q;K*6JL,;2UX MS.BWZ1S2&7\<)*!!0/MIW#D>H4I4@VKVQ(#<-;B)5%;)814K@6.#]E73Y=>4 M2F*'*6LUL5B&?7G3%=M-9XYN59DV#\;6H4TI4+=44ARHCT9>T=4GFB,;$)U0 M:42,"!RNLT!R7I^<8J[K;@F20J)D8%=D[)$U*])]T#I-)64!)Y6M,N(4/Z&C M!:2I))-YM4!>5NZ`68J#*KUL++)`IHC[L*COI&9MVRZJ-*#R.V&<_@-H[H.3 M#4S8U2_EV?6#7=./$$)-`\K]L-/TDQ,&+9V MJG6U"2B6$JA.$G=FK=5AG;>VY(3&UTE)>:W)\;4&Y[E6O%+;@L2.9S6V!RKT M1%B<:CV]:.A1M05UG<*ZQ&HALXQX99TL1ZU&S.8:'Z]IJ\VK_PJ5.<%`M96> MU;D\L#6O#JE;.6GSVNVZ18P_)]Z>5H.#\P9?R1-\F-2N)CI7R6^Y:;\50?E4 MNMK*@7M2"VYV-;C3UJ["'%H!J&"K#[`S296QL4`D:VZY#,+2K<%TX,E)2PIO M3ZO(2WP$U>3E00$K4B>;<5VJY?7_^!_H4W'[CEE-C&45K4K*`Z=8K1]K"FI' MD9)&(XYRE:K>O3(SKQ,9%1]]HU=CP,/%"W'JUNRGJC>?TER>()Q"J:*^I<,) M8_WIH:FZ\;0,TLKHZG+@UU<#0`[MH7VHW6:N4*K9+8X8Z=#D)[6[&-?N+:LR MU4QQ`IY:N%$])H"BB%",A4Z<6)4V&X^N0J)IJ/5`)I,7-PDO/1V<:VAP[AEU M*C>A#FFP1J*6Q?[;Z$IYM^[6A94E%N%B5T3+>=DIYY6\#Q3TBB`E-=FJ7OVI M=T[(Z$I%1:@1WRT_S:Z=!/:>V/RT<&BO%^T[S#:Q$E;%-U@%Z\;MK;WFM])* MX"I[!Y]J]9NU=_82OE]C>P'GB;VVK':/8JFAE1SYC]=<5$,WRF!([,6O0EO0 MZJ5/!T=U05V;^%0F>U.^B6\_+1QUTWD>U."+;?>)X@TM3-0J(RQ,&JZFLKCR MG@?J;E[GUNF4$38:=;J!,X"K$(\RP*.7NDI7]!4N42M-M4+Z9OMV7Z"%,7=. MS]&GRPG\$1T6!`1;#^;YP)_GE8;@$/R$;`Z@;36::>!H)J]#:"C-%"I,$F<= M[3R2%'+>!\FQ?B[P+B4\E=>6)*'5J4+8('V&`Q(1+ MC^AA@:M'*E&ER@!)\I@H?CET M4J`0],`($!%_Q$7L>9I0"4;!6.?KA"<5;&58,-B90F-'UIB=G==LN*C0!4-6 MQA_)G;I3=^S,W>$U^'/OC-]ZZW=T,$AJM=NG\`A=Q#.#D1)X-:DT`<`)/CM1 M,;M>'HZ+P'@G-4'`9P?!)>\\QC-3#T^@X[F'L*LLE5/+(ZC1FQZRC'J(3T;^"@68@"9CH>A&*ZA;.C,2"<& M6_&T(="CHMSIT&`\N(M[W%]Z6D]:N.0JTFU+C-&N8S2.R;@(5:051'.ZD]/Y MCQ#YEA)B0OS(#'$'OL@?<&-"32D#BE\-U^$UY]/,!*XMTX8:&<]#1UX-L$)5 M<=TW3ZC$U5,@%H_M2?/RBR]F>U M$WPF*@VVY*4L@%RD(5U`M,>RL,!#0E*H"F789D&8J1>R1W/@CRO(`LAT*:&XU+.!*- M`D,>#`(O81L-1&Z4Y6V:0\2(2G,BBC2/:-=$HGO3Q4JF]FQCM9R7;&!GRRF.-EQU%O?CH(DA61Q``4D(CSRW%+X`_UI*.K*Y%]9BBJ2Y?' MIM:E)+20*IUT1J2**!RM&TG#8TS4'8I% MI\QO#LXI<15UPV>T,^NS+%)`[N?Z?:=2Q(M^8(_RF7:!6"(?M3MYH"48EY'Q8VQV'\HV=F@; MI!E-Z1.,0*I&99UG)C.2\]7]"UJFJ@Z5'\T3RD*9L($W5EC1<*7[A/X^SH5% MK9_YC8;F,)/=.+->TCDIER^9T<;)=3239NK)C;E4"KV=1XUCZ]J<-';BY2#+ M941PQENEQCDHFC6-(>;$>H1AOBR*H+J&`B,>>)MF#2(RU9`)'BDB,&0U%U_0 MZ9$=NAE"IR3#8S!1K:[4<_838>JCHZF]L*_N3V\:]?C,88US<.J M7$)EVP3?^1/0=/A1Q[&#!%^)3L6Q.6+,^Q1I[L,#VW.L431+*#D&R.W#J>/3 MX]Y14=)3!K,"(LY+I<0/C>/1;+LAW,`J!GWXSLW<0S>8*CZ`6/>?/X MJ0Q>?K88R0@;:OURKO5ILM2ZVD3?HTT]JUOVQNO3QM:V?V/]N3 MRHA1@ASXUISO(7C*5/IP3F<*@#Z.'T0^[NC.$4ID$R;EL6T*E,I27VWG5@K%UMTYM9?]7K!TVVV>I[\-R9-H,-[8=-+3B) ME1AW:HR4WJR5>;,"-G%3768[8A^I!?6V.-?$X/:Y0U?1\E.:L/WJF38L8M31 M,*\'Y^0&=(&!4K!L@$\05Z6C7]DE@>4>]&&-9DSB32+0MR-YMV\T9QCNS>OH7KM7+;9F-5VY5%F=Q5)]4`764[ MF'(NL[&HQQTM*_?C=FUFBDJE;\O=RH1/I)R+MBR6\9U>!9\26P/DD-*;7?;K MOVIFC/![R#$ MXG,<;O?HI7E5Y5@;/GM7F2MG\3H?)L+'MN=V;FQ;LE4J%<[9Y!KY\X(5;@;] M.+Q9!'JWN)5&?VXPVM5&%1P;ERXDNZ%O-MI$;\@5J8^J.^)&W(C/Q1VNP7[8 M5*6T;"G3L$'WQ;T#;D^#MT>"6GB#?]OP.JQUBQ!+AL-C:_2INH;3L,$N^ZI6?WS/51\D-LBY MZSERWZ5[R2!RSH;`#8)6W1$G>W*U4^(9G<-<"V*J6`N`X*J\ZOA&*>K.=1;E;,&5NC253UL<8/QQBVZ/#F,7.0@ M:I%#'6W.R0.=\%%;X9L6I6D9[;XU<%:RR$_P`Z5S62YADPS.5.=7D4Y=6FKN MHMRWB[3GZKQGYBE]#LOU.36GS]I*&VR5<]<<)[AOD?OJ MZ44=N]KSI*[3#7&[ST3HC6 M.W6G4]5\,79VX#Z(YWIR_^OO_=S==PEKK]Y=5EI2Z>C=42,2:+WRWQ.,1,Z/ MK+>9PV6#,R5E1VHD%515'LZ-R?J/?(N6)P=8X2C31G;MM]=R,L6;]$9UY;VY M?UWYVNEQKJ>G>!1?2=FY(7_<+QY&Q7@S-8!E_(JG\9GM<3S*]/'J M.Y:Q.DVF''CM-=7:^U:VZNM/SC)ZJ-_KI MT\^UNJ2_])$^TV/U25^Z43E8WZXF]Z(G^(U>S(-FAO*"O MIQP2P/UY62_K%?V?+U:UOM4C\#QOS7?]K&_UZ`C6+_E-;V01.9%/])DMLS=T MS<9\I)6F8O.K'GY%]'JN`7%&EL-*H=S(WRWN4(>,]8W.K+J:F1)?2K*"<^\^/U('_0]/:6W M'31>AV9WE`IITZFA4W)B'PLCNLAJ]0_?FD?\053<<;U0E/B8/N-K^D@/8A]] M5-_XK9[36WI*?]5!_L(AN5#!B3$#P( MW``>\W#CGJ>G;SE2J(>W=@;T8;FEH^7RTH\3W"V)# M_HQ"X"MVK6LZKL3>O)V\?W5?G_H\;UF'ID-ED6S7Q4`NY_8XG0)3(VJ>!WZ/_\8'D"'G;)>\A?']^-Z4 M_3Z9(T!OOUMB=",PYO%]-BNRW!6_%=2/YM/$%5#+6]5]RWA=KXO=,QM8*VDP MO-NDJ#K/O"G=-UW>QYON*6[-K;AC_]226=JM5UFQZLX21#HK6V2B#![UZU$3 MB8RN\\-@VN[!/WCG9L+1=O-_V^BKRTO_/3[]NQ;UO_Z=Y,,C2K!6P0N_/W,X M&NR`F6T>'@WY>/E_V_T]&I[W80_.GO?2"E+O?_V7_@43) ML@&_>!JF'7QB`\XT)4T^];PM*RE*%>@"+1PIBKV2!6*!\!\7>+@%@3A1#RCO M=8'MRA>X^GF!K\P66`:F@74&[=<#4H$EC542VW1!]522H7>@?F0/NR1\$'*Q MB`R``&IT"J!7T\C!>TG+L++(1%M7RK+2^BTS_=V$E^0X@NN+N[+;^$7/7"78 M`6)YE8VE(L5H@";@<::E(#;[$]:X%VE!/(O MAPK*$0L&1H#.[B,M&7=,3P9TCH!VN*##@=G]@G=5+L@)]H+#(-/#"PJ#C!\Q MF`PB@^D=,X@+@G9*S4&%]*PJ30D"2%2%;(%;-H>W.7%:B;MQI7P;GXJK=J/8 M-1H9.&(..C&]2CIX#GIOZJ"R8]1D*8L/NH*%2(`?34+3WG1(`!0`15)=?4$= M[#;]@'CWSU5DMXU;989;M/N$1_P-@E-0)80)(=_1T2A6L53$T04M,@>?1%/^ MM"FJ("JC_&D[T9SUYQ'F<2!A"_<1BH3FS@EG$C)_SE]*^+5U;M+./#?0A!\P MSK*A!15^&?&%BF`'UH2M%7%8HV?:#.Q[L%@FM)>[3(Y"E@H>66V+5N M2]FN1!=]/LO-F9(6KH5GFUK8%K*%]-];*!>:+39>HF07_ALW'ZP2!`@Q78U+ M8Z;\A7,ADT(7LH6#86!(&+J%A6%BB!@JA89A8Z@8'H9'#9HB&0XY?J!>*!'Y M`.:&GL#,37A?(.O%^F$A#4X7$MG0+5!)G0$)QBN+#-B4M\@9]8U4.,V=@:\A M_-?3B(&S81`XS,V#JU]MF!O"0+OAJM,;:BN_H8ZS_G4T^HS\!QS*AKPA+8UA_<*H!(:>C)H"IRST=`EV:&%0]_,`+9.*NC(F$XS M`('&"%DT/QVZ-_4D*MW&3-AD&!_?U'MH$R)O04KGD8T@A4UA?H@?[H=1H5,( MVP"`3Z%1R!3RAT=A?Z@?&HCZ38&HM5&%4Z%H&@Z[3N"(%!U:<$G M\=LRDPC"?Q/@UY:L#'=MB[UB"U9SDZ!^A*:!>$M>B0@?G8@1AXDHGZB((B(M MV+;LAKOAG/;M_'?!3M/S`F)]H48J.`/Z;F6*D86WX8&8#7NHPQ%ZBPW_!\Z0 M?X1>^W?_O3*,80D8_]E_\I^0-R4ZB56B^W?^Q85/'ANG_R$[Z![HUFW-/6.< MI\.QM'B>CO56_86$+!R;^/R!,]'?]0>V4'8T7W4X??2!--R$N`"R7L10L.=M MM(=BC!%@?!PJ@N+],B<1BK<@H<2-S&UQ&]I6"@&(@%B?>Z(&JWZJ"+,D=$F(DLB>6+82< MGI,Q8HSDSTTB(K(J-,]_^)F4!<-<'07&HU# MC@#SW5$RXIV@DQ`U;^T6TK.;<2%68Q7EW>UTU4K`P35><:$'<.2HF$,72'B' ML90HX@V;%-&(AM-.9W2CW'.$$OZ3R@$IK\?<""G)=@!Z@)QCKT8O8V^+588C]I1H",]JAS6(_>(/6Z/X*/VN`M^C^)C]"BZ MJ$*R1D)U?#0;Q(C<<3(-*GD&(%)1G1ZX2%/55-%ERYC,]-*X4A<:_\@_1GVC M4?\XC@R0HI$`64#ZCP0DY\$_OD3W!X5`;O`\G%:SXK[<:_>:!3>2D47P5/+8 M=VQ94,>IPC*&2L]C\\B9R#HA9)I7ET&/T..8P3QJ9RHD&K5"BI`I)/"$0K*0 MYJ,>\D+.D#`D#=D\LH`Z9`KI:7V*9898]T`&5>X(QZ$^`1I]D;')&XUY60S@TX$E2:A29O3E:-' M]I%^6HZR26:2>PG5%:5$4-5,\`5'4EMH(@RP2.J%X->F\DN$,"!5+PI(# MU`-%2V93$M0MB4V!=;.+I[=N[$JUY"Z92\Z2Q&0PF4LB,\*D\X1+*I,=B&06 MB=@@W,<1R6THD>G9&$ENT$WW`;(4GZ"1[\:>(62Y--@-H]&)N`]U1B?BJ#E# M\]-[(V5H/FE&-QEO@%==V)]#2<(R4=:3G/D/8DX!9+[Y)03<`T?Z(9ZPKO)'5P4_.B6V&S^)&<33NW:`&FNV4>AI/J5/VE$`E;^137F8'AV5GFK0;HY2')O>(7W[(TY&= M)9((Q\$!8,6#:5Q5>11E9VB0P;4B42UT4/CB9CU/!Q0$Y22Q)8.)\Z4RF2I6 MD1PEVIP_7.06Q5>M)8.*0>14BB.PD0>F,HE<)3(H4-!7`K"7U5<]`` M-0>41FD\*[B'5=9%^5XN%N[A$1TBU]16PH@DDB-;^&727&4.&S94MB0_(=(T M1:3U&;D/*7<$:7&@Y0.6=?V6LU-P*7=HE,*B'W):DF2IDEQ&>=US2Q1).5(. M0')):G9/75,N)09"+3U^LH9D:5L>B\,'8)*UT\91(;M:`M1 MDQ&+=$%1"*QSA-4EU$]HE)1M<\ZE25E?-I<_T!(%G>%04XIT@GN0)K(':'?^ M^"LU)0\0@)T^T:!NTUAJ;-=40#4]!6"=#N(37+Y.KU,`%H`13SR;E8:0+3V# M9:I4AFF5T9-,"6X5F/&AR@)KU#)I3O8QI;B4[-GRQ&(&'CK7:^+80#-,(J#L**!BBUE]Q1KRH]`F9$X:IN2+B5/%4_4:D?E7QD!##(DI4;)N MK)2/:5%"CQD'61245)G/)4=I7V8>*I(A-L>HDJX(W5'_W'/@FW(I^BE-+2;" MT14Q&I#)`T)[C"K:98$)1K:!Y")39F5P&^ID:6#K0&OO)#J)9]:93)D6HK.H MDXU5&7EN%%[JAJUSM/$8WR1?5/KXDBV966AG]DJ[2^@"9!(NZ(8-8A>84L%E MP%3'-6`@$QX7#\(<&1*446J:5PT;.F5.?2B&5I359V%7A,;-$6OR3R[2(]-9 MEESVX(EE?ODGB4^-Y6MF5_L)T<%524J28.S6$NH;ST90U5!A2OLCK`."A%S7 MR,?5;\A(TQ86Y'%=FX./"Q0R#GCJ%!\0#,3E: M2,ZT`Z-)'Z61&2)K:D8%D9GX3V^636%N=2;@TJA1_A$M0I7I-=@Y-&C=N9E=?1FC6XME(MU5JE57! M2!*2#D571ELYI^38QO2<.!?/V<295$S,T`ETYIP3DL[)8UU<_\>-U7'=(?%6 M=M6TB)O-9O&QU9!)9QWX]WDXG;E&,%5\I5:G57]B6DU(VJ]+F)7J.6/R72J6]E8(_UO/"EJ$J$55U\W1PEM_9N97`X"ACU<*"@T`E M-P?NN4Z51OH&M'7)_9X29L*E5UJ<%:=>^2MV2$<4BO2G0(0L4)-2>S):8162 M96`]5?08;T:>\![>Y=L6>@A5?!>RU!>RG"-<7JEM@ESR)OEI?I:;Y^E@>*.6%K/@W'V@(VH&2H,LE"GJ"#DXCZ"$S*.5J.@P1=6K]')I)6$9\QD"2 M2`.I.!)#RXR).C/H6\\)[WGPD*#ZE-F2(YUC226C!TP MTK-PBBZ-E0?26'GNII4'902<*\G"`8FB&Y)H:+5^"5:7J'ZR?E49Y51[AF5T MHEX6I;%^!1U]AP*B:HHBY\;Z%6N48(#3*EJ*NJ+KEZ_13NE4P$DI*IS8HHA9 M+)J+ZE2[J//%$J1?AU8INED!HUZ560E6M990!XA"$4XOGPI=4NM;A8)0@ MSF^'$`12`SP?CR@V:F5E7>N9W]2-JDQM%C@JA^@GH0=!HE/E6K,H.&IY\J(7 MV[@1D>A46@T\NC$EDLT+/0J.VC6%5GPR`R!@CP@_"@\FDB**^D5M85D[6)HC MHB@9'6&R5+)9_11Z`R:)HY M'NF+=)X$*AA/2&J2.J(AJ81TB6BD6XSXDQA=I`:(TXFZ3:1>*$0Z(YU3@0H8 MLG<556IH$,?LP%\]";2Y?^:?V"91FFT6I=:F'(I^@J%!J<]YUX2>I*?F12HY MG5*I?I5Y@9ZCI^D)ZE5,7,Q64\3(,]$;BP3509M_!T5:*ME<4AU&>E*RI"SI M4G62@J0H:5P*E_(G(ZE<"@^FI$N52KJG:*0P*1-E5MT'QJ.OD6ZT1^E&Q-&) MN$6IQTWJ-]$?U%AQM8S42Y".30+?+4$&W*,B8JE`8QU/&N2(1Q;22IE6<9AG MJ43*;8I<'Y>(,IK6I"_G9^K_[&98Z!(#64%(&*3K%)2%^V3F_`:\6-- MPL_P*SMIL\%IQ7R75/0AIT$=F=UH"JHK%7T!Z$ZJ!.JY^5YN2^3CSX$JG0W*Z<9"M)4HW&5 M\?5U!J&+IXAJ\IBC)FJ)JJ*2J"RJ/]6B9ETLZG-R?,VH(ZJ+:G4YF-]9Q/0Y M,1KL%W`FI,@_0=>HDE`QI-Y7JK/8?%1KR*YQG^J=)A5B-3WY5L,7@H5D@4X$ M*FLUM/"GD4J+:?+87TGH,/2#+&MQ$E$RCWY5Q17](>J`*D%JT!5JKJDD53[X M*<5*8!>R&6XIIYP076(**H+5*.(9ANZ:H^;QQ7<\3>XFFPES"JIR4GH4;#(: MANJFE*CF1*KF\HE.B4[,E:KIB/:A,U+O65Q=)OL72Q5K,J&QYNN$='2JC]0* MY']-(E_-*-=H M(@?&Q"QYUZM9&NF>\BCYQ:^%C,<7H$H65F;ERFP9)+Y6\XY&>-",.*J4B>FN MMJOPZKOJKJ8C\6J]*J_:J_CJO9J#H"TMU!``/PE9=,+^$8O5)(-7%AF!3)'D M6L(Z;KP>7@<3R42Z&RX(RD:)T%&=V"@$HU"2X\E4HK%FK!SKQNJQ=JP@Z\JA2A,(U$C]C89&RW)`J)@;6+))DWPW0!``9'](ZBP M/5<&*U:6TJ0>9)<1=`0=KFA[6K_18D?)]\D827$:)G&T.\UE@VLY\WOT:^H: MR[1?II09I97)_SBN6&;DJE(237V&XCJ(8F"(ZYB)V&QB;T@$*5$"(CU0$$1# MUG=.5.GJ()I8^,N.=FB\'^[7X3%E$#OTHM;C(T$O;(>QH6JH-Y2)E2&5:&8L MSQ'&'=4>+H\K%KSRKL+KPR/]6&'KHZ]1AT$D`D>U5K;>0ST*#Q4;?2.M:7UJ MO;92`U$YRH6X110F!B4[J2J^7$!XD/P?U^>4Q(Q8'>-71%;$";#-2`'[BG`@!RP! MF\#&2P)L-`>X9&-D!@1)GA4\"A8-,VDN3`V/U-I)[I&;9`?+@]U5F=;Q$3]B MAD#+K_*/BBB.7TOFE9Q8O6O<2KR^L,-K#-O"2B6$S),1N'`Z\2$8DI9]1ZMG M_1J,/:U,JP:[#;4RE9MC(YKD'E>3OC:FMD4R!^UQT&@PR!3-09S-0C14[I%> M>6KQ4J"(UFVNP82Z8(&V/2$,EKT*]K MW*1B%]@$R0[V\H7$K!TD!I)OZ24Z$TH4G_VQ)%',P57UL2%1(,O'IFZ$;"!K MR,8?Q(80&[9>-`Q.FB*#8484XMI$%?&O#=G_6K4BL)QLK*'.*+"=+`,KRAYG MHZP!&\J2LF,*X$(K!&BS:P0YMW1FU9F,UA.]0D,1X?-WW+)*2ROV=^`E>,FG M]JDE9V=/V97\7$9#&OUZR=(A5&OO\<>*8`BD]'J*.+-<:\[J2M5;86OTJ9R4 MK83*PH._SJPY2:2RUBF:K6P<0IQ$&LXISO%TG92+*U[9N#ZN`%+I"KG&1ABE M.LM>N4RQ1MTH"W$A]L]!P>Y":)1Z9<&0T)Z\JZXC%X M#*I1A8!K04JHL1FM)^)6?N+2#I2I&U6TJC"T9NRPJ48!0#@MOY'34FX17JTA*4L9@/*'K5%[2EKRB*UJ*PL:O-PJXRFQ6.6F`41+?+!:=$E M*(>BU9CXJ#&'O80C<13UHP^R=F6?=\*.NG096Z''+$<`CD?XHV95'TFTS=!? M1-2BLZ'8^.J)@#^K9?5UKC$G^Z;/<7/A(A25V*EGC+.`!AWK@X0?,E;X8:54 M(:<,%9)D="*6U2R']8@>&\CMRC_U)UH-,7O*K)PO$4EI/UY@"58[>G70'!%6 ML:&V\$Y9I7[265)E)@C;I7-`9K34H2CMQ$1"&FZ$U2(VM92QP=4&:UWM=^G; MXEFTQEC[.=$A9^W2->(T9#('6_N+25NW%Z[JW(I:#X<&HZ_5HY$&]=5_@+-P M[=K1MQ!`IADD)H:0(7V(3YBG19VB6"PE=-@=4D:M`FQDA[281/M],G:6;=BA M5'9GM1H^DD0N:30)?'4_$666TEU@@RE+:8#!0&_+AD$L=A2KV1?E,<25#B:8IG MUPG3,C,,;1ZIQG)F":4)\N?XDYK3+9-DSDZ/9&D0B6B%&:C'4V^DN=X6Z&3H M7#G53#[9XOF35@\;Y9GMN6R4*!F(=%U;[N54CMQ>5*Z:6QFR*?B'J0$D13:7 MQL8A/C%LIU:&0Y.43.D3R#,DC97$TJ3[9A!9`2BH`1`Q:D@)G$$&F2\T2:J6JO50:F8/-=Y8L,:E/+F]="#+G1[;)12@I*QD2<9NZ#NAP-L5$W4KO("O-P`ML81 M!BHZ&K$'&.+7=KS]5H)+OSQ[+;X,F2P:2$ZE+- MNP`(J61J?2BDTJKE.H%*^BXYVK"U&6`6JO11D1L;7EYYH*BD;1P'*EJM),76 M,^-[E4J]UTY%%+56MEM?HI[P0D7,FX+\Y5MTZXDUMWZ\'F_(F_B`O$$)R=OQ MCKSKS**6F/0;D=U'IPO]G0D7U*>E;$BC%I1!\X(DN\;->^]^62N)A)0>G6(U MB29E6@F]D*=&V8W6)%&E:97T0IY+K]EZ9IA40I2:TYNDC^^<\\-*;8*#9JC$GN"-XDXG8Z M$U\L2Y(]<9V9#,F%^K,^)UT;)UE+5VA'6OAL>+%0],$+8B.#;^++7`&^=I@9 M%R<53377C@7N.CS?"WMU\:V!LL;\P46"8H?+^A/4-6(46L,;CR6C'!#B*_B> MOJ;OP@=24J]1ATQ#6R5VBFC)ONR4[!O6`".6#]Z`B#J:Y)"%]JU]'ONF M)A6?"+_C5P1"\SQ>^FF!&H$DJ/0N@TJA-K\-JH7ZH+I9TF^%.OVB6!L&91_7]0L!H^6N+FK^Q*/KK<@BDN&KZ>_Z^O^YO_$MMR;^) MY/_)_D*W\^]`*DFF6O"/ZA1IP1S'VI61OH)=`(@0P^Y^I10M:G2"S#SX6#'& MX0%J*%M=:WN*LB$9)"9ZM)>4&(_!65*`$@>>N&;('8Q'TIK*0&,^&ERFE?%L M'@C?\>3.DZF'S\0OT6?N*X8K`T^X'(A"LJ#9FI`&_%%=HH!&FWWDM'6%(UN> MA#J:'..5W.704D/F3)C!K5F1C14VRW?<'^[#!SS1"CK+;14,O%)(6+#=VFXL M&D_6K)&-B)#IK0&I,"4P6*L8;&N(D!'+E+)PM"1?AHY9)Q0?H1!E*<9D&<]K M=S+N:AX+2M6$IC"TY6NFN9+,&GW9K$$5^;?/1BNIG,(ZFLC!5K+QSU"S`1\"8HVF?'=L.+N(#+7TR+2=$7`H'+-&P-PNZP:"8+.H M#T/Y054[\TSQ)KFUBM'I>3IDNJN3QHOYFOA$)>#YLPS%(A=,P$("_\(GL`E, MVYT?0HC$P=C^'[[L1";+&F+@3S/,##_#18T\0GPN3B8.@\-Z;N!F7A4?WJ\JY0L'PQ";,.P.M\/M4CGG7`*9/@GK M!CJ!8I3'6?NLR(REAX%%[F)KIC`Q/%&*0")0>#.)4#YP,,9;1$89B4LX3`YS M((&:0RP1E\/CL$0,<00I=,+%"XBY9B/PK:&LO!OY60/R-N4$Y4ANZYSA9,_( M<@<#+VB`R$IL`[?$/1-+?!(C9^DK38P2Q\0O<>#&'5XEX!DY22?DCAWQ^6&* M%!\Y<#"D#%/#R?`R'`U#PTDQ4GP42R/-77:4_/C$$.5W`\FM4MQ9FN,#3:>X M\$OIR=(P8'&\UE\230:IEHK>4<'G#U4A\Q7"$P>=&N(4'\N(+31[?"&4&D7\$#_& MQ\Y=Y<@0PG0"_J$9U1L5[9V!T9YG]P%'FXK*@9:;+E=GA+2Z!DFK:Q2YH(8G MLBN&492QW*&![B>P,;4EZ+Y;LZ5%[!C;0WU);NQS^5QQS&F&&)T9@['(D>B. MP(G@,SP-,\5/\5*L'#?%R+%2_*G]'.*EY)&Y&J[7<&NL5)89;HYR_*ZYKD=L MR79E7%L]*D9+:PA-)P=\F#X"(`C8-HMME4-J\<4HQ7D;$U?=YZZI:\G6G]&8 MYAZO!]6D2F:?]T=3`GE`8J@4@!S4=$S!IA!V`[7=L(X_'ARVN81ZON^CQ#PN:6)BE6X1)^&@P MJA/98XU]H&]P<-PB2QO*DAJP:&8:DMU5?0'()MQ\`Q?E0>C/+/F5@ZRJT' MS>N"U`D@S6!#?*C%2.OS`\1M>)PR`8SF_*.@\BDE*I,H9FK/)>T\+D,-SCG\ M1"3;2#9G`^UKV\CKS<\D:U#)AW-5)^%@QRW\`;ZQ.D^J7]G3DE4#I4^Y7/R4_[,.W-0&30#S?%@ M+]=O"%7^JJ[QMWF*3.6)J674G8@58A6E<$"!(H3V?NAXULA[@N/@..'I>>FA M,$$AIO`T`%=0Q94BHHBH&XJ(@\F,M*^$1N\+#&8EFZV:T8E<0#J@H:/);!E0 M^I6_T''CL*3G1T)'%91G5JLJ2/=L%Z=&D_+"5 MP,31+:R((1QG\`M3JCT@W,NHEK$0A'P90$T)_,D^:47)/MY;A!L`T\<$\;-1&U43&(%Q)! M$VKE35E;09,E&(?/C$%+T!=T#`)!!R>!R6WV:R70!@79]GG86G86->5B7QIPC7Q5Q M>>=E"X)`F;,32N)3E2\*DNV"0Y-)MZD:WO%4;OM.E M0#(]XYQO2]F2?G@:,0CP7,/=!4*6W,O,Q3YP&&*IT)G47I1*_3?9U6>S'CJ3 M$:IMUET-HK6B_.E0?4V9(7+)V2QX3=5W=?1A6$-7775BK<,8UK]6(OE_Z:.- MR&8T8;6_D/71C*L.5;CJ;Q-L*6V/2A=J M;YE=;VC;47\*TS[CX',8!5/%-!1G6^?6M'50`R&YH?82Q0,J/:6T*8E%,.U8 M)RJJE%Z=5((EHC64]I]'*7,]A$"'@^55XG$M:^OF@*IN\5Q>DM1DQ_W!-_1: M;;OP5?F>H'29F)B6[;Y3%&]FP4P2[%Z/EO#UX*2I;,EDV+?'WB0T1!_$>-P` M9@AU9Q*J(&93!HJ+XG8I754*8))X2.QQYQ5<11:'DW<03^/&A='#;9_ M1#^(DQDGSBTD'J9#W5_'OIW>CO:''AH<9W_"'2,APBPOS&(G'R('IUDVOB#< MK-.J'F<_#=O3HHK`GBH+@T5P"D/I5752R^X]5S:_DUAHQW!0WQFUQ=]P/]\>]]RB!.S%+XVRP&<@2@28+5,BQD,1'SXW"WVVK3)*-O*28OY;!]-`U`C3+/= MK-$'$00"17QD=]H=\>$]<9(`!*JD(AY1LZJ<2"G$]I+]M()XSXHA:A^U&4`W M+;1LR"//5I/4=^#;-:V][0#=6B=WF\&E&9$DQV@\+]/:B;`+5I.\:2UB)46^ MO<(CCPJ48^589.ODK6LZV)X48LB9O\IB-)PX3JHD[2V;0QWQ:C,+W,)W\1US']^,F.WY MM9:U6!!BTG/OW<'R9/>I%$'2K'[[-YG1A=26(5U&*7*)XV0/([@RQ]#B77O7 MZ1F5$6'ULR-(<65N,+?>5ON-_PJD1/(C:9;U'WKWGUV;G#]N]&R[/,FW:L== M,"A424&&4*`ER1WX@H'KO9I031FE=T5?I>C.Z'/'$7X3A]\7H4C@`]R^R_?U M*]8'!EY].)S4!P=N2WK@,,??[$*+FOVEJF-3F>"UM;2EOU`?.-4?UX)WX+FT M^@7JR.#3QQR8R-G@!Q3V-;H`UD-XV8J)HI;P&8!!,>/2S9&$ M(Y*$97DJNO&:"=BM*J6(.J9IX$>$NWX)#2S7X#@T?0L8(G!$P?-'XL@G"RR@ M7@_'L#Q[I!!RQ#5'X:S2`%>E_"V"-&`YP)55J29V\JA@E\MB!8X$C6XI^*@Y MP"W*?3AB5(&W6U=?Z\VC&.)1^"".-17B@?AOHX@WXH@X=OF(Z]9X2R3.;EU] M:IJEDHE_UY3X;EUB,1^M=2".!;H[>+CB,\#M8G!X*=XZE^(DRF'DBB]9',S< M\]6I>1;(S8>L3!Q#`#5F)QQ5`/?MUC&6XJ\X&TZ)!^./7]>U-0OCQ'@*GHP/ MX\=3:12'A"7E&K5OK&PQE1%Z1,^0) M>4;^C*2.#AS]- M"!A)KI0/Y;6U4^X6(2LYN6YM$&Y;PB,9[;VME?4F.NY^93#@X<>Q:>3B1PN# M=^;%GDG'BO/ MY4@V-^Z6I^#!E`MSD-HMLIY)4X$'XHQY)ZZ*5^*#W-HGF4?FE/G/6)DG.7F. MG!U67$R7"."X=.OGE"'JYW,X]5Y25%^V@AW44)T%R&!WA4,N% M7G'>XT/XN6/8V*I->'FR"0+0<)@&SFK!*_V*V$+KHN;87K M481JNBR"!UTBRHT.ME12)_ASBPC[C##T4@2D@SLON&].6+[@D*HF_C>!**_. MK=&"G.9$4KHW3*5O?M7'>3('!EWPX):>I3\97GK7$F>,7D$ZU$.F MVWWY:#E3IJ\L5%P',X?@?0W<>!ZGXWV@TN)7IRO@E=?BYS6A.A84+-FGSY)[ MP2G#:10*>@:^]^[^A"E+T($T(=56)!$$EPQAY0S?(?8$7ULE:RI\,LM2-*9^ MJ=NAA3)$^%.SRPQY\E4B)]C\QC;3<]$Q$@[`"KC]A"\'CM.0Z+U/G"9C=O%TC%55_394N<_HB!OBX[JX3J[/ MZMXZK:1T5H]R2AV(?L4H6C)@?76, M9RF8+5QW0L`>L!,A1,@((IP9)Q:7R8CU]:JJC\J9?_S?RRDB:"K#O'K?N+T& MH51J)<8^@8ZA'_L#*K(WHA0HR%ZR-]_W2O@):M1J-5RS$6T(0IS1:!Z]`>M8 M:51:LUNEZ(O-#I5:I0MX5LJ`!YU*SX"WT^$S".CUV3_K4T"IZD0^3YY)$1,: MOLC2F]#_HGC\+U7H.7(K5Z]XM/^#P_'-PO47JG]Z[*-F)8U;\XWHN#3")=TK MZ0<8[G_G'^_N-P(L@>TF>]S.@(KL(7O=?K+;[0DHR2ZWX^UT^]U>@6[AC-`7 M[KK,'U/P-U(;QD?UK2=R+R]))NB.=J+OG?G"U[I&T MN5VZAZ:I^^S^NI_NN+OLKI1V5?P?',@XPGL/C(5#O#,N9-Y6W>_<.N[ONNGOX#KY/2.%FU?>]>^^\ M._INOJ?O#*@QO;Z_[_.4!B2^RHZ>(4\\7'8WC`>$7<6-'5`+T)2'E*9G5=:# MD3)HBZ^$Z1ZQ5%LE+HUGJCD,H7C4P#/P#KRE;M[,2!FU[[%\.FNUY\LE)TF; M0RD,G$PE@J9[=&*42&^DN-@! M#P\BO#6\2<65Y/`X?`_OB,[P0#P?;5(YTK3IV`R.L/!+AN1^A*B&W@F0]++J MZC^[SAYT]NQ2?!6OE4Y6BIAWJQ@/.P/4O.C.$@4 M4V-<^Q67AZC*782/\D!RA^X08)B:$O@LA(LZT_"QO MRP-S;/OF_S.4HUC\US\]^\->_-7_.=::CBS9,H_-0Y M/]X5_52:$(#2T&1.`7D`$PQFYP!B`&LX&,<30\ M!5>2I#`19`0J`$:A(Q@,6$:-`1O(!MD$$*`FN!AD@$`/!WP%9T19@!8($S-! MO6,PR+V#@Q'`!J``04`*H"6)`2C`%9`"M`!G`4I_$:P!BX-)CQ>@]%-%"B`' MH`!C0!TP&+@!*8`;@`+0`2F`&8`"G`0IP!F``L`8-CU./Q;,`2G`@`6GU<+P5@ M"&:`"S#7MP%U_5W/!KSU@P)*#P2D`&H`"G`&M`%D`%=0!J0`V"YO9,&``'V] M$A`>9/:)@AR0!PSV^XY@+]33!(H'63\\MP^:#W3O+!3W08`8H#&$`:C!6U^. MPR`73%K_U5,!MT,*,`88][S!;R_4)P66_>BPW'_U8?U8+]1S`6#]4G]&#/90 M0&;_,90$;\%AGP3P]).]DN`>*!%<0`K0U)/VBX%4D`8H]G9`9K_,Q]UPG/?UW?V M1OU@,`>H%6+!8-].[/@QO5&?&5SUX(%@D`9D]@D^&8`"L`69_8;PVW_U^OU@ MWP:$#F0`59`"C/9#O1G@V7/W,X!;7]S#]\N]DM\@&/E)`J-PY:,`8D`*8-?[ M"3S]5R\5,/43A[.T%[@$??T=D-F_]"F`DD_4M_A1O@+0XGL,)0%3?]]K^4;] MB?_?(QXKOF=_%80!2H%VC],'!52!C&_4]_//!`4'G;_D4_7% M/8[_T\?W3KV(3^);^$0]AX_?M_A=?8,_$CSXRWV+;P:P!:@!A\#JWP%HP$^0 MV5?V1CWW``?T]A,3]S_]VN#J6_58_1Q@Z:_W M18:M7P4@^&?`F:_?PP>>_;*?VNL('T$*H`!P^%%^BP]/6/9``4]OU%L&F<%2 M/]A'^4*]/_\3,/8G^63^_V]9>]+Y/?1_LR?\0,5AX*.(/1W]B%_-D;R M3_<8"DI?!J3X-L!IC]DW^);]@9\[N`$[?9-OV>/W9@!I[_`K]3M^5K_66P1" M?\P03VSZ*$!3D.RC`&<_Q'\1Y/O%_>A`W3/U\_U$8#2<]@E^KZ\>1`H=/\R_ MWMI]7$_ZB_I&?<`_V*?Y:[YE3_##$?"!C._F\_1*_HG?XF\3%S^'[P88 M](Q]B[_ID_8G_F1_W1OU@'ZCX-D'^YJ_B2]&`/O3_V_/X4_\%?_%'_6[]@I` M7T_\W_5`_60/]S,*/+U0+_NG_>I^"J!X9$ETPON0UK?X,8`4`-M@GH#1$PMD M!/`"?`&.WKQA)M`2P!PX"XQZ`4"/#DIOP=?3^^DY!I)Z?S^GWP(0HR/>,P*T M^,!\;[TX`TJ/TG?F6_0I!09[+[[`03S!LM?9$^HY_Y9\\($E`54@PS?9P^M] M]71\/+XMP8\/WG?F*_:1]AQ\Z+WEGE%O"S`#Z`(8__``/;^?@%2`#7#=D_@! M_;)Z"\"P''0OYP?SV^IU^(!Z1KV#@*LOJ]?B2R8`]$=(&7K\$7^BOG0`>6"#8 M`,N`1+XP@)'O+>#66P#N&>@-?;T%'Z2`(B#CJSC$]:)\64#PP`F0P]<@2.UM M_2R`Y0&FGI]OJ,?3X_/%]XQZ]KYFWTE@YE?[B_9%]B9[U+ZXW\5/J#?[*_Q= M^)Y\';[Z7]ROXW?=`P-Z_DH$TS\Q@)7/$'BDF0*F`&-\DC^I7A-`CE#;0^F5 M".``2;\-P8/`^_?56Q08]+)Z8$#0P0)!U9?]0P'0`.]ZJ3\0GU#O_3?WRQQX M_UI\Y`$E@:.O1#`8&/MI^81ZT#YI'RK0JX?]N_@M`(L,>(&^7LIO$2CCN_3% M`@][\`%1H"RP!F`.1/H)`5$`)D`RX"\PGF#IZ\ZD&1"!*(`G0(GOJT?V"_\9 M]=8"A@'"W\]@L*<>X!YD]6A^I$"$'V7/2I#R2_`Q!+5\6#Z%'XS`LW?_8_59 M!+!]"C]H(`H`P'<1L.QQ^^@`WKX%X`JBO5?N0PE>]5@%$3Z6WTSOJ'^ME,U!ZE<`60`I`)B@GF`)^!SX&AH'!GNVOLN<)_`<> M"E(`WOF/XF<:;/$9 M^@"!:C[2'L0/!0@F2/]M_AR"*(!M`J3`LC?]O**@1+`E:]HJ#R;VQ(+V!3H#60P&\ M^S0%3$"M'G0P[O?M6PRN!UM\6L`U'RG0"WC8BQED]JZ!@3Q/]1 M!$F$B4'C'E:`#"#8NPPZ/TA^W$"@P-&OQ:?X8_P!!F6!)X$8(9\O#+`&T/`Q M]1I^Y`'5GO5/*I`*=/BI!V=_CL&NG@L0X#<2J!:L!ED$-00HX;E/#F`=Y`RZ M^LB$IS\9WV5P^:`3Z.M1"!F%&@2U()_PL'?B`P66`>Q\6<#!WW2/ZR?DLQ$" M"H.!,T(GP$G@40`C"`/H`I=[F3]GH%-0)9CRJQ/6`!&#[T&"7\KOM`?4^^K- M`S<$>L"AWZ10'S$%A/3U_E!Z?$!E8$=`L1 M`3]^&<$.X5#/LK?XCMB_V9 M"L=]1KV3X+G/GY<'$`%T"RE[3\/Y'Y!0#B#8JPJZ`6*$D\*]0,Y0<;@%;`NP M^L!_D$"_X+101R#8^^I9"7J`^;ZVX+9P(N#H.Q2<`DE[T;['`05A<:@:*/SI M!.-ZGT.7`$HO%!CT,^J5!M>!P<-\3U]/[G<==`Q@^Y)_ST+/WNWO3@@:O!0. M]N*$VS^`($4@2UC[VQ0*]>:'$SZ^81KP=/@_C!D>"JY],3W7(4+`Y[@")`56]K1_E\*<'C>0J<<0W!!^^L:$^;^J'PK`9@#S$Q_> M]8"$J\$Z(#$P^`?:VQ66!RB#8L(L`>*0M/="A".<"`6'@D-)X8#K35$@Z.O= M#BQ^@[^?P/%0*8#94_+E`4X%.`-R'XM`9K`VV!`<"GY[^SX4`!Q!2)`6;/%M M`6X`F;TN@*6/WH#24PB^]7:"R4.JPL:/:&@EZ!UV^-R&VL)_(-8PF1#4VR`: M^.R"G+V6H+Z0AUB*F`+^^X:&]L(M(+Y0?H@Q1`&(!*M_9K_7X7900!@>$`N$ M_;A^'#[9H'G0LL<%6!UJ"&&"C+W880Z1=F@[1/I!")=^%X&F'WCP:LC4(R5: M3::`\D&@'H?/)X#UR^E-%@Y^VD!!8H7P:;@[M.R%`T<";CV/"N7OD1C78_T= M!!9_0CT38@)Q;)CCVS!($UV):0!;H6^O"1C^X_R-`LL`(41+(,O/(8#2NP)L M_"@"S<0!%UFBC?@VA/,U#4U^*``>(;O0;O@J7/PU`UN")43TWK0O;D`5W"!N M"D&!:,$189-O)%`)9!&D][YZ+(+X7P5QJ"<'6`,X"&>!+8OC!_V?<(PR,!(:#VL02@<5/W'?8$_YQ_U"&_C[X83IQBTC:N_H) MB[1^A;\8X5Q3J M^\:*&@I+H2QQ#V`12!O($EN":SUU(@H1]!/7P$^C.K-",U]-[W.(!20WL"OH!-T M`%M\%$"XWL70+>A./.R)!5@$?<#!'E"`WE=`3/99"^T%K47P8*QO9`A#S`T2 M^FB$L3[)P&$/OU%SF)=CW1H64O!=A)'"&V!&.'O#XN(6-P++`W?!YB M&/^!C0*Q`"FPH]A:;!_:%[V$H;_SX6%P`QANF`*V"9V$6,7R@&5O+9@RE"BN M`Q^"VD*6GPC0*#C6`QVB!\^`Y$+'H!W0;[`#M!KJ%^6+1L(:(&:OT$<+O`Z^ M"S:$3KV@H6-@ATA=!+RY&.^+\[VN'KO0J/?[@P,L$L&'_T3@WHQPDU@O3"@R M"G!^+<"I7O$PIN<\7`-.^\(#+`)8H*Y0"3@5F!BB`'2",\((HMR`>OC6:QJ> M!WN)4SX@7X)O`VCKH"EV!-<`(H`OX9%0J*&0(C`:Q##J_`J(6$,2HSZ0(LCJ2Q=0 M!5`#W\![8.A/(\@Z#";6!&>'5,;1`$-F"GC5LS-^&3%]?#\:(1=1/(!/#!-8 M`BV)TST%@`M00Q`84.Q%]M*)54./H(&0,_@9E`;^]HI[/$(EWW<@7V@@+"*6 M!'J&N$(3H[ZO2]@8_/"M&G4S-$5$H>Y`4<@9'!"2!"^+2T/@GD/`?1!-@B"& M"2N-ECU'8^//Z=<,9/8Y_81Z13^V`09M-`!NH/%E24(",X'WP1W`KR<%"`)` M`9``+$4XPL\PV:AE+"@Z&>V'64#!H4I0H9CQ(S'4`=*`9,4Y092LKU=<'`"N M#0J`50'L`?-!`3C@:@"&],R#3<80%TIO_I6.Y% M`9^)?;TYP+W0ZXAXR"J*!T>*UT(9WS809=A.1`U2$X^*B$6O@(71B0@?/"76 M&UU\\D6CWM$`W(CZH_E]`;&/0@(PXJ'@J9CZ^P1F_#:*><4@('!Q+T!0)!\Z M^PZ,G474(!O0B\@R/`XJ``H.-$4*X'PO,]`W[`C`!FF$\SW_(U'OH4A^[/?I M#U-\CT;RW_]\\4?+WK'=\&)P!G@B_L"`4;YW8&11?S@O\/[Y M&(<(O02JGK#OO9@\A#B:#U&0><23H6RQYDC4.P-V&F>/I,+Q8_WQ'VA^?!X* M]>`(:P'(WI:@$L@]F/(-&N,,5A.:HM'`X%CD`/T M^)1]X(%AXZ"OZU>#-/OI!?V`OST;Y/+P_ZA^I/Q(HV!3$B5(/N0B_@:=`,J%2^0)+X`8YA'T3`%+##2 M&Y.'';\GX60O2G@_=!OB(L4#_T?GH5P1/IB-+!R2]J*$CD'\WB<2[1A#5#;> M]6R,AT%[XJ(P1%@9Z#I.!/2!!L@91,Y0SPA4*`,(]AZ$1<',@09R:"AOC.\1 M!VV*X44NHP@2<;XGDCRVG/2) M]02`4++#.#^VWS1Q+-#Y2P-\ M"\V$%&F,+;[B8GTQW'=?%/*) M)8>%-0$7(QK1+%D?7.[-]WP":DD.@5<1S+CW^T%:$:6/JL17H&?/VE>2!.H% M$G5]1\0P8!(Q=E";+".J!,F,[T>/HH`O(SG2,T""P&Q]0@`(@HQOOD?4XSK& M_LY])8&@7V5PT6@-5"XZ&V6'+#-U'?3U)XF:ON":A\\ M!ZT`(4?O`@H@">`$B`(D"=UZ0CT08(&2WRA"9#NZ^*:/&$BXHH?RNE>0Q!U` M&4E[K\1,9$+1LPA(%X%)04[0I&=%A.]5)]5[8T86W]C0_H>?+`F\ M"N6'(#X>XBF#Y+<80!AJ"/^!),3P7XLO0FG42U+:$(^,4D("(B=03KCKVS#\ M`P%\BX([`)12TGA?_.K1&,V'A,7#H%-OUD@_]!&Z(6^%PL0BY7A#*EG\8P6^ M&M^,;X%&@6BR[.`(3.L9]68`QL6X7A%@\>?32_/=]!H%$\&59,NQZ,=#G$>\ M]IP%'\#XGI*/.[DH]$[6'EV"_DG^(8!RJ)<2=`RN^N1](#XU)-MP)DDP].S% M`1,#X`&689[O_K@;?`^F&@6!G\([7TEQOOA>[.^))F<""K(I(`4P+WA0L!PV M(BV))<*>X)\0=B@F+/WU%!T$_44U)4*1&[A.E#AZ_P:#P;_=`'F1JKA`4"MZ M"$F*_<(]H=Z/ZTA*#)GU]0`#4L$-`9"P,!G7`PP(!MR$UD=\('GR7ZGXDPI4 M!A*1*8`V`Y@15K@3S.F%&AN'B<.')+[1/_B6G/O%)5F2'T+:89$R"C4%/$RB M%>-Z\TKYY'P16IAA5!HL$'*'03[LWXY/IUUU`@.*]%]UL80X,92_;?>4SS0%-F'Q\@AY79R^AA,C"6R M(TF!*L94H*7RKW@K_!]F$[^%SL.47\_P7CEX3`4"*%.6+LGQH'EO(KE/!#-Z M^*)^(@>QX8&/9^GT(UHB'M*"/T4=H;;/0)A+'/5Q`124#S_8Y!0`"@#80SYZ M%MD`B<;])*12'7@8S#KR#YV%R0-WI$90K5`.#/F1!L)[?;TNH]#/,E`AU!#\ M`RV0M,OW(':PR&?IY=?R_2@K;"^> M#FV28[U"W[KP!YAA-!&._H2"^3_PT,;2I%<@B`YZ_N@`@<5WXK"O/]GM8^S% M^29&S@_-H"B25=`CE`UF*F.'=$)39.EQQ`BO+!TJ&AM\VD#LWQV!MMCDD_SM M!!V,<3V7I8]/Y0A@Q/FU'`&([TM^Q7?/4I@E$$LN+1.3?4I-XT/P=TAO9`,J M]MZ4YD/_(7S01+@)[#+F^TJ6&$).8PU0%A@IH#\6^C8"*8`VP&JRP3?@FT1& M,'M:9L';P>)/R=?9FRI*'#60Q3VK8'RP]_C]^Q%J^.0`Y`&[X)T0^X?ZPQ'F M`:9_4[[(GE%O(J!9E"H"!.&(P4+>94[/Z6>[G-KD#/V(&,N*PXPPN=B^_.W9 M`[V7+$EB)1#16"E9<%#>`_6%BLN])$=`%ZGO"V)J^1:0H$CB(/C27\BZ9`O> M`VV7@3IHXM?RNI=+O$%2"S6!Q;UC);_O;XDR_.H=!$I\+D"M8[SO2ADRY")J M*9N3@[V.(*U1C(G26PZ.]"Q^VL+DXP]2JA?E`V2:>?IZ/TRG7^9/'J+@F46T7%P$@P0DC54P`$ M*-.6&\!YA-@0!"A$O/;U*EF`OTCV)3"Q-SE,7&-6&`-^/,+M8_'OJQR$@R!%F8YKY)9(O/6WG8>UX^%JF+M0J$X"V0J/<@=`PF M#Y^5D+$1" M!.,&A$$T),Z/19!L-"M>%EB.JL=D8K_0MH@AB3OZ*X=Z;$6NXZ#0#2G#O"S8 M--^`-$2CX>8PS)=3S$/R%(UZ`+Z6XH[OI1@_-/9!]DX"'L5>(3RQ]9?O&_@1 M$5P`J;ZU7MX2EYFGC/+M_Y`E_;\I(*K2YWC1@PB`!0R`Q8/B`71P;9`2*#JZ M!(Z.G<7YH#WPNWA3%%*.%[>&CL'OHUG`FJBF5$I*)*T$,;X>P=QO5BG6@U&2 M]"Q_Z,HP(<_/Y_2+#N0A\,`XHD/[C@%K&F^*96.\$0Q0#^3WG@;3$'^ M#XD&3\LC8P`Q;N"";!PJ*RN$88"CG_OQ79F"S&"2`5N8ID@7(`RSJLF8[/'% M+BN2,,FL'QS2XW=Z;!%6,:.&:T2!E!*,56B`#*3$L"$8(;/IS<2X`AD M*U>96KZ\7RMQR$?5BR["#_E[-TE4(TUR%-@NK`$T2ZZ+7,'EGH^QHNF2S`+" M)#.03C]EWX(3,9EQ_+/H!!R<'"B M*'^2)$>3(\JQT`?DPS>R"\6-:4RT)381=2@3G(FT-D&1[B.47I=1&#F^?`X* M.;F6Z$:K)>Z`^/C;S#=R^((`3``F@&>OR^CH2V?6!9]__,$*XU$R>3A(O"8* M*G&/I0*[8$Q/)M@+S%!N",T",\)D(?D/++DVO"FN_,B41<[W0>AP-1#<]'+F M.%66`DT"I=W20.D8U%UB!&E[1TMG@5"/D^ACO&ZR#;:$3,.4I/.1WD`AB#L> M)K&;/$:4+42LK-GJ"R0UFV3`R2 M]G27Z<2]GYKR6ZB0I`@H">J/)L.[(RL080B;%!+J*1.*N,X_)3<2/RB^S%!R M()V%@\&I8K-16`G?3%ON52:`6CZ@X':1@KB,#$`*+V6)[4A!(7S0SYGS6PUR M#;F7B\Z\@'6QK\?JM#/>#I.'M]Z4QCH0*1SWD1%/TY M!'^=0D'VH_;12@@+5'.>]MH)J,64X]V2<"CL+$K6(&^>5<+@99%RW=!&Q%4. M,1&(";X(Y(9SQBC"Y$,._L2.L3^]X,]O))#_HQ7P&*!D.\<5I99OE]CB`W9N M!$F!W$#,GO.PQ@EIU&!]8 M.E64G4D3YTB/"A`3`%&N!HE_1<4BY<2!Y.=+D/&M_'&* M'&$"@D^B9(OO6&GAQ$`Z)FF$\D8-(O91%LE#]&R4%5V3O,DS'XK0A3B7W&"" M("F9Z\'0GZR3)A#7NP)L"N.6YT%?)ISRL+?BG-J\#Z`'?D]PI0)@Y!CX%$H. M/F,"GD`/W_?271G\F^[)*1N.FC^>WEKO*"E]3/F1`8%_E<\?8#;1)S`'F`N^ M/%F08H#,9W40O-GBFPCX%U^0@L3:(1RI`F.7`0F%:.0JP%=Y_GQ]"FA2"F@ M0,>$H\^IX,=O[*D\3/DE&Y62KT[UP?BS.W,IFP*.*/D5PT\5F8@`"4K_'$HF M/PF?<$ZP8&``2DA1Y$]&/&^)/\!^X$)1G>GMPVG.#EV2Z,!-X52Q\V?L[&CB M]R:%*06$H`E3\E>F5$F6+[6)'X/Z9.OQ]TD&/!=>,Z=[+D!#IM.O0)A@I'PB M%".?'(+/9QQPAB@&!5^6"=>(M8[VG\LQR\CJBR:2(I^*S\A'9N=3DIFE+#C2 M(^=^FLQ`8?V/-LCAVW;"*7.:\DT`Y*UQFIE)[$=.&!D%WDU_XZ10-Q-W/$HN M-'.!5KY,Y3KR1T@NM$)R"LF-RD0IYDPR!KKG$W5"%%V.#,%+(;6P@;CS:1&" M`#-_OE!NYA\TOCF&5&WJ&[&4"DG^YZ/O\%F;A#6&*G>>;$D603>Q"U@A7$]> M]Z`#=\840,B,&!G?(R7Z$(F5]<#C8KT3<'@_-&+6`.4`=P"L0)DO\F?28X%R M!1^7Q\5+X\Q1@QGW^SJ6&!^4=\X*)S6S%%KZBR+N&&F?;SW;YVK0,A!M=#38 M.HJ%S4A'W^2R'&B?).E%.4.0\M!CI3RSJ`CK8PW>&R6>KTZ1XFU0-YF+S'X* M$[<*],,J9O/0HSDIS`FT"%><+U'X@YY`!WK48T7H8X33#2#8RS)V M*VF@!H.H)$W@F4@5/0*4'$^.73^V9\VS$RI&3!TN1">>X;^O'I13RAD/73_. M(1>82`230`NQ,4K5LRV6)UJ;E<.@W_G/R<<_M`P,_F2B34S-WUT04F`M@`7N M+1>6/&RZ"=,!(:"`C>`"%2%V M]E*C7[T(Y:`SV@FHI$@2^BA\Q$R!8W@2U1FHO`>&,_4,.4QK91Y2(QF='`7N M$Q6=,DV47CDRFKG-!`0V%#F=^,`:J88/YV>-O$KN_)0$&#[IIUNP+9E&W!LF M#XV:UH)TI)8OG$F-K/-M/LV#%,1HGQBA)RHHA'=.&&,'IT6@II90R3<1(`_@ M`7"2B$YA*$]OOD<12&U.,24+88#RJ.1S$4C58P5*%$./?LR3)1O0=QE0Y#B& M/5D$=KZU7H:OJIDA,).V!-&D+CX:`7@P=KG]/(YR0P.20+WFWS200^GR+/P= M#VV#;<\=9;YQ1KFN#%\^0L%\/#WRYW91_T?F&0WD-?MZF4F^)M`Q(^`2N&`0 M'1F`ATVGY[+42MAXQ#6N'=^/$$T3H;*2MJ@"I%\:]2R5V[[DYOD1<'G7^U2B M`FN)A$L:XZ`3BC`G712")0>>M,:UGIG229CO8VW2.\.,K-#[(0?3FRBVY!^^ M&)>",<8W9\[Q7@`P@U09$=/19D;`/#)`%@`-`"M*=Z18HJ3U(P, M"NB;!8(7IYY@Y&@J(`,`^-*F*8"E:)+RHE@0,&1HYS4DC@1`.H! M35,"J$>H*8!Q9SI?!)<*``^0'AW-QU24#1J6Y`SF)SD$6-/?H-KT/KGFPPJD M$3.G?(/-:8?O=J@&*$QF,U(*)+^!(:S454`WDXW?F(#14#+<,>P&%O,0IH&<8>'4LJ<_Q5[B M).$>'DEY!.#T^*GI7`(0`9(`30`O)=64/O@;W!M:38>GV0R+5LNSO)$C;6^O% M$:AZ3=-6$D+P7#C-!/!-4(UZ+@!\(I`4!0`#2/5=,.^HIE,H&9X@%2E?;/%5 M4(VH!TA7%P'.)8!%W8&V*+6H@\]17Q<54CCH!*/B")6G8U2_8PL5(&A)S!TR M"?6>,0A^ATL4<'KI?!^2_U)]U<_2)`#UN,@$0"/6"=&<*#V-)GQP"0KM=!)2 M*/V.K#Y='S:2;;J-7!1&$;^*"4$.XC138$I5O&2*"`N>7-0T``M`#3"FO!DF M)V>I<3TG`#&5[B@E)*#:*`%\SX48P#15;7HT30.L`,2IU-2\:>H0G#I-W0/` M4/>&[%0UP#D5A3H3``_1*3U[^M/[GN.4OQ"_K!,,SQ0`5%%7JH[@A4I/37X0 M%$NG,M2QWSY5P'4#;3-T$@*J^,E7*@I@:2KA.YQR')`J?3VN*7\A?7)1C7_" M.,NF*`!Q*@SUL*<_%3^&4Q.I:`;;!9*2;;I1-1ZD5(T'E53]H1G`;,H"&*ER M^$JJ8*DA5G$I1/?/A5,VI`4"+Z@H"'9%19:D: M##BJ2]4-*MET-4!3+:K>5#U[I5.EJN)AI9HHO.B1>5RJ<8)!@5"5JAI290', M`!:51U64J@6C4MA3K>S]5"<.F@^`:HIRY`A2/9N>525\5]6OWG0AJ>K,;"EQ M57VJ3U67:FYF@SK_E*F^4-&F:%5G)4K5)B`VU*@&5CMZ*PBI*EV5K'HVQ9L: M5;&JYU1]GT653H`0W)OF1?D=G%56JHJ2:^I(/?-E)C>K%@Q5Z!\5!?!(I:)* M4IT?E-2Y*J;319D$4!*V^$2HG-1>ZI)O\7?1TPL,/V&KHE7!*<74DNA8O0_T M4X&K8U4SP-Y0H^KU9(C1!)Y[,=6Z*2R5NIAD6*-N3UU^$\=78`)UPME*W'Z" M`1682$><7T,!9KC9A)2J,IV9CY<6H?JS,E@7!&LN^1Q]6CX%`!U@)5A11:]R M'&Q]6D`(8/74_=@#$*F6!@2B##%^1:/!Q2CSU":Z2E%\[U468`I@OCI=9:CD M"::`;-3C(M_1'MKCFV:.#ZB>A,>VJI)2E%C-S.KUYMX'F8C$G[%PX.W0O0 M%-6?N-3W*=@`Q4I/M5V$OV![J+\Z099$-],B-`+D#MJ+`$#:7G%/QA"&%/JQ M,@>>V\FY)`S2%>GQ2S(.]8BB',E=Y!SU$DD5$%*>2(=Z*5+Q(!E14?H0=>NM M3,.>G4=Y*R=C3W/IA+`!4#;^SC$!@2=/,$*X'+/IU[3#A.3\^.Q$FWXEO`3?H8!0`* MQQ23%$"S9GTP^_G5`PRH'L..>\58ZR&0VLC3*^Z=6+T]I2QQ<:#=2YAR M^M@`;CVCWEC/KC?6^^H!^)`")T`OJ]Q/;0K;O/9M25>4KKXDP('`LB>+%.JM M6U>4M@7!(&DO"%`'F$+.!R&`PK%OI>9#;+@%\"W*]!J6T;XD@7/0"IAHW1LR M6HNIND>`(*15_-C6._--6G^6PTA,*Q3@V'=K5;5B`X.<5\2@0*(@PTG-U/&1 M#DRN:,U-*TZO'ND&V&D2!I.)I,4MP89PU>J[K`#6`[^>*06V:DCQ#O#2B_C) M(2N#2H+,0`40Q.?34[46%<>:Y$>8JV"@U6IR-6W>`].)P5:I:VC/V.KT1.KY M!)6M:85(XQ1`U-JWJHU;Q@B50.&_C2&,T#)PNUPUBB+C+5BN]2N%<`\8+QO^2<'S+7" M1T]\XD<&Z,)`#-!SY;0B']<(%P%Z(V&`V[=T]1D6!MQZQ8$MX1,@?]GBDP*$ M#AZ1B#UJH;.1":!6]&8"+,^N7T^RA-@0G7#=F^]9"+J3GKTG@!3`"<`$X`R0 M%@A<'X=_9$(0-8`0$$L6][A[#`&Z9-@3,VEXC>M]6\$#)8(38-P2,"`'*$E2 M"W^6C$1,*]]1Z6JN$[%'*'_OU!?#74Q*B M&SM_2]J!7PD$ZP::(J@O]SH(C/<5]R8%0;YGYZR1I9A[%32F M%LN`/;T/805P"_MJ[<)*^9BN^D,A9;_512B!)05.(:NAX=7#GA$V"KJ'Q;7R M-_5]L=96DM@04Z`I2+>N1X>2BD?*7HP1%F@![57>#FNPCKXSJSOQP6<:C%V^ MP?*NA[TZ@MF!']CBFP*$!]8`:DJ"J\$56/B!=?0=`0I[@\RAY.+5(4"`8ZON M72^5=[U?:N?4LL?=J_05*1F.4\`M@`W@!"O7H^NI8*V3,\*VWN;5:6C.7/-9 M$I$`+H`6P!#`!4!Y+>T=#8("W,3,*^G5C5`E/+VB85>MJM>I)6-ORZKOVR5" M*#.PDD=-)VCR\JCD2_G)-G.O_LQ88.^5O]E`C-1`]\(4'T!$[.%P$7OF:\06 M'.D`IEC`7B3VS@&$.J[8B]H2;V=EEA],2V_M2:HMA)`2D6`K@W/,5^ M%E*Q=4^CWBN6;QF+S76>.$GFQ2P`Y`!=@XH!(7`AL M!;&59TZIWA-`T3I%=?6-9/$F*[WQ6BX5#<`ZR'6"_(Z+R-@Z*J8U91!5M#MJ M#'*O--&P)E>7(BTRWSD'N,** M_(`IJM!]9_[/*+`E_+_N!864#LNUZ]:1(HLA2+>B`'(`],&M@&7O;,I2I+N* M'V,1,X"%0*U#_BK=RT8V!3"$SJD<@)NU7:ISI"DB$N6L)8"_9M!13\`")6R. M%?.L1KT@@#GS58D@?)6N)$MZU901X$TR/+H<9.G%]9"CAP.T@=?97&+EN:15A,Z(TR.XF!Q$'" M/#>$J5$%HG`4C+@0_:YB9D&E[T'%`!;1TN?@+!ZL&SJ`ELN)0U#/63#?&U$N M(PZ0QUF@H-CTVE#BO(J^]<)[S4NP7@7S@OD"!`HX)$L"R4[S9(712SDF!%,B M10VIQ5EY1#P0>AH?1212"9NS>-%E1/%@'@$XG9N"]8(`'E00ZEB0!>KT!`V@ M`26<:46A)Y]SL&?V%`2>-YNG^,7O;-SSD[J;9'"6'>JEUX;D9='RRL><-:(R M]`H$P+;Z+.#3*GKBC`D0X*"#7=?-X8:`BPD^+1DJ:/.+XEDT0[W4QA>F2,ZZ M9RNTSMD!:YHAMDJ=C7$"1N$F+]B!YG;6STB@74A>9R\8*XBFYY6D/;N<-0+` M9_.>%UH2I:55;IH$S71N.LMQV5F802+VEFHO?7KV9/E\MDR4H8J61"F#0,X2 M:;FR1UI)YX"5R2I7K='R0#&I1$GLK-/3R9IDE>H%"J:$;\^')(J6?.IA5,/* M_U:#`E7PX)86W'"K'-).:-^S%EKY+)^V3&N?Y8'Z1=]ZZM(U+7$`TFE*]4L^ M"O=[NE5M)*`R?BC^]-%N:=\/7MH)+9@V4#M@Q=0R:2VIM%4EX?(!1+LEK*4N M**^SEUFJWIZRMHJ@]:5:9)EZ;U::(NSTHO$--/")K[4`6BD1>P% M`9H`10"OX*"@UA%W/(.&"(BU2-#IK'T630O5P[]E9X,`M,03:]RR2(KN9+8. M"L&?!K^RIR9[4X`L@!/@U#<%2`)<'LT";U/Z[&DV:KHAM)8. M&VFH8ST%P!$5XS"?'7Y^''V2QT]F+8(5?[I<]:0"!$&IUT%1ZN=4P@>P50#8 M.O:HBU/JX"956)L2L/%5/C5[C-,O*M^`9+N"X`LH3C>**=N+K7<3P+^7H6610D%J`)` M]ETTQ6;1>03S) M:\.10S]HJ6\R;LM\:`D,#Y6G`-9'Z-Z6&TE`O;)6:YV5*U$OZ;MR[LF=^T$G2)T5T4N@:ZP#.]_`44P`?:!*@U)<$>`(X`4BV+3*E[980!#A'?=P* M*BV0P$Z+HMX/NZ@D[%.2;<62J]*1[;&UW#DUN(?&#]N,6#[O+-PS/-L>/+M& M5N&O9L'VHKLUA4@&,,^**5,`C-"XGA#@NG>;6#NR/<&27%.,*TDQ#)NG97!> M#=>;)%-M)FP2R[K(1(H";]T'DL,98OR6FQK3XU`B'$^*7;UJ986QO;G5/!)6 M^_2=:X2B9)MOJN='3*!^&A-^)=/1+=ARZ->Q98[F#"&:ED6T(RY3ZDAU/61J M!8N<7\%(K2P6KZ@'E=D611F<;C[Z(7;2*TE]Y#A\/ANEHSUW+=)T98NS#8@^ M9N4`WEMG"4L@[@@>`"5&7Z^"TN=]G*43+W#0OBOX:>QW4OR M"D&)U\Q:J6F0$KB\JM/S"3B4H\V!+J]U:#?]L]7QFWPT>AW1+* M2,^7WK\BHI6R!@R8+K]P9P)6AZLE..."4Z&4:UPFI:QO;=O!92?9 M;BVT%QV17]=!0SMR=`+H4(627%DQJ4:R05LMU.$>^5*Y(4\+H/?T-LH:7!*> M!%H`H=N?P.\QDNL5G!'*/`>YE5R3`!H7D[LE9!_&%V%^W]N40MQ1*0FCI8$B M*U2@:`;Z[*?VJ,=$[>&&-GNB9LMN)Z)4Y)F5G!00;E=9!DO$'E7A=QDS>#^^ M`,.3!(/9Y,02&QC#11WV;UN(`DL_9G$1R^83-W"*I1"SE:%Z>Y$EM8;N'4XON?=[^*Z6FU$QL9_E4(X@BE'WR\Y`7ZU"]W(*K*VKV>L10`6&9*U(,<:;0?U@]H$L-N2 M:D6O"<(TH+ITBMMAU92B#K&U[,J+YS9!#TC:`^1>24>#_D%Q+4.O17CBF^\- M`;RUX-G58$]W__A+NM>6]O2U_%I_+>"%N<9)Z`^4#;Y4DF M=CFH^%G&[L36B[K#;0-4396GAM3(ZL>V9VOL8YR._@ZHX-R2K5TR9%NP+>ZR M;"=_+]OFKA3TA#JTO>CY:0T-(5BD[=?3MW9 M!-ZQ6\*NK9V@:5C*M"**;56X%EENI0,W3(H^K3#.2-NVYLD4H>96+]!8K=O. M">B[\;UIYN)6+-FWQ7$&N6I.C%O=S*;D>5(H=`ZWSO">"T;=H^;3V3F;VI;=7V:@N\[3KT M:1FT3+T,+WY7?KO9Z_!>.Q^W(-[5K6EPQ`LB?-W:(T^\8,.4ZB<7=ZN[Y=WZ M;O^WB=/U:'S/@#OT-'W&#/V>NJY?$6;`>Y75LM(7!6(_F?;1G* M!I5\*LR[WLA4_"@;E)].+=V45CZ&[KYCZK?9C?<=(P>YSE,QKPLWS)KOBQN2 M)P\# M%WV94M"W8#YS)OG2U'6>%PV]8,+W8)LS6JE>M'D"$"N@/TU(IK@TUD=&=>)* M".Z("3Z&+C"%IBA96"/\#'V1:UU:(BB1*ZODLYH^3F.E+;\Q[J5OD,O=3>,6 M:9V0*$`XKK/S!9BG75AR>9>"*]R1P#57#YOJ_0>Z&T=]B(=7*4.7^T#RPT;: M=,NH+=Q%;U&P0>!,K4?NYV^%)]/5R6GYP3V4JCK'W& M]:Z]JMP.+IR!K?HM+.X)7[$"Q%=I9:&OSEL"Y.<&_8I[;,GH7[MW9OONE?"5 M_0RYFESYXB)SL.>!A8#"!0.]I-8I:(;QJX.@6<,MX%%7T\@*I`$6%4>FG]3_;%/7Y&MP0%)N>#=[>K[W(J\3-[G4+?WI M'1%[?H)7YJV0@+H8&+*V^*BZS5+++AD@3-'K=?<:>_6\,%\K88E@#1C<-$9^ M;W<^MK[I'H?/Y:N6I0E\1Z>!M%+2IK=T2GCS!?"I`(:H3D^KHD)V^,=]Q.KZ M&[>ZJ<,TXP#2&4E/#&Y:!+*A0C\3(4-W"G7]5!P$?%>:'5R;0,[0GXC9F^PF M#Q^=MU*GH'A`,-CWS?D^?/>$T\"E8N%7UC?Q_32.;;N\D]I=Y^-V/^G4:Q`$ M<1F`I4H2KY'7D6GP52\.*\(3F8>G7V)O#1?W"`K^%]SUHYE/0J5OK!/!I.K$`(%2JK74RKN<"*/_N M#9L`+,KP;P@5?1M`A-86<8F0"U^_+^`7XKN=;!E$"MZ42%CRXYJPEGN_]?K6 M&PV&U5Z2;?HA9W@&+30DI`!?N&=?NV0CVR[MA0<\NO`/E"4<^ZX5)+Z&J2!.S65:A2 M+.6Z;E;F8E]V"FACO>O&3-U'1D[#K&&SK^NK!=8*@`^[A%V\:"@762N=3>S: M9^N?'=KMZ6.W*CK9RXSJ:4NT,\J0H1[37SA=DE[(\K=KG&6W_&PE:V" M@>^?PUU(H5%/7TD'T)SJ;'F[`MN3;64@9MNUU=PN=[.[*%LI*"6X;BL3+(R1 M=TVYYEVH;8Q7O9=F6XTZS\MVOK4O7OFM;W?'&%L&]9ML1)BA26ZNV79BV M;4&A6D,!;]9VY\-6'06O;(VL>-MB(#Y7]XO)-/$>;,6E;>!/J:DT#7PC?'J2 M!DNCGCWLKS9QAUA-2=J6%2E^HLIVH0RB3WN[K4PV>:D`O=O?K>:6U`#AO#66 M=OZY1*62U6"%4GD+>43HWG?7?#.8A^^2KXPK].3>ANMG0R2 M;%.\/=OFGXEV>[L:%,^:;,V@6N#BK"!D,Y)#Y=!" M@N.6A-!G*C;3'\S"/6V21(\&%<7"98*6(2P'=@^V->&##U[@K6M,%8K?R_S- M-=NY][T2WR28!\G"]37*`M]\/CZY7^%O`@N8!)2N)&.WS^"&8`'48OET#(%V M+[7"8K\.KED@9YBBM2)B#9NXS,<<9P-CM8B<0SRN<9.)B/O>8T^@`NZ:[T),.6/"6`!EO5F M3_'"58!1(_,2M0GQ6_8Y-KN(DH&:Y$BW:UI-J?&A([K`/LDOL$WXJ1<34.D" M@3."&4:S9S_XZ5NP_1"2(1.\;=QV:?3Q,>RN)*!J=4N!%%#XH%"WHDO470_* M,WU[Y\TH8M.PI&!M[Q:2JX@$UO\I M@5FF<@>7Z<]QSAIT#''=!?:Z4^"0GE\7L'L%+M9J@8^UAUW@\-ATY.@7E7$R M]X:.3D_W+\0449L,QNQ:,W^C\$7^KNXRJ0LNK#`R-L.+VMK"Z!*R!,D-'MTS-U0<'4W2"QAU-RV M;*/$3.+JKDZ!9NO=C9Z&=R_!7-H6H6VW3VDDAA1PB=FS1=N])EX8:?O>/06[ M>,^[45N_'BN8QCO@;3AR;0^H\]T987W7Y5<+Y@F+;//!<,E=<']W)/PM+!`& M>`FW\L[.JF>1J9=62P8K>)G#W`]R`Q64\K\Q)4,W!J(*?0O\ M"BNB4CVHGQ51B_N/-4/*+2.,50%E\*V8/NDVC-^>(RVFZD8JL4J056SF^PY+ M%G*'U=0!5^W6ME"':^\7,MXKF6/5;LFA?VZ M3Z6F8-X#ZO26W%D0)ES:@3>Z3LI4H)946$LKX%>(#=>W9[Z(9QAV@$LG72N.0E@%KE"/@5L&1@V6,K.V_BV$(%5VWYD2C>O! M%ON(AL?*L`*7PP>Z[(IJ<#&$0DP0+KC/^[H>?`.3-'>9.,B.)T%4P@GR]!/3 M(O%['5P[0;FVX,K[B7J/TB#NDE#!Z=UM\B.*^:=QQP3=DE89: M:+D/0MRR8$R51XQR1/;F?\]\O.(/)$B12'DU)&9N_&)CZE(T#6"6343> M.PV2AM]))([36;CM=775^02AV.'4YE11677`@>>%,Z#[<)3#YG).='$VXUT;_*82FBYM%T\ M3XT`@E(E;E:OAYOJF^\E#V_%.&!RY#:53HK,7!0C2N62<@1:KLZ/H5NKH/,Y M"P2E\3XLH:MO\!?>[>Z&*5!ZI5/VP)80WOL^1@$X7M,GGF++0/T85IH\YA[7 M5V,#+5KEI?6X78L_]NI!CZW'7-D`\JN4>]S=[1\?^>Z\6\(5`+G/III`+MI6 M3_NX`$"WGL:RDSF4;`E^6L'&Y('Y(#78X/?/PQ['`C_%%&0",B%W:>GTU"`# M`)'$(4=+*ZV5,TCRW7$^*$E[&,A$G][8S,)4R$[>3>0#!@&0`*S`&:`F''&%"4%//IM=X"VK+/1YK M":^&.E_'(.[Q='QW%%2RC?U\F-;2J347Z2>;91OW<85Z"-5,+B5R[3LHY(?. M']]`,,T(#:6T20?JKO$W^ M+S5[:%W8Y.%P2EL65G4:)#-^16-5X!0R?4FO+3K*4@.Z18"!;D%7`AS56PVW MAB'&(@?UZ)90`PQ%?@]NAW>O!=36H^<3YU?QG'M>3CF+C%[MYT*2]-NK77&> M0+E[-]`\KA>XL,HT+NHY/=?#CL>T(`^XZV#K8QM+$5W&DC]DJ@98BE@7O"(; M-^.'2WG`10E&:CMY/GE)//V]A^F^1F"<7GU83H``Q`\?AFNU>-UZ:CT50+RK%1!7 M@0.[PUHL,"@703S\5"8'AYNT40!-9X'O99:=9=/6`Z5Z(V`K+5DT2]NII&96 M/%V?Z--1Y1[XD^L'CNVV=Q?!A&"?;76W1@Q)500/@GV[$%N,LD8Y$BQCU7R0 M@S&=;V)5<'IWQOLB#KI".&/!"-[!\02P3PPIS@4G#]&V*67@XEEOR6L.MA8[ M>=7!D55@A/MX>!LF!/W6E%M^6F4$,`]TIYQX7"`X-%N)M^2#HPB7T-ME%4_> M%46][\6&9]H88'BLU-SB30B*:4^>WDK2"'DJ2$K.$/M\,T]XZ*!8**PO5+4!;PSBQRG"A20S+;FY@X%HKUIV`:@`^N2> MAG^WJ>%6,J:/-0Q1KH"H#YV>L>%S7VHVKL<]N",L]6S#',)FI.EP-ZQQ)/.8 MEBO*.F(SK9.VP&<<%OW6/F^@209;W^PT8WE/4F\-P!0S6G?NAAP.`[V!MXCYYY^``P:N&Q/F!)#-T* M7'F^5U,P,6UW29P8$-EV@I6[2F)-<).YNGLH"#*O;)_$)MLQ,67XZE=!_NX: M;??'8.(K,9,Y9@NTW6F>B;W,!V8P,\NV2SMF_MGB)\_,7N*?,7VU?]HF7MJB M@E^\Z%TY,5&99$O@/2KCB6W+8%M:\)*PJ3S1#!2+<'G!NE2,9J=0/8C:C7N^ M/DG,G$,^LX4W7I@K-?#*@O?$HJO77L5W#`4 M3"MS^YP)AG3B\%`2B>$VS%5H6+Q7@3:P8C1/+>->[=.+O[4OX-/NFG5/& M]9+(SL&AHWZ9I4L6C1F^:8^WR6$L*'R4>)P*)/?F<]&2.\W>[TD@:#PM9N0F M<1V+^P[0;14WPNA`)=U:1&&-$)?!1=YF\,46C$LO#MZN M4*>6`E$5J#SY5*Q$GF:NA\.P15(5LR$4DVA_[(IV!''))D(.<6&9?,N1G?-2 ME7.W5F5L,5:9QE<:CA#3$2N0W4OCK:DWV60(:$W+"H(O/V^"TE^;D`IH5,YBCQB-D'F M8=V/Z4C':4CP)C@2Y)9V='W'AV,?9R)0N=JU3&?^A?V1KE?N96\0S/J:%%)Z M5T7*P4+FKBCJ M<-T`>`#,#I^'?5PSKC#WDBOAFR-?#?NK+`"[ZM[P+0P#[B06 MD_&B9H$BJ-T9L7M1+N6>.3K\,1JPEL?NM2MX<*PUS,8`WH$@B;^46!*Q8'IVJ`5@`=4)8KANX>=S'K1YG)K6$T^,'6^$W?)'6(W;\^7Y?K`A`[3 M^."^5]`#ZGS//'D"#NIV<5^WL\1_H*P/,5H?(.IN,JV`M&0B:?(B6>Z7Q+WXES,;6("5C@:/GFDXE0P!-NB*EA.Z$&4&H`;U-)M:Y@RF M2Z^3%+_:L)`@@/@!U@U?GI&T&`=G"9N!WV%17A#/5E.Z\,]JPLPLI[MN& M"EFYEMUI,KHA3UJ4/-'B)OG/SN=W;=TV^-O5%?K-DP65Z6&!(1"8Y?@F'0+# MA].9`>:$(T'971KVY#CHA_N:"N;^L!W77BH%;BA#F`G$$V8#,9)V"YP@;D+K MEJ,`A]J_B MF6W-<^*7LE$9%OQG3BI?9VV-3&5SL5ARZ?P/9*::/17-(X83+R-JJS_'[&B[FHC\YU\+X#)CR(+LYVEONB?U$D@-KTF.G4.T7C M1?4"[X.'-.6T%9W?DRSXG"F[:66#,982RX?OM.`6F@&T@X*:0-P1XLB=#8UB MA1^"\=-8]-R/22B,[IP.4O>NP&(X'_B4@N@R)ERF&N>@V[W#+]FW9WP19!Z/ M`7)[RN"$M*[Y@0C*E0G32PL&4DT M@%M\/6A.8*"4-OQ:'D*;C&7+RT3:-D@L_8D.P\_NF"/">Z@>D%ICW:X.=-'E*:@9HGIX#9TB+$-/,^?)_L%Z<@#0:3QL1NO>*RE\_>65Z'\YH'Q'$#`G M@0O*=6A^H!-XP8<]&!WHFS8O3![@3'*$4IPPR+:EHJJ?403 M^,C%B&5@:J0X;5CK13:"$3&FI619*Q^X5_MB#D7'F(6W^TGM+I?5(CT$M2FK MHG/+K.@6)9^R20F+=@/4_?:&SF82[]XP"6@FO@3#B+'$4^83X0'5$QQEUC+; MIVV$]&(H<9NYNIN?!A'"F6VV:>:<[74W[C@F-OO]"A_4:>(O(?*N.87L3>8%RWCF^\ZBH^+R^&:-`N7&-TV9B$[>MV8R.C^IQ=Z[%?=A3B* M=OO,T<=S-+UX;HNDK/MEHY/!UV1"IV&F7<0#Z\*O;%-=6 M0*R@/\13;0U0QJ?D0X`RBUW%*EAH,?MW5ASBZNLUJ%>`,0@Q;OCO3A#7"U)J MI@>TLN=;J&67(OTY=O7*A?5]Q=NAIVCW/G"D[.M=J:O31T98<>UO*AV\;0"O M>%N\YEUI]"J8&CV8+E4Z/?G30#T>@T(TP5NE!1_;/.?32NKM-'P0+OS$13Q\ M/L'3I&'C:866'7U5!MZZG'G0W`.)8?PY3/HA')DR(P6_*5.5('`2F]DJC91: M$&&/?T3^X3?ZN>DVU%!6=X=\1F=N[Y712XARE8X>]AJ1OU92M=L6TOP>!/"E M/T'.X%KR:1*VO2Q&E!5;1(\`68%.(X5X73H5UM^:,=NX6\@OM9JZE^O2XS<; M*L=^3,[*KJ-O1(D0YCAH*(R?A54EZ'VW8VQ/(`4B0*..%P$XPJAQ(KBJ%=5" MC*ND3E\5-==S0@K?^U9/,R'5S.'D*`OS=D"4U6TR&N&'3V.V,NVUR$@(!-Z* MGFV=&^;*+CC:2%T\KB<>];:\V>G.J1HY^L=?+)/FC!.3N\J;,0W`^S<`->+" M![N^;,4.;JV"M.N(U.G=>IG/?^'E^%@$YH+MXD[ECIBPP032L#,<+NC.F6-FNI;ADV M"&E[:MJ8M6,@LO?US,XB%`?0&E_S8L::A=NWG8;6>_-ZSI*1"1.X.MU\?@ZX MD`_4Y[[(7O`WV(MH=W^_1UJA7_M[*((:FV^:+;61OL@L6#OI)-='-L^?!Z/92RER`+9P2 M+K<)CT(A0;+28OV-=G7BJM'6M`)VL)?ZOGB_W>\V"$*-3=R3`+UZW2R[Y:+Z M)I&5K5ZTY9A2BH@2W1FN1,NC,='`:.\3;&0WLFG=]3DL&H<391GD=7D$" M*:G34VJC']W5TKHRIHN2<;76CX%3G M]5!ZCE>+:.DT_%#!AF`+&2;8UL7_=0<0@[WSLP_K"<`SLF0)=G&1S8".$&%? ML/W'EKW2:4T`Q$J`ZP"JL-]Z>J8'MI"6A$WSPV&7)E'8#6`)-JJ23ET@<&'W MCQVO1KW2J6NLB%W#3NL)L1&U&`HC-K`W:RM4R1D*)J=\C4B_,0J@8@BS1"?# M!T^M(\4V;]:PQ",^.&(NMD(\31;+UAC"(: M+Z\UMCZ&P-YULNPICNQZ^R1^W6D=075YWMO5BTI>#!&*CTW+\[M69IU&WDXJ ML,^/+\!T@:MZ6?F--L],:@2VD6FE*V=PKBU6I8K2P\". M]#S-9TX[*S#W5MQ;9NDR*36"O-3M]<19^5@EU%]C`0>Y?(`#]6.`#]"/!@2* M!?W7%X===6.PN M?(&-3,KJ9?0OZ*LC)%ZKT4A^S%\L@95O:"C4,_KB%>F-T;Z!HX580JFO3CCR MLE&S/ME8M6I3$^KLVQO>1T>*:5YI;P]032G\\U^;YCB5V,"#=2(WCOOKA..& M3W<#'@,AY:W8D,WS].P!K`O:K+Y@,ACP@(@=/=((;\VO-&C4<-T:H5L!5DY/ M)""<86FB7NF6++WC@RTS(]N`&4ZU]"R"X^@(5!`?HB.4U=_:-&)2!DP!!A67 MD\.ZI\TTM%EWOZS6I1?O;1""TSTT[VY:GQK7]4W;AX'3&MDX*X(9#TT%``]8 A"T8"^[ZVP$`A!BA/:`LD%$``1,)\`M$@H;`YZ`L\`4`5 ` end From owner-mpi-ptop@CS.UTK.EDU Wed Jan 27 05:38:43 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA09209; Wed, 27 Jan 93 05:38:43 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA25168; Wed, 27 Jan 93 05:38:24 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Wed, 27 Jan 1993 05:38:23 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA25160; Wed, 27 Jan 93 05:38:18 -0500 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA16655 (5.65c/IDA-1.4.4 for ); Wed, 27 Jan 1993 11:37:24 +0100 Received: by f1neuman.gmd.de id AA16117; Wed, 27 Jan 1993 11:36:31 +0100 Date: Wed, 27 Jan 1993 11:36:31 +0100 From: Rolf.Hempel@gmd.de Message-Id: <9301271036.AA16117@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: working document Cc: gmap10@f1neuman.gmd.de In an email which he sent to me directly, Tom Henderson made some comments on our working document on process topologies. He wrote: > In section 4 (Overview...), all routines except MPI_INQGR and > MPI_MAKDIM are mentioned. Is this intentional? (I'm not objecting, > just checking to see if this is what you intended.) The omission of MPI_INQGR and MPI_MAKDIM in the overview section was not intentional, but I didn't try to include all single functions in the chapter, either. What I had in mind was a more general overview of the functionality. Since it now turnes out that only two routines are not referred to explicitely, I think we should add them for the sake of completeness. Here is my proposal for an addition at the end of the first paragraph: ------------------- The function MPI_INQMAP is provided for inquiring about the topology of a process group, whereas MPI_INQCART and MPI_INQGR return the size of the process structure and the position of the calling process in a cartesian or graph topology, respectively. In a portable program the number of available processors usually is not known at compile time. In the case of a cartesian topology the function MPI_MAKDIM helps in setting the numbers of processes per coodinate direction, depending on the number of processes in the group and optional constraints which can be specified by the user. ------------------- Okay? Tom also found the following mis-spelling: > Routine MPI_INQCART is spelled MPI_INQCRT in section 4. Which is the > correct spelling? I do not know what the correct spelling is, since the names I used are only first suggestions, and we still have to discuss them. Of course they should be consistent, though, so I changed everything to MPI_CART for now. Finally, Tom made the following suggestion: > Would it be a good idea to begin the subcommittee session with a > combined meeting of the Process Topology and Collective Communications > subcommittees? Maybe we could all decide exactly what a "group" is. I think this is a good idea. At the Dallas meeting we already decided that the two subcommittees should cooperate in the handling of groups. Rolf From owner-mpi-ptop@CS.UTK.EDU Mon Feb 1 12:32:41 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA29157; Mon, 1 Feb 93 12:32:41 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA25164; Mon, 1 Feb 93 12:32:07 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Mon, 1 Feb 1993 12:32:06 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from super.super.org by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA25155; Mon, 1 Feb 93 12:32:02 -0500 Received: from b125 (b125.super.org) by super.super.org (4.1/SMI-4.1) id AA10551; Mon, 1 Feb 93 12:31:59 EST Received: by b125 (4.1/SMI-4.1) id AA09570; Mon, 1 Feb 93 12:31:59 EST Date: Mon, 1 Feb 93 12:31:59 EST From: lederman@b125.super.org (Steve Huss-Lederman) Message-Id: <9302011731.AA09570@b125> To: mpi-ptop@cs.utk.edu Subject: comments Well, I've finally had a chance to read over all the topology subcommittee mail and look over Rolf's latest draft of 1/26/93. Some of what follows is small details, but I thought I would send it to everyone because you can just skip it over. Also, the volume of mail isn't yet overloading my disk space :-). The question has been discussed about how to deal with two groups with the same membership but different topologies. A few thoughts. I think a useful aspect of Marc's proposal is the relating of processes between the groups. I see no conflict here with Rolf's proposal to optimally map each group. The thing I don't yet see how to do this with the proposed functions. For example, you can use MPI_CART to define a 2D grid of a group and then use it again on a different group with the same members to define a 3D torus. Then, MPI_INQPID can be used to get the PID of the logical (x,y) coordinate for the 2D grid or (x,y,z) for the 3D torus. But suppose you need to talk between these two topologies. For example, if you wanted to take your 3D torus which has a 3D tensor and project it onto the 2D grid so the indices aligned. One alternative is that you have to use MPI_PARTC to create the 2D grid from the 3D torus. This then raises the issue of how they are related. I don't see anything in the standard that says how this is done. Will the group ordering stay the same? For example, if a node has logical process (3,3) in a 2D grid and you MPI_PARTC it into rows and columns, will the third member of each group still have what was logical process (3,3). Am I making sense? A second alternative is to figure out the mapping between topologies. The way I would do that with the current definition would be to get the PID for one topology and search all the PIDs in the other to figure out which was the same. Is this the easiest way? If it is, part of what Marc was suggesting may have been an easier way to get at this. I think it would be useful to allow the user a way to relate the different topologies. This also relates to the question of doing a collective communication before and after a topology definition. If the user could determine which was the "first" node that received the result of a summation before and after the topology s/he could relate the two results. I have worried some about the definition of trees. I have talked to a couple of people here about this issue. I pass along several points that were raised. Trees are more complicated than tori. If you restrict yourself to balanced binary trees with 2^n members, then things are easier, but there are applications (probably the majority) that want more general features. A simple example of the problem is that if the tree is not balanced. Determining if an logical address in the tree actually exists requires in depth knowledge of the tree and not just a couple of values as in a torus. There are also applications where the parent has more than 2 children. Another attribute of algorithms that use trees is that they often change the number of members and the depth. They often add children (leaves) and after doing this enough need to re-balance the tree because some some nodes will have too much work. This raises the issue of the cost of adding members to a tree whereas a torus generally has a fixed size for the execution of the code (or does not change frequently). There are published algorithms for embedding a tree in another graph (such as linear, etc.) and they are much better than random. What does all this mean? Trees are used, can be mapped intelligently but are more difficult than tori. I think they have merit but I do not have the expertise to propose a standard. Does anyone else think they do? If not, I would propose we either solicit help to do this or let it go to be dealt with as a potential add on feature later. As a general comment, I think the error return codes for the routines could be more informative. For example, MPI_CART could return: -1: Sum(dim(i)) > # proc in group -2: period specification illegal -99: other error Do people think that multiple error codes that specify the type of error more specifically are useful? I am will to take a first shot at the other routines if we agree to this idea. Now some details about the draft: p. 2, par. 6: There is no need for opening a channel explicitly. I would change this to: There is no requirement for opening a channel explicitly. (I would do this because channels might become part of the standard.) p. 3, top: On the other hand, a more precise specification.... I would change this to: Additionally, a more precise specification.... The sentence before implies weights are unnecessary in practice and this sentence implies they add complexity. The "On the other hand" implies to me they are opposites but I think they imply the same thing: don't have weights. p. 3, first complete paragraph: I would something like: Also, the mapping of grids and tori is generally an easier problem then that of general graphs. Thus, it is desirable to do the mapping as grids and tori rather than obscuring the information in a graph. p. 5, top: Add something like: For example, the exchange subroutine might contain MPI_SEND(...,neigh_pid(1),...) to send to (i-1,j). (We can be more precise when MPI_SEND is finished). p. 6, fig. 1: I think it would be good style to check for error return codes on all of the MPI calls. Also, it might be slightly clearer to replace the value 4 with a parameter. p. 9, MPI_CART: I would make the hypercube discussion more explicit. change: Note that a hypercube is a torus with 2 processes per coordinate direction to Note that a n-dimension hypercube is a torus with 2 processes per coordinate direction and MPI_PER. Thus, ndim=n, dims(1:n)=2 and period(1:n)=MPI_PER. p. 13: MPI_INQMAP In spirit with the rest of the document and MPI, the routine should return "symbolic" values instead of specific integers. Thus, MPI_NOMAP: no mapping MPI_CARTMAP: cartesian map MPI_GRAPHMAP: graph mapping MPI_ERRORMAP: error occurred These, of course, would actually be integer values that are defined by the system. p. 15, MPI_INQGR The value returned is from 1 to nodes. It seems that lots of other places in MPI it is zero up numbering. Should we switch to 0 to nodes-1? Wow...What a lot to say. I probably owe everyone a beer in Dallas who actually read all the way to here :-) Steve From owner-mpi-ptop@CS.UTK.EDU Mon Feb 1 13:26:29 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA29899; Mon, 1 Feb 93 13:26:29 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA28026; Mon, 1 Feb 93 13:26:10 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Mon, 1 Feb 1993 13:26:09 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from fslg8.fsl.noaa.gov by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA28018; Mon, 1 Feb 93 13:26:07 -0500 Received: by fslg8.fsl.noaa.gov (5.57/Ultrix3.0-C) id AA25619; Mon, 1 Feb 93 18:26:04 GMT Received: by macaw.fsl.noaa.gov (4.1/SMI-4.1) id AA00489; Mon, 1 Feb 93 11:24:59 MST Date: Mon, 1 Feb 93 11:24:59 MST From: hender@macaw.fsl.noaa.gov (Tom Henderson) Message-Id: <9302011824.AA00489@macaw.fsl.noaa.gov> To: mpi-ptop@cs.utk.edu Subject: comments on Steve's comments Here's my two-cents worth of comments on Steve's (probably at least 35 cents worth of) comments... > I think they have merit but I do not have the > expertise to propose a standard. Does anyone else think they do? If > not, I would propose we either solicit help to do this or let it go to > be dealt with as a potential add on feature later. I agree. I certainly don't have this expertise. > As a general comment, I think the error return codes for the routines > could be more informative. For example, MPI_CART could return: > > -1: Sum(dim(i)) > # proc in group > -2: period specification illegal > -99: other error > > Do people think that multiple error codes that specify the type of > error more specifically are useful? I am will to take a first shot at > the other routines if we agree to this idea. Yes! The more detailed error codes are, the better chance I'll have at finding my stupid errors. Also, error codes should be symbolic (as Steve mentions later). > > Wow...What a lot to say. I probably owe everyone a beer in Dallas who > actually read all the way to here :-) > > Steve > Yup, I read it all! Are you serious about this beer thing? :) Tom From owner-mpi-ptop@CS.UTK.EDU Tue Feb 2 10:20:12 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA20600; Tue, 2 Feb 93 10:20:12 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA21021; Tue, 2 Feb 93 10:19:46 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Tue, 2 Feb 1993 10:19:45 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA21013; Tue, 2 Feb 93 10:19:37 -0500 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA01293 (5.65c/IDA-1.4.4 for ); Tue, 2 Feb 1993 16:18:14 +0100 Received: by f1neuman.gmd.de id AA11127; Tue, 2 Feb 1993 16:17:54 +0100 Date: Tue, 2 Feb 1993 16:17:54 +0100 From: Rolf.Hempel@gmd.de Message-Id: <9302021517.AA11127@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: Re: comments Cc: gmap10@f1neuman.gmd.de It seems the discussion in our working group got some new impetus thanks to Steve's mail. I would like to add my comments, but since the Dollar rate isn't bad these days, they are probably less than one cent worth. I start with the easiest topics, and then proceed to the more difficult ones. 1. I appreciate the corrections to the text. I will work them into the working document. Almost the only problem I have is the suggestion to check the error codes in the example in figure 1. I agree that it would improve the style, but I did it this way because everything should fit onto one page. I'll think about it. 2. I wondered why nobody complained about the (1,...,nodes) numbering in my graph proposal earlier. We had lots of discussions about this in the context of PARMACS. The problem is that a zero up numbering for a Fortran programmer is at least as ugly as the 1 up numbering is for a C programmer. If we want to keep compatibility of the Fortran and C versions of MPI on a subroutine basis, we cannot please everybody. In PARMACS we ended up with thinking that the C programmers tend to be more flexible, so they can take the burdon. Any suggestions? 3. I agree that more specific error information is required. However, this is a more fundamental question for MPI and does not only relate to process topologies. A list of possible user errors (inconsistent parameters) would certainly be useful, so any suggestions are welcome. 4. I have the same problems with trees as Steve has. I know that they are used, and that mapping them is fairly well understood. I just don't know in which form they would be best suited for the applications people. Trees of the most general form are probably best covered by the graph functions, anyway (e.g., varying numbers of children, to be specified on a node-to-node basis). We need much more information from applications people before we can s t a n d a r d i z e this stuff. Perhaps we should leave this as an open issue for a future MPI-2, and just mark it as such in the document. 5. I now come to the most difficult point: the alignment of topologies. I first would like to comment on Steve's examples. If a group is assigned a 3D grid topology which is overlaid by a 2D grid, then the latter is not necessarily a projection of the former. This is because the 2D grid can have as many nodes as the whole 3D one, and not as many as a face of it. Let's take an example: the group has 64 processes, the 3D grid is of size (4,4,4), and the 2D grid could be of size(4,16). Of course one could set the size of the 2D grid also to (4,4) which would mean most processes are not represented. Such a projection could then be constructed by using MPI_PARTC. Steve's example contains a misunderstanding here: if a 2D grid (4,4) is split into columns (4), then only the third column contains the process which is (3,3) in the 2D structure. The third elements in the other subgroups are (3,1), (3,2), and (3,4). We have to be very precise in what we mean with topology alignment. If we just want to be able to communicate between the two topologies, then MPI_INQPID is sufficient. Any process in the 2D grid can look up the pids of processes in the 3D grid by specifying its group id and the 3D process coordinates. As I understood Marc, however, he wanted much more than that: he wants to be able to specify that the process (3,2) in the 2D grid is the same as (4,2,3) in the 3D grid. In this situation you can only do the mapping optimally for one topology, and construct the other one as a kind of indirection into the first one. The alternative would be to specify both topologies and their alignment at once, and try to optimize the mapping under all these constraints. Do others think that this is what Marc had in mind? 6. I just want to say that I read all of Steve's message very carefully. If the word spreads about this beer thing, I guess this one will become the message read by most people in the whole MPI effort. Cheers, Rolf From owner-mpi-ptop@CS.UTK.EDU Tue Feb 2 13:58:35 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA25972; Tue, 2 Feb 93 13:58:35 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA02369; Tue, 2 Feb 93 13:58:11 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Tue, 2 Feb 1993 13:58:11 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from super.super.org by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA02361; Tue, 2 Feb 93 13:58:09 -0500 Received: from b125 (b125.super.org) by super.super.org (4.1/SMI-4.1) id AA01734; Tue, 2 Feb 93 13:58:08 EST Received: by b125 (4.1/SMI-4.1) id AA09873; Tue, 2 Feb 93 13:58:07 EST Date: Tue, 2 Feb 93 13:58:07 EST From: lederman@b125.super.org (Steve Huss-Lederman) Message-Id: <9302021858.AA09873@b125> To: mpi-ptop@cs.utk.edu In-Reply-To: Rolf.Hempel@gmd.de's message of Tue, 2 Feb 1993 16:17:54 +0100 <9302021517.AA11127@f1neuman.gmd.de> Subject: comments Well, while we are on a roll.... A few thoughts on Rolf's message: 1. Ah, practical details, ugh! Even a simple note that checks and comments were limited for space would seem fine to me. 2. I think we should ask the full committee what MPI is going to do about numbering so everyone is consistent. After we settle this one we can go on to religon :-) 3. I'll think some more about what error code I would want to see. 4. Leaving trees for MPI-2 is ok. Would we dare risk posting to comp.parallel (or other appropriate newsgroup) and ask for help? I am willing to send out such a message to state MPI topology goal and say we would like ideas/help from others. We risk opening the flood gates, but I suspect we will not get much response. What do you think? 5. I'm sorry my note was somewhat confusing. I agree with Rolf that only the third row/column contains (3,3). What I was wondering is whether we are going to require that the third PID in the row/column is on the same processor that had PID for (3,3). I brought up this one since I thought (at least at first) that it was a simplier case of different topologies being alligned. I agree that this is a difficult and potentially confusing topic. My guess is that Marc at least wanted a mapping between the topologies though one of the topologies may a poor layout because the other has a good one. This also raises the issue of whether a group can have two topologies. We had better hash though this one in more details. I'm open to any ideas? 6. So far only two people have said they acually read this. Thus, you may yet get that free beer, but many more and my vacation fund is gone :-) Steve From owner-mpi-ptop@CS.UTK.EDU Fri Feb 5 15:59:44 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA12320; Fri, 5 Feb 93 15:59:44 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA27071; Fri, 5 Feb 93 15:58:58 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Fri, 5 Feb 1993 15:58:55 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from super.super.org by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA27061; Fri, 5 Feb 93 15:58:52 -0500 Received: from b125 (b125.super.org) by super (4.1/SMI-4.1) id AA07077; Fri, 5 Feb 93 15:58:49 EST Received: by b125 (4.1/SMI-4.1) id AA10426; Fri, 5 Feb 93 15:58:48 EST Date: Fri, 5 Feb 93 15:58:48 EST From: lederman@b125.super.org (Steve Huss-Lederman) Message-Id: <9302052058.AA10426@b125> To: mpi-ptop@cs.utk.edu Since I have heard no objections (and one private mail support) about posting to the net, I propose the below posting. I had mixed feelings about whether to have people send their responses to me or to the whole committee. I said to me for now so that the real junk would not bother everyone. I'm open to suggestion about this and anything else in the posting. I will post this message Tuesday morning unless I hear that people think it is a problem. You can also tell me you think it is ok if you want to help my ego :-). Steve ********************************************************************** to: comp.parallel subject: topologies and MPI ---------------------------------------------------------------------- The Message Passing Interface (MPI) working group is trying to create a standard for message passing. In other postings to various newsgroups, people have been invited to become involved and the proceeding of the meetings have been posted. The process topology subcommittee is attempting to define the part of the standard concerning how a user can specify the topology of her/his algorithm. Each vendor would then deal with trying to intelligently map the algorithm topology onto their hardware. So far general tori (of which hypercubes are a subset) and general graphs have been tentatively defined. The question the subcommittee is trying to answer is: are there other topologies that should be treated? One area that has been raised is trees. However, there are a number of issue that have been discussed. Are trees often used? Are binary trees sufficient? Do applications that use trees vary the number of nodes frequently or does the tree stay relatively stable? Other issues? We are seeking input on the utility of trees and what characteristics users would like to see. We would also welcome any suggestions people have for other topologies and general comments. Please send your comments to me (lederman@super.org) and I will make sure they are passed on to the rest of the committee. If you are interested in seeing the previous discussions of this subcommittee, send mail to netlib@ornl.gov and in the mail message type: send mpi-ptop from mpi General information about MPI is avaiable through netlib under: send mpi-comm from mpi We thank everyone in the community for their help and support. Steven Huss-Lederman lederman@super.org From owner-mpi-ptop@CS.UTK.EDU Wed Feb 10 06:46:03 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA12556; Wed, 10 Feb 93 06:46:03 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA28434; Wed, 10 Feb 93 06:45:31 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Wed, 10 Feb 1993 06:45:28 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA28373; Wed, 10 Feb 93 06:45:10 -0500 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA21176 (5.65c/IDA-1.4.4 for ); Wed, 10 Feb 1993 12:43:29 +0100 Received: by f1neuman.gmd.de id AA14624; Wed, 10 Feb 1993 12:43:01 +0100 Date: Wed, 10 Feb 1993 12:43:01 +0100 From: Rolf.Hempel@gmd.de Message-Id: <9302101143.AA14624@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: Trees and MPI document Cc: gmap10@f1neuman.gmd.de I'm sorry that I didn't contribute to our discussions for a while, but I was travelling for some other project. Well, first of all many thanks to Steve for taking the risk to be bombarded by all that mail from comp.parallel. Did you get any feedback yet? It is still unclear to me what we do with the alignment of different topologies. I would not like to see more than one topology assigned to the same group. This would not help in defining the alignment, and it would make things less organized. (I think the 1:1 relation between topologies and groups is a good feature) I suggest we discuss this point at the next Dallas meeting. With this note I send you the latest version of our working document on process topologies in MPI. As before, this is an uuencoded compressed PostScript file. I also sent it as a first version of our contribution to the global MPI document to Steve Otto. Rolf ------------ uuencoded, compressed PostScript below -------------------- begin 664 ptop.ps.Z M'YV0)4)`F=(B")DW8LJTD.$"AH(2)8;(*1.&SALY.D"0L9,&SAP0-5S0R`$# MQ)`WBP*9.1S$H[9>2XV)C&*90P9\K,R1C#!DRG0M[4<=/3S9FT>#*6+%DC MAUD<-&0X+<+V9)LV9=S0F?.0B)4D`_VV"<.621HW5#5R]*CQ:M:M'-&6.?,8 MBIPW8Z:4H9.13AD\+N!\?AA"P0LJ9;`0\?E21HV2/<>\)%/&#`C8LFF#2,C9 M#8@73O;P-M-G^7$A>\0\)J.Q=Q\GQZ?L.3T&S?7C6/9,`>&DCP(AQZE(V4-' M#N,Y;"J6^?XB#?R/9H#.*8K]A9TY:>A15`PQ@'!#3FW4$51_0*Q4W![VL3'' M'EN4U$*!!*++IH'!H`"JA`"R?N"$*"096D'HF^&8ED MCL;5>&.2+_X78%%0YB1E@3Q11^5O4DQ9T4IX>%B''!,)MEA[:>"A41UP[-DG M#2"$]=)G:U$'*'=H@``'BWXV6D.@HX%`*%L@@8"HHBPJL!\==\))'Q#QL05D M&'"4L4=]]X'07AW\]8'>"T`LYD8=0)E11AED[#$''175,4=N+[T@*ZULV(JK M`JL6M>B(KP+Q(1QIB'7J",]&^Q$6KAYG1!,5;9'AAB!<""Z'_1DAA!`5CKMA M%_VY85Q)3KB61!'@]J>;'/4>1X89+?":!ANGN@L"#C#J1AX(`A/WF+8@TD%% M'J6",,/!+QC1,+=XQFF&8'+H66[#Y[ZAI[GH8;?K2FX=)T88^X&`!0A,N9?' M'!AU9/!YJAEOO/'C M:%0+UD=.,8^H(_CLJL1O_ZR`7<1>\"!$L+&1=WG MW;5`1D5A*.`YV8&=X7JE<]0A!J2GOQ`[ M&F6PY!+LJ<_.:Z-3N;4[H+W_+BC.!E!E[Q"`##AXF3SOSN>\^_.Z)YH3^8+"),,-FI[_:NQDX,&^\F$`G_,252#ZC09]+4C# M8L+R/_?Q2540*X-K3O88W4408:9J6_V$UQX0M`U-9*!.>4H$-];EZP5RN,,) MY3"&$YZA3_!RS1A@2+$CG)!Q%!N:'(K&!M[(05=AP(K$8"(I%05E/+=[P')6,U)]I'[CV2#_OO3IZ>25.]Z8VR!^LS@ MF^$5KR5TR-_OSLC&-*ZN=8E*4QK7Z)M-_:0[N"H#1\90E-C=<7=N--Y+DN6M M<3VI0"UP08'HV$??[1&-DH19"+O`/@6"93Z>#,MBYK`&28G%(A.!&Q%.-886 MOLQSR+I@X?#DO`M"9@]=P!+G5,:RHM2OE1[LX-H$1D4ZV.R*D1I/V\9CN!@H MX):W>U_X"&B^W]W.B6R!8A8C93@I+4N7:P-C,/$%0A&:YU5)@%T+RPD"(L!- M#"C9@Z^R`BR5H:0%:*A:*=?@AC?0Q M"M!0&]Z`%8L)$^98@G0DVYJXL490[YO`,ZSG__4%:_D0`>SFJ=$%]42 MCI1TN!?(%4S&>5D,8,!7&,GH!89)P@+KVC3;F*5K,\-.20IDFQKL85Y2)"(% M4W:[DDSLB7I:VQCL8)QO4@U?-JC!#7`0`Q?8@HQ>QP MJN5.968XLRX47S;35UEA#W"SZ8]X10=?U5-AQGF!:MX0-8/=@7@3R1Q*U/:9 M]N[F@?#;W0T@M8>2[%=0CS&-'.P`%#^A8#9S@$-\\I"",L3!4GCH#K@`!6#! M9(7`;%``"IP0&RHT^,$7H:D709`?G6(II_NA:8E3C%.=6HHM>]C#&>:P4#:% M*Y.1+)!ZB%C`($E0NWK2[LQ67)3=5F0/0ACNB$)9AE&6\@P3*6HJ:2IC&@MQ MQSF^L8[;!&0%"/E'/N8QB?539(PA6OG6;X`;P/J$;!#<`6YMR%RX M)G;O5P&4WCGI]W'>H&]Q"?P%<=!W@0[^PSX%_)S'F5"?^`TWF/[)WG#S7TG& M3`W"'+R&38$I4]A>.<;)G109M:X/$*(ZLK%P_%1IQ(]3]L+6SP# M&M'$$3BS,1AZ)9:#&-0`!]NC2PUF@`,8W,0L>^WK#*3>5R!\E:W:"N#$:J"7 M??$`!E/GZPWX:@0C\!4&8Q?[V=..]K6[7>UP;WO]F[WL9H^,'+P`=)0'SEP4YZ&(!^ M[#$8`O`E[WOB@S[SW.-KYP%_]M$?W_-\97STI5\$ZA-_"(Y?/>]A8`3I&^'O M1A`\X9.@`'I5['>@M]_$O#[VQ!MA".^_@>/GW_C3GU[U,)!!_7M?A!L,`2^R MYP-U<7AB)WG!IW_`!P.,]WU%$%'?)WE%,`.21P.C)79-AW87&'O<=W8:V(%O M%P-I)P-A-W@X@'GOYWN^1X&!-P,Q0`,^%WUDYX4;J(`*L%=F=X`;:(9C)X$("`,U@'V\AP-NN($5:'0(&'C?AP.X M-X/^5WQKN%>,1R`&.(4QP'01=3]-)WQ"8'9[&(:,Z(%S]WDA.(*KEX4&F(*` MEP,L6`-Y^'D1N';.I'D+^'F2EW:8IX56"(9<^(4_F(HQ\(<&Z(J^!XNM2(2S M.`,*T'2U>(&,YX2,)P07:`,@>':MR'8[2(P_*(29!X:N17QFIWG&-79YQWK$ MYWLU$`-$T'286!8^L%<3(X-@.(NQ>';3Z(/D6([F^(.#R%<*(()-1P.8YX[M MB'DX4'=Q%X329X1$R(>3=WD`:'F3AXLD`88P(`0Q8`2`IP`Y4`1[M8UL2(#B M6(1#"(ZB>(X469$6V3CA>'8V,'J-IWFGIWG<>($AR5=KB(/&>)(^.(3'"(9Y MMY+X.)'".(XO287-5X,TB7PFZ8-L-W5T1WEGIP`/J)$>.0,1.0/;0UH).8B; M*)'<%Y$R^91%.(X>284M.953B9(Y"8YC!XQ>.(_11X;1IW](%,IWE,AX;[V';^F(8`"'PXT'1_N)>]%X8M6(*^8B,>8SVB(SYB(^2F83)1Y.)V9AT%W5%H'\*R)F;&9;*YW@W&02" M:79%D(>S6($B"'I"DW\S$`2]YXZCV)%92)N/EW;2AYMPEYM@"(L]B(^^67E( M%X6W.'BS:9R/%X4?J7BJAXM_>`,M>(DL*`3;N%<.N8C%-WI3]WRR^(>M:82U M^(H8&(YSN(O.&893&'OIF7F2%WL5Z)YP!Y\_N)V5^7LF.8V#&)TKJ(VY!W9^ M)WTT8`,R<'P4B)SS2%J2YY6SJ*`*^9O4EY'EEY%_"(N"*'D5ZI]$*(59J'BU M2872)X'X>"":!YU!>)(M*9`HNI(N"94`.HJ9#2)X(#2I(2 MB'D*()L5"(\_&I=\Y94Y6:1VEZ(JF:11N:0_.9,S.97/]WD$HH+9J(E3>IT& MB'2\MU?&-77/V'O&5:(7N'?))XV^UWUKN796&'LVP*8`*II:NITGJ'I4FHF; M:'K?:'QZVCCX>)FPQZ&8:7>/*9#WN*>1Z7>3IW],IZC,.'O"%WX$ M"0.&Z7?7)Z#ZQXT4B(),9W:,"JJ39UR.AY4_::I(FI*JJG_`.('<9X2.%Y1\ M*0,)VG@!&G9Q:7R#5Z=C&`3;J'\$R*:^=WQC-Z!\F7]&R'C(J:R8QWAO&:+Y M.'QJ!YPTV8#5>JV@>'>SZ7EL%XK*)Y%$J:1`R*2N-:;<,Z)W9URE-Z#XJ8$I M:(W2R9]?-XN(MWEY]U:G2X)Q3:TD+`2A*3,X,I.J[E6)#F2+(C"Y3FR'HV MV7PW2(YM.HX:Z:)`B*;D*'PJV7?,YX-C6(+N^GD\2XY3I[(^**P1[-K^;1%")?N^GSVVGS"V++;=Z8U6'H%.8VEQWU#T))#2+-E MYWUGRWUFF[:K:9"!MYEX&%H.B78G6*^PMXJ51P25IX%`N8M!2)1B*XP*J;6M M>+5":+C1MW<^:`,YJY&-JWQ0RX'2=Z*12W8W6[E-28Z8^[ABIY*,V[,;";JL M=Z*1IZ\XV)[XF85_JW@:R'9H^I*L5W8VN7?&]WT&200)*0,VX`.,&X,\\+21 M"[PRF[.*NW;WJ+6-ZZ*-N[Q_V[I-:[B-:[IH%X32&[TM:[U,F[79NY;;>Z;9 MFW=?V[G:*[[<2[[>:[[A.[U9.Z;4>[WM.Z81FGG.R[+&6[_FJ[S'6X/$VX'` M.XW`.X3K.*!NF[MXF'O"NX#_2[[$>X^Y.:B-NWPE:KY3%U&-&W;,UZ7BB\'J M&[;,-X2`QY2NY\%I:[P@S'B,UY(7;)*,UWFCN,+<.H0M+(RJ*\/JZ[,U7(5B MJKY..\,VJ+;`Z\-2VXH%>8F;>7O]*79UBX.Q:YHXB*8SH+?TJXY06\%!Z*+@ M*+T9B;CSV;.Y";KFJWSDVY)*>[.6:XZ86\;E>,:9:X-I2[-+[+1N\0^:9(GV'EM>P.XNYDYX`/0>7A2&\EW M*[5X5Z*6;+\WW,?FJ\F9?,F;[,D4[,F=C,FY\'+3+YX MG+^8K+SQ&[BSG,FQ)[5+G,L5W@E]=9RM>^1UI7W7N` M*Z<$.]AG2K`D6Y9VG=B'#<9S;8!V_=B-_9>*3+%RC7;7.-F==[67S=F1[=D$ MFY$F>[>U^*=`V(&W_NVGV/"W@'7@21&\*&O.`CK(X,#M4B[.`4+L()+M4& M7N`16.!#(+01];7_?8/]+11 M;M(G3LG@.[A$W,OVG>2PK8$E/N,%G81EV+)A+N2("Y23.MX8WG\(7>$-/N%R M'N=TCI$1CLH6CM`7KN$#WN']7-=U[=2#2]37.`-7BS^)%]Y"CL4G?M_QC(P< M*8<^^-2HS,\JN7?@++\PB\4P(7<86?,)TNWVN;EQ!_.NY;-OYY\M;WN)?9Z9$779SN;=O279$ZA"C MY98:^*DR#.9.;9/#J[PC3,@HOK;=C.0M"I3ZB]_++I5#,(T2*/`!WWQ#L+IB M>KV9.XO]I^V`;03@"[&*G)1$D(=(#GN3VK][R'S^9X6N",%]!_*1M\Q%\+VD MM[TV@/(H+^54WM(P_?(T[>0MS_(R7_-5#O,J">7RB_$[K^T%.8R!1P1<[N)\ MY^BY;/0HS<I+7_54__1:W_5]9B[5-/+'KS'Q36(2D/K0RF\@:7<"_6^^R"WNPGN-*''HH M/7I2O9/OS;(U^+Y7O^M.#]%1ROB5J=`K'>G)3?>U3>LJ:KJ;_GXSJ>S.M+)- MV?CV7(Z`%^GRN\ZE7XZL1\:H7H1J;L:>C^[F:(Q&2I+5:)#A![>CWO=.K??\ M7NTU2=MW=\_EC.KM:JAHW/C%*_DW6,MTWW3K'+XP^_G1WWV@3[>DS\8C>\;. M=+0<&/W*;L;>#]&:&_[<__TZ:_[C#_[J[X/W7;GM'X2V,<2WOST^@`,Y$(,: M6S'4H;OV\RA>%W>Q!QS%/2<4`'>:91I3=H<>':.'-J@<$V0Z1@I0`5ZFR%3% M:!![2E-2[.T,@5YV`V["'2H\)4$&E9[6!9>,3P6T0#YM,CW`!DCM-D\"?(&- MB05")A5(DP@@)[J`&,AIF;<-2(&&P&`"/4N)1<&D(8BICL\#>D]+**IIL_M& M?"A8JP-SVL[73;0D2-2"3P4J@B))".:C(4@@HA/N@DYX2`'DGN!#E!"<$6`] M)&OLV"C_DW@ZSR$:+3B@&N$A&@`HO-&`$ZT'%O5W$]X`8)4Q;^1*M(H%(65$!3ZGJDSMN)5W"H0,DC1[;PQ%05=RX=*4:A*MKF> MJ.6>%!&$ZU"\Z0?IIBVDBER+>LIVYL+&D:P6Y`6S$1[*`?C#&\E"7\<+%8]Y MHD(63">IJD$H"!^@0,H[6"D*X2(+57D@FY?;A>7+]PPH"H2[6-`SS`$."8L) M/]-5E6R25;*'^##S1*`,9"YP$:D#43?(!'6H*X23/E3H,8CSZ5AYJ(*(DTP1 M"A2(_LETM;^)E@RMH;TR.E:P`6U"(-B?`ASMBEH?REN5H@Y5BA@B$=II3RL4 M:KNGU9HNG3;,4./P[&&H]D<+H:#Q68<3SQTNI<&EY(C::4MT-FH[0:"&2!2K M4%$LB2>1$UVQ*'C?X%)"JDDGRDF-L/:7VM+6'/*%:@?AU22MV)8RD/JZB!`O M"LZ]](41;9\S9$C=*-S-KG]WG_Y>*?Q)%.DCPC@^)K.#1O5`.9D&/T:6_QUS%!M66H:*!EU%/@ MJS)J1IAT&3GC9CQ47ELT+';3#?@$OD=O!>\`MDV](8J2EP=KF[6 M\+A00*.-XZD'`4*)%(&F4?\Y?`7.Q+FH\.._'"`@O#O6*ENE)P*W^R95EDI.:#3AG2Z7#(>14*Q4&-K/`&N&]3^6E3[$D8LZ&)1 MQPCT(96@6/Q/L*]$0B-A"(::E4DTDPLQ39J@^!2?%*+FC,DAI^J41'`1CI1#/(`,L5\RLY@4D3),"#V,]:U?=2:(<2!D/$F M29>'$!S.KG4Y M^W@0%UL[\?(&"`&Z>"VK5D3RDWMG0W)$'K#$$"5IE%I%UGE^'+(+:K^N6.Q'8;Z/4D3:&V-)6FTVR:4/-I2LVH M"34[3UZ@0#;@!^(A[4$#\)_Y,0)%`5`,J(\EO%1B>H1-]^/_"`$1Y`,`YE^S M86\S;AZ?U8@&G=H,N']+2&WAP+C'=^YF$@0]FJ\$R4V11#AM&)V:>(.'1MDB M.\B,RESBB5T9KS;*26%&.859"`MAD+,IZLV9J`YK63L$0=OH43!&J3>TIA_0 M,IVET\F1H\VG[E864T*!:4DG>3XRYH1BI^Q9@. MXYI:HP7^_+T.*#ZQECOD8.IH*LD=]4F*5A9"%%FR+QC"STD'O,Y2E.1!U(I) MJ:3\V1=55+9RBYL/"L&L@!83`6@[(X3KLVA%R@3Z$M7G9JM!>4=^0J;XN8-4 M)>E\?@=3?RHQT&-5[*-90VE!LKW:2HHY907E0R%)1(>LG1JX328OPDP$RHEL0/"'1&WJ3 MA"1.XE%%,8K*'6)DA086;N),_4,Z/]!7]:?!Q7K0(!%`0?I'_]`?TN+IB`]?2HUA[33=S;DU ML+`HP!))9T?_Y)W28WP*Y];B/`^P,BXFRP0#RUGI>4QY9QIITL2W2)W5Z MGW0:`:SLAD4'UL"**(A4!O2?''#4E-(1HZ$*"P(I47TK#,.I%S:@U:Z-. M+?'82XNI2:6E8W2DLE/V1YH,HQ%HI>".`SG+Q8.,BN2VDD)/E"2>RO'DGA2B M+;1`6LD0-J,FN%2_$-\<3ZGI/)4GJ/I3V],%5$]#]3WIG:TJAJ95$J))]JG[ M":/:=QAG3TXE28GTMWTHXO,N69"ATCP$U*>AJT!%J%:@9[RK!U"NFBNAEN08 MD,C,`6:UR]VV(-2AHD[8J5?MB%$U'L4Z>*A07I`[`:KMB![GYQV?U'C;G6_O M=A["AX@"KU#H\:QA*/9XQ?H#CK(/\>%':!7Q?1ZRZ@,#:S42E9,-[40IR=-( M\12:'(:IZ`Q-J'`TBWKK;NVM1Q3L?"-9^B-,2$`&`%G:(EAM92;8I24U+Q?)T928D%3QD0 M/&E`>PV>HC&7]BRPA(/TPCSS=#TK5A6L^SI4/T\C7%CNBC-I+7O'5T13[>2+ MM7.O+JS<:$4O$!9U0@_V%UE`B:0\LY#LLD:E*8#Q+M$)LC)7A.6@,(D&'$M# M^#"O3U]\2O&F M+)<"#6#'>*Y5H:8J%.+RRE4W;%3&"2#4*B22)! M2U$**2=M=JVTO]#^[-6-.5,Q4.-K1O2S9X':)D9,Z55:PJ(?CO%@T?1J6*=1 MO7*UF/91&M@:I&"?CQ-JL&*G)05`BE@)SQ2B36J+UM`*0)%TCXCM-DVO)5`M M)=M8BF>7+3N]M1:0TTI;4@M+1^V^$[6E5M6>VFV;"%M^VDQT^Z*$T8T-1I#()MV/1-JSQFZ!+;MLQ[WV*_P1L/(KVXF7 M!;9(OLY.A:Q4$MT=4"J)2,DF'W3N]NZ).G54?57M'HW_ZNGJ[SW-?OLUH'J6Q$6+9J7GN?XO-@L)1-;%!^% MG(GTJ*TX8(9R>^&5_*\F*Z+,4A!T@6:O$YJEABB6ZB@_6E_CGMCUO;*TMR(L M9_*'AF\9\KS'%P>=RE5K?T8+\]U0SQ<#K=I6.WNC41SRK_YI]G(BW,N)Q&[W MK:\0+Q[$0K=3(K>*+7!=QR;Y8MBGXR%P)'/YI['5BX`I,A]4GP76PF!*+:AX]_-L@:6V[I';'-^K;$[I0 M-60]BP&D"03AKN\C1Q'/#XNJOPF';BH4F$>*XJV;AB6PN`VI%#@5*UP>RXIUKBL>Q>FU)3%@H*62&EU=&RU3V.))1A#9 M?Z$M['0FAW!@:2`L.HQK)C%.J^DUTD):FZF,R6_<0Z60TME*XU\FS!7,T4QLB8$0&L;ZR66I;^VE<3%N&N6M6$<&?ICT*X&Q.6NA:6ZG^8KQHJ MG;/4/ZTL"321[^C0LLA*S`&;8M)EAPTJ[D1`Z[@=WZD0.GGR)KT4M`(4&UJH MT(,4":*][*RJJ$3*Y)@LEHJ/?@T^I2K[MITV='E.S_]I'%C18AW&&?6"K.S7 MN;,*CZXZP+IZJ$11&;JCY!0V]48Z2=1,4-HQ2&3G*E_E"<658Y$K*H?"Z4N% MJ;4&V?PK`HHY/U`3';'HA:B*J:"=JC`XS];2&TN7>6;.'<`X][Z*8[E3KY#L M7_-(V(E7#F6,=<1TYSW;LQT35>%)/`D87Q5E,LE-2!=-5IB(H82D0WS)F#FW M:F94!)K*D!ORS#DYH+FAL<.3W]!/EG34"`'YP.]S5H]K'_I-X$D089[:8SLY#)M7ZWR M+FK!N#I,A6FD5,S8^2621YT%'J7S2CUK^DL(B-`2BMYDDH!JH8$T#UG?2VIV M^2-"#(8SV6#R5O$$GBC42J[,3Y(J32$I5%6;)&#R.VQG`\Y9YZP7OZJ?A:M8 MD?\8/F$T>`SC?E++\\HUDM&_J`4I93U$EL5()]G9C22$N.47HE8@.KBFY!%M MV>XS4CS10E4!&5[0E':<\/`I26Y(^A2\Y?-Y'NO$8\V%Z8@Y6M$6_H[G-C1T M.JK`OJ,1E';2'3@;5)VQ+C$=R%.;F%/F"3LOB2`9)!]HP@1J,)U&HD:%V)!8.R*&24I_'T@]$Q$3)-4+AV[""BH[&J> MPJM]*E5/4E8-X3239[K5H(GT0,L:1)I695@;`J@I'@]C#MWAGD^!TM5XB@(Y MO^0#H$[1#Z*[P["0^284^9(>5+7V0BI0'$XD)`161?2?)9+VQS(362K$E;`A M7QK7F&>-.FL;L-D:!ZYRL%-J*-N&8'W7=!QR2)ZD MJL<.1^K)JF*HD,TTB2I7E4^W\/F$U0*$)8-S-LJI[IDD"2@")7]@E0"VV>&H M4RVHSG%!4K]*!>T\8HT@EFX:R40B5C/$%3 M)Q/O%0ME>L:.Y.F@X1?&(3'6E83L$6Z*/>538@*@O9*)QQ"5M$2?,YS.JS^; M5QLW/GQ^4RW>U/])\ZT`6V`@>UPN*AHLFV\ MES#!#D]J.B+H-\.C*>2.B!#=;4U$ZMY*;S`4$Z,.>-K6$\G(2A[MO7]@U40Z M?!,,W&H/ESV1E)U]QZO](X:AW#:/_=;-.5O MV*VPP#4HWE8Q,0'Q;^M]MVNI$0)5+0CRJ@HQ-U\,(06#[,_;V3V1B'A2?)O%%AV!R;Z#9Q69 M=\:4*$OS).A`:'I$Z5#BKH:1;>\NHE?TLH__@4.@4WLL3NKI(V8`U!F;VM-C M&C*HABNW7<'4VRVJU(G'^$;'@Z,9R/"W(_'L@) M>1\'Y+4LE5V\VGBV0-#]$)E!0%,5'GCXM@G:4YN(9`[[_69KZ?L2>2BTL)]< MAG+RU#?*T5224W3%2]'UJ!>-PW3811S>4<+BT4C(KJ( M$V\&I-R7/<#<;Z2S4@YF1+EW]R;*TI*R+K(^?DZAWM/H%SV]N MR%ZY#8)9I%P`HJEPOL_'^3LGY=H.H!MS!);:0N$H.D&)W,+RL"0'%OE7U*YH MI>XL/7-%)LW=<=-615$Q>&HQWW?'.WK4HGL]*S6*6G?>\,JV5&,Z@X^CRR%W M10G[HC\K1:V+MO5"R)3(85+RHX5S)[?Z/F&V^^HG3_?H>]QI!<]>)\C3N2"_ M>//Q>H$T#)L:BT!%WT1:=_0HRB]7Q7JML?QR(>W!^3!(B!UW+A#]ZF0LK)>C MAL?5Y:^("[,5O[*K%52OF//)H+LF)LZ1\PG&Q1Z\Q,T;&=&IE1W'GZ+Z= MHZ:6LN0O2-_2KF<FJ(SG:<)4=M8^LNL9P^2-1CX@R#FB-NU4O07^ MP-=UJ>3$&/R`%^74JK]'GS\6G""\B':RJGI33;!K:5N[6H3L>_%OK MO88&>7T^A(+\%BOR]%V^W^_\?J*=&,5B\O;0B47Y:RC@N6[+C/(IEXUY]M3G M@*;\S27E5DG*J]?[7>7#/)E/\LVG;Z&^,Q;?JSS08O-PGHRU>;YJU:D9#:[S MO>R[1W.D0YVX>/?Y7PE,LW_TJ"6+J;NAAX*(/@X/>LR^Z`4]HW_T]ZW0)_I& M?^@5/:2O])1^TJ,X26_I';VGQ_2`/I7]+T>N+<,[(G9DB_/&O_GS2-27F-F4 MY5&0LS^J.EZ^$EAH]O=_/7NZM/->QO$;#??7_OV"O[ M84\P]9<@1^G(Z]GS+`^*P/Z>6]>-S%X6'O!LSRE#(9/M]M)>;46C<><6NSFY M_WW?_"HS>VZG[MVBF4);;O$V#L5P_^[W.!2B?HT3@.)[>\\/I7BQ[_?\$%G" MX6-?Z?A]P0_W&YZ/3^:$3_`57\-'=0T?W1&>525]FO.EBN^DSY@D5?5&HXO_W@A=FV.V&&Y(>/K-Y]N MM;,31,]9UD%GZS^_4MYS3C[FB7Z5]^QUO:W?=;:^U^E>A$+Z`IWI&WT"K_3- M^5N_[1_Q(\8Q8[[SM`/M6I, MJ^>?M7G__$]XGP$NJ`-"LJ*VXGGM\MSP/!YJ=?,IYL]5W4#W^0D?S$,MQ=^\ M]E;C3_Q.;/%'?G#&^!$_=/^(E[_P]_4?C^"]7,M<_&K^S M?GF[]6\$_ M[`E_R;R.`D_R)__'GWE"/^1G_L_?S4=_^F[B7=3E[WWTO>>C^$%VY!(\*2SN MM1W"]SH#W\4TXY_]D_ZW?T\K_],_HT M%/H%?1HB_K6)CZ<\B3283BO"S%%%Q%WLPGM\?.>*"X)TR%.5TT&7XTE!91.[ MQ_5MTKKPI;,(B*@HX7\*4E\$@H8+H59)Z`7T@*J M*+<=`-.`/V:#E@#5B.>'(X(`VH\P6!/*`0V`,.@3:@ M#^CD&(%,8!$(/)TRUU:=U\S!/VO3/.+4W0!31_U3!_%^5$M?A679.PD@J?/' MS%>PTF\G]]0S!=>6M[+4=NR3BV*H\3/+G:63M&/X6;F/(&3 M%/TT8L&!/102*/N(0)TU MW]U-)=[5/]W-E47[&7J`#9G3GL`Z!QTLE_P8ND<1I38EB%5S@K!XD$U!)\0( M`9=((E:4?31GC-QG'5%V[&`:4]G%?=/49>4.MH.:RSMH#]*#DXY"XOQ9+J[' M,>BW1'3,(!X'\54Q!&%Y@\9P=!#/=6001_S9#X'\^ M3/-BZ_1SH<>/HR@I=`&A-H@3IG+P'MOG$P)U@IQ0I[D,=43A$K@$&H,3B5L' MGB`FS`BUPUR5.0B83-(J'37D8."A>KE;LAZMY_EDAOSU>XH+UY M8:%7.(?],65A6KCF"2S=QU<(>U!(\@L;$]*@+2()Q*/642UXH>43$#*$]\Q+ MEQ&V.W_A6"C;'&:#86"H%Q:&`=]A"!@FAGNA7XC3T3K58%[8&)HY;(E9)\BT M?%I>5G<&&7;O!V>7$JZ$#8B4I;O<*?$=342_)"][2_E2XF6$W.#WX\_P?*W+ MTS>_Z$+5GOARP$R!15UN2-01=<7+UH<4?FIFSDD(D[@Q9`Y@HQF6@<((05(5 MLE**E$#5\W6"XXWA\K70$\YQT:A:I@>/CE^"\\ MCZ;$*IT_[IQ*!_RLA^IA>Y@>OH??(4,6'ZYQXJ%[5Q^RA_"A4;CUR5](3K\C M?\4$*B%+6,/!.(#-SD/6B4@&3J`CMG2&RT^">.CL0\5+A#C3^3X2XE13(5*( MQ]"M(PVN)1&7_Z(@WCMHCV&W%9IB0XL#]@6:.GI@>7C*)$_W#FDD$09F[D?: M$^^T/?C-C!@).C7I4@7#'2([D$T72.$(+MM#.3C-Y3UD3-I2/*F#!V$^B`_* M@_7@DFC9.8E)(I.H),Z#\Z`[1_U$;0[(;_2X=!]TW5RT%Y9R69U>N/F%B5\B M'B023H-HC/!C)AJ$:B*:N,:TB6*B&X/]O8DFCINX)J:)=6*8%M(:GS&5)T+>$F4B0=?I5?Y9>^7()>761TUR$PS\X:8[9$BI,B&E,I2HK& M#J6H*5J*FR+?XRFV,3(`.1C-'34NB(L#*>Y)9DO\03"4=[`A8[B+17Z58?SQLR@B M58Q[Q'KD3:_ANU20^"($%__B#PV#<)PL)`SN6GLBMW@8ZCC?(I_2WTR$Z9`P M""[NA?1?<7C\C(OK(IES[*2+X>*N8RZ6BQGAN3@ORHOUXN?1^J"+]F*^B"\J M.).C&6BG1B_P(DE(\F(,JZ,>6+*N!?2BB[CMA@H M`HH;#LUH[PPH,0##2)JXA$D>5;2&W3!S"I7$&O*+M@['(P)R=-KAT:@+%4Q. M$+*WMRV-3J-M`S5.C2(@P>3>^3.*R[PXNI"+TDMEV"/2,;M2>4C?62,YHY#H MCC5RP4N&,R?%A"3A32@0#D5=B3`+/`@E3G938I/HU4F.5&*H M=;FH@U(@LO+(686CH8O3X:`LOT[I:`2@2X72-3(T(7M#D][4.N)+'%MK)"7* MCKK6GT8YRHX+X>UH.Q)_@I$VPQ+Y+&6C51A\5"=;H`P5=E`TJ:$-5"_%=,PC M`I2`H2E"7?<7T%5"'Y@4Q`]9?PF8Q+0JS5D]6/FD+!8DL,D)W(ITH M,L:,+2/+N$%ZD+6CRMA!CI`?)`G))G:0.`Y,A1BMWI#"'$^)?O4\XEQ?ZD>6+ M@K.MG"^"9,5"_122&DY&M)+@>P?)(IG"H4$B3%QB'R$QHT4=Q>"5.=3B<'>F M9)([#XICDQAN/6'6$DJ*52]/^'BY="^'T,P%E_!%^K5WR%Q>)$P8R`&%M`WI\AQ&C)]@`X:Y-%?S',^R M!BH^O%5"R*0,,FH=;'/O*'/7Y&QC&IZ1_9(:B7N4']I"HM"!:0\$@_:DKZ@U M#DLGHM_=643-@;)]C%K.&E>U52%3.4@\.445*:X+<6=/[I,Z"#^I[N"3`*5M M%U#*D_>D0&E0$I1L5SW93RZ4F\?,>"[I']B(*PE#,4:("^.QIU`FLXBAEE0E M58E9TY8+;7RI$GRVFZ@Z(>5(F469E,-02EE2GI0LI4I)NIV4-I)"-A=F1AHT@!AY/<2L74`99KJ2)$9!190=F3!&5265!^.5A.04F/ MQ)/.VE/)F$"5\F14:5`FE$[E5:E5=?ZJ,`YGDT MI?$'X(D@N..P":.QSIE7]65]@.&'_0'WO4_8%'[1]ABO^!D)@=%<\@Q1@94L=6(?7;=%(J&UQR40D^ MOTUVPP!%;C(04[:4'2/93>FQDVR6,@SX4H-\4NN*,*)\D5^UU-K4?T1S!(([ M5H,`+*74:VGV=#>_#<8&-V661!`G!5H.EZ)E"Q1:@E))EA@&7&XMHT=V,XV8 M8>/5(54IR98476TI4"E4V&5$I5V"5-RE,D6Z9)?=97DE7EY3E8\PL^*54,73 M0D5>JIN.BX2ZD"?3DCK25[@?\(7P@EO5'.K57 M2$^B26/Y?S@>%8_H^,Y]4RY>X@&;['Z%98R8=MP&HL6!&8`49D+71<5;_F'" M)?2Q7%HMJ"6'6=N$4K^E2'+L^)9;2VG98>(I'Z9FZ5(UEYUEXL-*XB?YAV%$ M6X)._4E%J0(E<5^8LX:B,5!&&ZZSGHQL5U61260>F4-FDBFR*9GHR9(YFTQF M1EN4B4,A1<='YV$D?27)GS`BP%PB@=6OY-,TB)6;K*97C9F-6YA)9IJ9PI68 M66:.*1??J])_M%8X9K@SCTDF8`AGDG;,6L]6Q,6R!43Q6'J2G]F9;!&.=6,] M5X20FW:BK#,R23:%4HDFIU7":B"'(&YFR904^<>4U)4Y45Z,5Q)5)#"`TPA(1M_YIK"?\<(_K',D98'2/3 M4IW)DN":^!OTX=[$-P3:=^0=84*#1TGW78$[*19(%+4`=W4A]<+^G)M"EFN9 M;DHIN([OT:K8,/<(W!%UW%C6F8SU;:501B:3^:%0++O:O:*-69D""?`V1GX> MO.:WN3ZF+@_6@D7F14EQ&3O9*'%:S-FD@]:87",260+[7&6`2D!T*F5I36;' M2;*1;*1(?V:T"4E4)I`U[Y$>*!;/-;X-,%TF#Q"1J4*2IJ09:<52G`D>QDY% M9"L+([1@:6/YF%\&6\94=]'95IU95'49PB5H0ISF%D>RM4\5A7QLF MKJ*-59DVC)5)6;U-1=*4"20U+'_)UT(FBHFT2P*7Q]@QQ^/56;\1(1?E0[)) MK9W;E-5IP&%?]I7GD642G$A-TZEC_EG'4H])5MI*/Z6[-E0QF1]GOKFM;#PL M6RE">HU,7PG^-94E2PO?=59,:2$PG*ID=(UA,AA^4G"VG!$F][!`=FF9Y^99 MC4AP%^:@M$_%.XRB1<+\O!ZD)^JA:>$?A)NG(YI4(UD<)A)Z)D_G':5S/*67 M7=KQ9%[2GFCGQ..:\"[.Q.@$2RU0CUWQJ4*Q;#I)]`-WT5LWR_VS/]U7])5K M67;=7(L7^'5C71[C"/8Y4'U;ELVQ&=U9=?@8]L6(A)_C%81U"!U=(=<4HF;5 M6;.A,7Y!W-U[9,HA-5?B6"\9][$SC%P&D8%TO&B@[=(.*OO183+H)H=D?5NF%>?E?:XD")#XN8?U1>:7V^%$XB>)V(U& M97%Q]N>CA'\"3:GD7L5Q,EAGEB-T:Q55B%#M!(9>G"$-#D6&YFWX0QEZ<'V: M:*@:*H::H4Y72'/<_6)=:#A":_I?LUH2FAM5,0?H]T*F_`R:@8VS%4J M:8TQ6`K6LT5KZ4Z;EB+::54@N]1XDI#-+IY=N8*07:+1%B:J.]F9NQ0GBHS5 M3GBF7P:*ZDX7*)YR:V6@=TU1DS,:;D1,03+J/#49H2$*:!VBIR@C:HO*2L=6 M@H4"35N9:"^*5%&B!(V'YHO2FKSH+FIT'J/GEBJ6<"VC#-G"%3EJ=5'7-=.G M@2QHRH-7VV,#XJ`^ M:`_JCH9V`Q:51YY@78]2>"*+Z6"=ER,$9Y%@-,FP^:R9EMP2^Z5E05F+V#JR M1@Z,?14JRH+E1@WH(=20:J`0Z1WZD#*D#FE%JN1,I`MI1IJ!\)^,B$;:"'FD M'>EH19%"I"/I>,*)C5\#8$]'6K`V6TXL^0U>H`#8`RJ38J"J3@!FD\ZD,2E- MJI/FI#QI!A(*M3Y:SNVSD/0G8J,*=6\"/.$6N&5O(J5M&E-Z2SJE2VE2^I1* MI5&IIR.:H3M\4^"%E4Y7FD>I4GL%/XOD6=*.32])";A32Z(N[EH=:FM!E^8G M[`1O`63XED!6D&FBLI)")CC56'4908.)+#M%AP=SU'@P(INGB7`Y(;1F$W>` M0E4YC`'ZD=ZD.RDSDKT@C&'1?P*"GEP\)"EI;A6=1R MJ6:*AYF?I.FQ4WWV'K=7;N257IIF:=#:$#"CV%>WUBJ MU0AI7N_IF]5QJ2(FR]>8MK!VB4?\0]\T8G9='_D8;7(I:+S)5]0``JI.*7XJ M:PL6@CJP[%T+JB"27;4B'\ZEE7S5'Q.J5[J@>EJ7%HAR:0DQ&RH)M7]\'^07 MB%I@@4,!R?[QQ#AC&%FDQ8]--OL8;)EVM6K%%%H3=JAWV-=(('4?R9@Z#G+4F6RJ6NHM:IK0L'XZ2 M^JEZJH=-IEJ<7:IC9*6ZGLA9PB(O(CT)'QR(_E$^H1T/)OV54TZHFIU"MG-B M())/NM-)GBAM7BTGT"E45H(FI MI$J$"6`6T!,4C)XS"I%@:@%147#HLBJN?E;7:KGZK;HJ5\AW>F=]IW.*LZ+Q M#9^`2RLBF&PFSH2AZ*S@-6C4O`EYD&M5)[SIKRHLS^:SZ3[2(%L?K>>&D8?- MR\O7F1ZE2\J?Y;#R*;4*VFG#'$NW4!C"PEB40\I_TXXI8C6E+O;B:%6SB'%Q MDI(JFMB'BE;18PT[HFOU6'-@7F9YW:SW%*G+\9@T:)L*>"#:RU83FO5 M&K4VK50 M/12Q)JX0"2S%H8ZM#A4<@_\F;LH6G%K/YBK M#HV;G/XIQ1VBK8OQ<0@M(@^6!G4X[ML/6' M+]*&P)5PI?]!E.QP@.5?Q9IQL&@*;ZFN@"E,[!+KQ#:Q4.P3*\5&L53L%.O$ M8C4M+-KQPMY4,2QJLOC96VMIGZ)M,B8&FES-#%0J-`0Z*2[2C"Q6E\K-A&'A$E9(O`>8;08P";2-E2&E6(K#"4 MR':ADXM+>2+E5G<4/A+PC!XER#W"==`M'6N.22*5+O])N8-YP*IJ!VY6,^5F M1&3P@XM`(=@IQGH162+[R1J9+&V-:UJ]*I2;:_\&EU8B:H?7@OT)9I\2-K.692I_C0WF"WTI2AFM!KG!'6(E MK::;S;/R;#UKVQ4].JP!*X:X6R'1Z-8;"45D4)JD59EL!.U`.]"B2=12`R*7 MJ#7,ET-F4AUVU=`L":51<1]<,B+PE+`6+=:"W#PDVEKBR`R918G3)J+'9IOK MIQ\[QOZQ]DCS%GT(HY[9AW8P&2%J5HB&K;VT-.U,:]/*9SAMA2?3RF?J'>U2 M%GBNSRR]5GJ9-=B2*&@\)33V"-E"7(Z6QV53ZZI%0:H7-F)11D]'S/&"`:$U MKJ#QM*;43/9L*-O.5E)=+3T+UJ*F&PC2=+=U/K;.!J2G/473Z"*9;>ZQ)BV+ M%-?6A@'>HZ-M>2$+W$SEH=))0`@QHN"$-(2(-.L;\48.D'7%BV`K!].4]'Y< M6,UL_+DRX2UJB5-;7"ZUDBU3"UHJM9-M9KN4C3!K4UH9>8BTRQ`@`XS,!%%* M`X*@-()$I4`B5`XJ1!3`%K!!)JVMBK1^QK8KTFL+(\&V3!EMJ]M.>+CM*I+A M-#RJU@SW*>V0=-)!H](J("QM/M+2YK1$2$Q;TS*W.ZU.^]Q*MS2'G;Z?`M2M,7XH$8MI!.3)22=;1JT/I2 MHT4GY6Q.Q259+C>D5S0BW;:KR$I9@3"R:6R`.YX`N%G4`/;(RJ?`28=SN#1` M)VV#VU5926,5O%+.0K,RT2\2$\1$75@*MD(`DH.EQI%&V/*YFB[9X M'_1M-N*.%20DP00S=O1DE4B2BX+X68CU"UT>^=&MWANGPN<;*$,[-2!,=5'?H=V.^%F,\YDYR7)4F^, M[@^"`$4=&F5-HH>QK8B7*,)39F$>;(<&A'!,$2?XI8[P2^4L:Q1_TKD:;MNA MG/0>VAJN>7;N;T^(JT(OU2J&FHCU;#4>]^H_4J[]([0;>S*/.!]WU*AV+H$= ML$FG\VU-:P;3H4E2)2SWD:>$B4RA`9.?8U5ME!2LU6*\FB%"R"Q5\L0QSU2@ M66`=FV0:6+J$1E>\U6[5>E`_.!+5MB.UKZ3%;V5B.;J-;MHZK;@=9=BD*^\* M(T+#I2NXC"/>E*;;Z?).#:WG-5#]NQN(LV1VA#0TU!IU,\6XVRW>L\0T*[@: M3M9XY24S$+43JZU@2^ZP\H>@;\A;VO$]*2`$HR52SAZDL&E9@GH`:'G(!'.\ M@8_S",5826X?C$>%69;^*E86]90&$!WV@YE@!LP@=U`)!*T,0%MH7^+@9D?% MDJ&5YPZ]=FYUZ^?NN4BOT4LD78>$$C-KC6`C8TL89X7AO#MBJ?.:7DNHRTMB MM%BAAPV>9:VA6;D3$=7;BBR[78HBS#1.?Q^UAM:Y(ICMM&:MK;U;T-N+VUYK M.*W:B]O.6RD2W>OVWK1VK\Y'NX!(!,[5BZP\O4['\[2-=$V,D1@&8P:AT=NXU*I M;)\4>0.6OIB`3^;[6WY&25KL"_LZ4I5;YQMW`&(>IN)[6>(@H&\KFD^)GJCG MHK37<$M!BV1%LK(ALH<[Y:@5'H3EZ.EY_E/7R/(;H!`@G*4=@]ITM?'L6/O5 M;K]B+6."D/8^;ZX/1)!H(H:'-S+0V&!"BA64[MI"ME"YCDT+5LY(N M"A@;I@)SKQ-9P96<#2T%%S#BN1!?B\OK%)AZ'CW+#4RLG"5PJ&G8\URT0ZKU M%=JL-S:3E6G5-C:RI"S)']E4B)%2$UJ@@R7AHSK*3L%2X64&4"O>!ZNU]6=VX0VZ:/]M/2%.[U"!7Y`Q0.[7!;_"'4CN53XS+ M:LFFS,%W\&J)K5*,XL@5-`?[P:LEB.).,G&KT1PL",_!A,O<""-40CQ]YY&9"V_# MW@HW#)B(1$OF4,5KN3<;Z2MV7PVLEA%$F\$KO$+/$'P@I'*4`5AZ>J7#H(G=Q3A!PU`+`K2I1"9R50_I5Q M,6;7$?LWN0`N!.YFIH`@34NQ;A4)$R5=2GJT**U"VD._1"!H(GAL>0<)/7&` MVH26IGW%^XSOX0UJ<^[DSU:?';TN;5H\R*K%R^URJ['U5M^)(@*0=#._AZXB MX2)&.=75+DX"XP6PV_&,ZG"/1??07)B+V3!G2H/5+)&^. M2>T4/<#;#L?^*L#;W7:W!9>R@_%9B[$>QF1'5:P8:P^89OYIH3Q%'1S&2N;* M8X\BY4NH7,##,5-6&E5'WT=FDS,5=,H2TB4QM=')3=;YQP>'N?",AQI_M@:,?9P?X\==<8I, MH0&!8/`15=2BOL['"Q>3N*WMBY9%IO'!V4D6]&5%O7L*2@3C#"8(#)!,%4?( MVHCA!X*%4%Y8@O@A( MHN[ZKKZ)P19\9$8$6\&VIQAL9O*VQD2*R6?RFAQ%KWD'Y$4)DBG2X@F?'">(ZF1TGB>MPQDP4%O:H94;*5[&5['V,Q4+.5'6A MF,6^25NLW+[*^VVL3)_19Y5'7`Q5_<16S6&Y,D$@J5RI<1"X0?VTX0\(=CR[J'4@7@3WQ;& M=*QDPHEP`M",'5:-N\I6D%4'>BK*OF3RLJS@!K3T\KP\+Q4.U2BN%5\3$`3?/`Q'=25&H^B6ZNP M2[>D$"/B;U]]^YS!I/$&A09R.A#+."K^Z$9+#]D3SC!*^F)^<^7%J"T+E!;X]$%.-3SR*QR5<\Z<=V\E@@MX0T-3.0T M-L80>V*R`$3ZT/C\P"G/#K/YK#Z';.-S7?,+?<^M")A$!@5`,`IJG#,:)BZ* M";,?L<]%E4E4BKG'GV+F)N] MQDPS8BM20?])0N.09#W'5MBS]AQ;&6[9L\H%BU)W#?1N[(),RSQ`05+H8"`B M;U_2EUB+=$N$L@MI3AE@P?,S%Q]?63HK./F77\G]$'@0"(H;6;BT7ST'NU'_V8_T0("90T!L`D05/[*F:OE%0-W>**_F[T$V)6Z>XCA:9*0 M6W29.?QOE4SL3^2X:M*;4*DM6.&S2;F=(V>*%F1.)`&0R';`^<)V*)/D2]]`,=/Z=+_%JTT'`6MZ M&-/`B+FD/&MX)S)D_$0]GN@Q90-];"J(%$=BQ;#2X-U?(X!$SYVGK\JE==,3 MYM_A`ZG1(JA+-),`/GP8W#1+(;O:F/#! M%UD>J.VB4L'YTY;;/XUT"-3;0T!M4"LFVT.LIA<@)XF*HV:!]#*4#5Z@J/`5 MOLACM7J`'BJ<'J1[Z9;^+D:U?K+"4Q175[N((!1=.[:[^&YSBSM3DV@?8`B' M%ID8;0.8#)8G17?2BODDE=B;EZOAQ7@=7HG'D7,*P[NS`X:)`4C`LC[K58Z1H+IH532M4D9MF,'R5$SK_EYE2&%?&N5HJ4HWHJW1I[: MV('EB5;6R-LI/5Z-)P-99FT!<3#=JJ_CP836%8@"`%IOU@29!=19H]:!JI[* MC$[2K36@FOX@8!')VM1*_TN.C$-PA8E6%]!J5&>*7A'Q(+8'9:M,UB7]8FFE M&,GJD6@>GX,PF)9SLJ/NJ%"-C@;0_^NHV?X,FLTU!J)"59*+T+=U5-$@($SR MT>H`65)2#5(93R,H+GC66\]4_DP>E56;U*-.THH-CSA<*!UJ7]?7^/6MI9;N MU[\8,6-LY=?_M7Y]7R]R`W:`;6#S*?PU@"U8+]@9B8"M8#?8`7;7FF"/.)!I MV9%'W=:+==YS26[.AXY4(GT0L"V),6U,#Z@`BKLRNYK8)38.,KO>6J45B]VL M,=``RE5:=PEI1E=I%??@7=SKC0V-Y9CO]DF2YM]:R)9 MJ?5JJF>K;TB6J45:W;6>6)R]P^C9TI=?MK]"2UJ4,H9H4ZX^I\GEJZ"%VF]Y8%5'#=.4P/>1T+WTZ;SQ5Z:AN=(9U(2I+6I)87 MG.J00IP%]%WS?T:LL*M/*CY'CZ_(Y(I\IH1NL%/'-55AU*^H4L6)O/F'!I.: M("/$3=,REH$QRV:ZG5:IPW19\8?H%7;#7LY?;Y#8U M#2-BPDJYA%B11A9=327U_?P3RS!R"0B+H21Q[*-:(H.EKL%TR#E%#41LTN,! MG\17F=H!=J`$(X&7"**>"@V]534B6M0>0!`T1(#8CJ_36ZMRLU0K-Q""U;Z; MR"W24DRAL0OW1,+`L$@,SPUT<^_<.G?/?<1-(0B1?/*3L3`,]<_RHH@@B5PB MUN?Q2!',Z`'8$4F[[&5J#JO#V):.5;.>)%]I5YJ/)%=6U/!CYLS'E\C@P3,B M(R'D*-9RS[5G6MH=JA8R1ZZ,EG]H;-G*84O>+DET]Q]4=^M&9/=Q9-@"<3XO M.+*]\2;EDY,J4Y+43MV`LG2O?-J*,6NPI61PV=U]WO&2Z?#4[5%.+(J5IX./ MZ&^Q5+Q%2N8GHD7X(7:3AN@LX.16=L#=;_:[_9(Z86WJK?WB':MW!&1A)3J\ MV3^=G#6<1IP+:1S%!!*MN>KYQDS;' M.Z*]9)[#@K5D85N8SY*ZI"X/D]1\K3PA)R:GQ'6`(.&''"322HDI=]IM-R$"4E]4QG>0BU1 M&Q*]@KFG.1(I\D;(<8]KV1K'!SA4BUASL>W-][W&$+8(>')\@"]E9JX!+H$G MQYRM_MVK-2`,R;!HOMF$EDN8XIM9T%KLJVP(E4@L:$*D(F4KZ^?_]#Z=2"ZX M3P.#WULO>)E$@_LT+9(,;H.OX#?X]5V#GRU-S=ER/X`?@H=**-(:;M(@WD:" ML]].%_TM?]??+65(&7_;WU"X(ZMVX-][3>J,ZU''K5A,E5W.UVU!TD M4!E.@YCA%@@:KEV/X5X1=2:+)+8($)DV`HK<_1()`MI.-/@R<')UHFPFT56Z MT_#`8D%/34SQ!'Z!NVH5.%&)SQ+A@P?TA$+/-8+3:"'R9BE^E!^5ME:-&Q[? ML7;KNL4'74/77!^N^%_3ZS:G<1+M(7X$T5\NSXE%/3?-S3?R?_LPW>HG'H'7 M']]I,VY^->/C:;@&BK6Q9A$#YX"W'RT=F^6$V]_KIQ4>A5?AVSA+J127)>V3 M(8(A0TBDA18^>"S=NYB-(N#B<+74C0)W&%B*MNI)A2PF76M;EB`!(:N6%I6K MB1T&V2S5XHJ?!+E5ZNGVG-Y80D[PWI9WFDP"G8"TI][X!@4W?.^5""Y,?6#GEN"JL5(D8""0B`(3FRCWD^ M2SDYTI3S79@K2Z6_;#JUCALF!19S*HFW-7UZT.-5J-I1!S4_)]<5D=4S1!(EK0W_> MA@B8@B?1K.4ZE3WT-V=D/1?W-.Q=.F^U5,H]@2LA'+ARN;3E-1,K6#K+-!H7 M9$=K.]Q]VRG&"BJ>M-A0/G!^YL+IFY*_:8MSO[$Y#'1FWF=[7.,C1[T(Y2,,@2(9X8.[^*!5\-*G6DNFBOGVOEFSIU7)?]6>-Z=<^>4.7F^ MG8OG8R1X/IZCY^?Y>NZ>E^?BN7H.G[_G[?E\/C]BY[A(?N(YB6>@$Q9X$8L^ M^1+#)0JGAQP=>^7Y"`'PSUF(H)-=!SK\68.S(OIO9E@MFTKZ M4%ZO?-),.N`2=R;I4KJ3GJ8PVL@8R8*E\YR)1V\^7G7I3KJN!:DDZ9A-6_Y@ M861DNB)CIK?E9+I:$]`PZ;2V6G,CM^5[H'U(ZHQ("*F0/FG.`$N',^%_4.=E M=E<=M?UT$8U:(Z/.VLEY9ZZ6"$"PDFAN/NMH]J&+N MJ#O,L1-YGJB+YLS50W.H1^5:'K!W")[J9&2JGOJAZJV@JAZ:M^JP^JM>P+[D MQ#EEWIW?ZK9ZKCZ4&Y,Q:Y8W:?8K??J?7F/6/P!3\O2Q5D[_(%#6T[RAP/0` M3<6A6\=O.KI.!=6RZ@`=T"P^`U:G1;AHZV<)M^Y:8LTH"+@N+,)%VWITG<:< M@M.U=*VN!]7I92D85)_K)SFZ[JX3U>2ZO&ZOQ^OXNB-XKR>!]/J[7EW_Z^8Z MP%ZOY^NN#\$.KY6BD5,+.)!]!&KI]W5A97A\*=S M+B]2X7"N0LWD-#&EOFY&ZNFFK&VNTU`S>=T30HMH98X"L!^.HS26GR4R M>W0-H0.\Z(^A,X\EGS)Y8--I":$"#%I^PP@PBZAVCI/OJK3VT4Y#[7(2*G2I M/:O9DRM:`[7+83Z6_-J#CB&P=D2&96?M\/J2OFF+54DY,\6R)>6]*\\>MK\^ MS,=1;H^OY4OY\`DDM>WGS]L>?KGMA\U4;FW7[8_0W4ZRV.V.X/W#O_;M?[L9 M4Z*>/5H[2T4'US\CW)09"4`B0)9YVFTI1WZ9XXN;827;=4P/5-2+L#JYS78=U\!T\2H*23"73&K12?S&18^/ZNMM,6508-Q*L**H2D=!H;`?@QI,Q M.=L1?6YJ3XJ3,J8Z_$Z*F^[P0+Q.^@T&Z$T'$<"3.'7?Q^VA&(E*-!?2YT7. M8G^D$7*D`9@%^@`U)N'C]Y2>5P! MPAZ*B@\=YYKB:`@%?OA@?91NBP_=N.P")>W`'$A`(J]"2I0VE`USK1W!+O?1 MC$GO_[40&AQN5OMH66TS'5B]O!CJ/GH#S9C M"<8B*9I,TLL];NMD>K22\NF+[,+.RBY]J#PO=M3R[@G3Y:50PA]4>]%XI&!5J=N#$$ M4@NJ4WPWC\5;\;)U`'NTB(U;'?&41\FXFXFI5DGN=:N1H40\DP0;#VRR\=QG MV[6V&YIRI+BV:*K89Z:,?6*/V(>F.!;,A^_<(Q+/V\+=NFT\4EM6>=`G3I7; M47W_+<>('Z)L2I_2Y^PR_+6;5=(L"1UVDO)M^(7UOO8SM8D,(EI"4ZE2^H!1I0RK M<]])3_?0?2?]G=19V#WGI-T[J-L].(N=AB->U%O:KM1HZ\BD$@00)!2N07_* M37Q"8X%3%^;S\/UF-=#C7`X]?:_0W_?V?7Z/ER7T^3U#C]_[]_V]LBKP&IUR MO,KUA2)+[%52"F=OW01-![PVAB])#LYHT6\-!<`C.``$8`&X`"!`$IP-@D!J``5T`* MT`*8!4_^1;`&5!!-?D[PY`\7*8`<@`*,`74`8.`&I`!N``I`!Z0`9@`*`"*D M`&<`"L!S=/D:!B$Q!\P!*0`:@`)X&%/^<*#FPP%=OIJ_X5L+37Z0CP(\!BE` M&H`"-`%00!*0`I0@!0*[EA/,#$;`&H`"2`$N0`I@="@`3SX2T&3P^7$`"I!3 M-/ERRTP)`N0!*,`1T`00`2Q`G5_J9P6R`IJ_850$2@)0H.H;"10!"Y$"O`$H M`!K`YJ,`0T!@<.N;^5D!FJ]A?!:%/J__!K0!RT*P7P:@`%/`&!`MA`AE0`H` MC$P,J1'!0.J;"HO!OZ`#!/IZOA2@(I@!+H"FC^RG`)W^I\^2/_E`0`J@!J`` M9T`;0`9L!='^:?$H^"6C/@I@!*0`2GX9(`8L%TP!@]'FSP"I?IL?$V`BBWX- M`"AH#P;_J(_DK_M!@!AP,H0!NH&ECP6.(4":PF\$&/I4`/&0`HP![+[9P`:X M^DC!KN\Z1/N&/J*OZ*?Y7,"A+Q;0&&&!J@\%_/HL`X#@%KCZ2<"8_P9H&`)& M?D!(<`$I@*K_&$0%:0"L;P?\^M'^LF\?B/QO@*K/&W`!*0ADH.J'`:X^T]^" MR`&[@JK_5;`%3`$9(/';(H-(!T;J$_MM/F`PYW\2JOY2,>=C^6W^>@%EA`%_ M01KPZ\_\9``*L!;\^BW"QV_HC_RJ?AO`.I`!4T&#H6&$_(;^P#\#5/KK?L8? M[8!D4_"B`&D/N>_AG`&!3[40&=#X),#"W(U%'OWP&_OI6?`L3] M:[Z=C_Q\$X6)P`'0#3#BD_79^8!_&@:H'Z[/W^?XZR7,`]5\8[X3H/AO$QA"./_Q M`S]^RSX>H`]0'Y@$K/8I`$A][<#1P*]/$W@K&/.E^;:!DL`Y0*4/$R$V*3KL M%R!]=KX80`J@3R!5Z/%%!G(">H$"@9`/)``/&?@=^8P``C]I'SSDR5?S(_.9 M^3Z!7C[[@AUP`+&=L/C9^0Q_EK[[QY-/]W<`C/TE!:Q^O`*LWU-AUT?L2_/= M`^5^^X$?PE1@Z(?K^_09^L1^+#]3@=EO`W@`9/_A_+(".C\@(`I@"S`#Z`*\ M`_$`"@:?0%2`#P-/O`)CX(P.:!KQ_[#]@GQP@+7@`!/^! M!3\)F8-#X,PO"+A4<`^8'+R"CL&U7]O/+5#I&T#L*NI]-;]=`45`ZX>1>_+A M_=I\3L$SW[+O0P#ML_G]!.4#=#[2'T)0!JCQ:_.%`.E_:,!='_YO#*#_,P3J M`]-\DD#?8(M`UDCYTX#-0##CFL_,Q!:V"ST$-PUI`(EA>4`46 M^SX$I\$0(9W/SA?RB_N!!.=\$\`FW\3@R9?Q._J5!3!]7SX000L@!9`EM`O4 M^]H#+(/!@*KO&ZCK$_T5!U$`BX$4P!P`!:`&N`@$!@>"(+\U8-J/*0`4F`I\ M_`QGF\`G('RP_A?PVP(N^Y:#S4%+83%026@F!!3N"2F$C;^N@5N`0+@%K`O4 M,K@.C\(NH9\O`I@+A`@B`"L"`H8\@ZL/5A@A?/H]![&!IH$6X";0_@<@0/;U M_=9]+`._7SFPS7?FP_5I`@D.!\`\X-T/#3`1,!7T!,1\``(009(0RL`Z6!&6 M!+`1];Y[GY(O#T@@0!1.`[&"P;XTWWLAZ9?V^PRH!GV$DS\G8%PP"KCFDQ$F M!!.$/`=N$/T):82202=CF8Q>^ M^QB%\!"U\MB8X@`)`S&#JT(H,-2H%60*E@&X/SA M!NT#4<%%8+_0:\@-%`FF`,PXB$(G`(A`5Q`D"`-X!Z-]P\#Y8)VP/I@A3`-H M#KN"Q,%9X0L0"A@8-/1I"%L$HT%L(-00:U=TJ/>5"LV!QT*V@EQPJ8`^Y``N M#Y>$@T,VX=K0*J@J)!B.^2:$'P(%0>U0&D@$_"R4`5)]R[Z,8+%//\@??`.P M#G`%",(20?CP6<@1U`>J_M:'4PIR(&#@/=`D1!'B#'6&#`99X:YO#KAG0`8N M"Q^$"D,MX#'0Y.#OTS!,!.(`.P-5H+O0?M!XJ/<9!'H"7`*K(;X/!2`,!!38 M!C5^RSZG@G(`$-CT\P+>#X6&@D-7'YWP![@"G/R5_@*$T$&X8-D@*N`%-/0] M_D"%"D&UX!P@#R`;M`7J#W6'NHN2`%]0X\?Q.P2>`1*!A,#&X(PP9E@8-//Q M"W-]6;^/X;,0!T@Z/!)0_::`/D04`!+@,7C_L_2=!\&!=+ZO(*D`6C#S2_,= M_>Q\!P&GGR$1:<@;J!Q@$&<4]3XMH%I@,/!`#!\J$9\%`<(VWXP0D%CL2Q$2 M#?4`)D2&89?P[T=C8`.T"F5^G\/U'V*09L@M]!J^$6V!8BI!'Z"<$*/@$%`5,@/Y@)A/E-!*:`YT&D8/KAR<<;`!*L M!.``7(*[X;90P\`%X`)8"0V`@$/F80RPE?@`W/#E`40`",1<'Q^1(U@VE`/8 M#A&"ET/=8?.FWE=+)`R*"D!^!T.*X2`1"H@,3/49^G"'9<%1H*R09&A40!`N M!B"$&@8FX"WAB6!+Y`U(!\6$>CYEXJ/@R7<<'`$J!U&(S4)UXDS`?6A!=`P^ M`.6!(\3_8?COZ+M@&F"$:'?9+I#[#GS[0'EA1Y`B:'/8`:<.;W_VO<)@SU/_A M`F&(;41BW[(/IDA(Y`#:^9R&+P!4PBDP(A@QW!X&$MV(!T!,(!D`((@KK!D^ M`!>%`PBC@TN!U/?GFP3:!1.&S4*KH(]0US<0Q!FR$'P'A,&K8AZ0^.RJZ!]J#PKX$857Q!^@C/`.X M$FF'4$.80&C!+$#J(Q[,"9"'/@'['U(@6G@'S`-$"WH(WC_(`-``)=`B6`Q\ M_$R`J4"Q@()`4FCGVP+<`'Y]70`*X$X0!0`CM/2-"2F&Q`7XPJXO2#A`7/^Y M%(F`(\1;83\AL?@Q_/5E!_&`/T,18E@1.C$U5`&*_MZ("0;"(`BQV(?W6_8A M"3]^K4"NH>J/3L`8$"6D$0^`V<+;X:ZOFGA-E/5Q$Y^$WD1P(AQP3R@'=`/0 M`4.)-,/E8M:DWC=1J/HM^WH"@4`PWZRA9NCJFPA\",B&1+^?89O/0'@92`'L M+'I_MD4]'Y[A(&`+3/,Q%7&$6L#ZHFZP>_@]]/1=#AN`&K_:'QK0P<`J@/9- M`1T"3[XK`!$P6QA61#KP#CF);*4L(-B0N&`(=/7-%*<"MD!#GU(14PCQT_^% M#MB+)L52H'$P4G@TY`*R$$`$D(&(GZ'OTJ`V#/[)`=8`,T-\HJ$//C!.=#$V M#`6'HL!S(J&OQ$A:J/>%!2`#IL%B7Y1!+9 M&3>#)41HWP@1V?<)A`P2_12![3_['[-PED@-5"WN%]-\9H#/0(.QXS<[[/1I M_10`@$;!0[TO9L@>(#C$#/.`DKX'X\R/3R@MG"C0#K,"TD*NXIXAGS`5V`,0 M"NJ'IT-;7X20BB@QO!5^&1U_F40\8*%@?:A_J$9$$L=\<3]OP:\/02@#B"U& M%@<4LD9(7YM/!J`3U/-9`7R&5\*/7SLFH2CEZVY4^?A\7KXOX21P`-&.<12B M!%$`/4&C@Y[OI6@Q=/5=&;,"_L`O7R^!@`C_B_,!&7.&T42(X7)!8A@N%!66 M`ZU_&D'4G\.1C>AL#!A"&\M];#X78AD`!K@?D`YZ`>F-?+JIX151,4@)7!@6 M_T@%:(!>H*!/F3A;'"ZZ&B>!_+^HH-8O,0C^0P$J#9N'K4)\(K?P4H@K"`OX M"'F#0<;&(,4QX^@@S/]%%.$A$#5R8.10;H@+%!&H!2F%KCZ*@"Q1=)A;+#!N M-!"%;`S%`"J M#T6)8$6D8X""=]@\-!U>$)F!QK\IX&Q1IQCMBPD@"K6`P\4TWPLQAE@&Q"7& M__*`J#];(HYP]??ET_\Q!B`#=C]/GQ11+B@58#!D&%``O<`Q(=_1@)@X1#IN MW]R'ZD$40#51[W?V2Q]:&M8`(@"[XX]03I@5'")*_6@`D('"XLTQ!HAW1#G2 M$KV&D\`A(XMQ:S@1Z"LN`D6(0<`/`9U`+/``9!G._.B-H@5RX-!QV9=J//9% M'2,,CD>W`);O*[CR^R0X#]<1^P6>X!608%#ELQP2!#>(RL*=HU_0/^@>5/4! M'1&.=L#U']AP,S`5T`T8$7N%]P][#:!@C!CMX_B5%[&+;L1N8@30\HAY+#D& M`2Q_>T7&(V]P/X!4A/;9$Z^#B<4P`(O@V`?K8PZ&_0:):#^LHI>PI'@?C#R& M#4^` M$2U]$XS[`06BWH<4K#TB%D>*KT=<8#11QCC_BR:F^=J`?8*!VLA!_H`_8"OE M!&0$B;,[0*E/"A`$@`(@`9R,8H$T(J^0T+=SU`(N%=R*@D'R!.<+$(&O@(]OB&"+T/)M#F*--\';H=IQW?CDXPBZ M#MM_.,+47])PZ2CW:SJ:$">$#<05P8+0*E@V*"32"!&):0"IWPP@*_!==`&2 M`7R.E;Z\@)[OS5_`H+#_> M%4F(AP838H_0#5@U9"%:_>H`^<5GG_(0N:@M1!VR&KV$D$5\(`22-?@\S"/> M$!V".D0EGS!P4%@.Y/CE`]6",$!GWP?05VC&P0!&%\M]SK[7HYWQ$_DM%""Z M^A2(8;_T8)H/O^ANTB^^_LB1Q#^*H7,1JYCFN_ZU^?@$YT2=XZ!OE5B0!"VJ M(]^,!,=>I%#1_?.VT.?GX^PG'AUX$02GRAX3!',"0@)L`,T8)@/R^)`54`XH$`(#Z<*'H/>OKA@N/$7B_#R%%$#NXYV/>/`!'$3H M^9Z2W$-.(-#06C@:R"GP#,M_+8C8X_KPM#!_A#:""L>%G4#?X-P0UU M"]&3*LE[(J01**F@A"5J&.J&FL6)(PDR)*AZE#HV!P.+L,.V@'Q`,.@S1#^" M"%J#;Q1@!:DO"``,3/5A&!M_ND8BH?IPV<>>E$/B_MR(*X$CX*WPWV@:O!S* M(1.+DL9H@8#1,4A3+!'6)L&`P-I%Y2,"@6\!E:#_M^)LC[8M*/+DECC"G._.Z)G,/#(1D`QTB:Q%&Z M_>Q\[H&1WRDR_&=Q1`@`"6*3',`!!":BEUA=%`*^%Q..6\"[HX8A2CE#O`F( M$>>3C$C.8UD0:9@F($J:(.-^]T(>I7GQY(A83`S^!-@*74J+(W[21ZB'9`I( M!R"%&4@2);IP*%@3;#QT-[Z2^D`Z),+0_P<[A$FF``8/"46-GZK2`8A>W#]R M*7.0CLKO7TIRS:<`R/TE^E:"B#/_W%YA^F6$!`$@(?<+2DA"FA1@A^AAJ/U9 M',.,^,(;I27POB@6&`.L!*9_9L;(8K2LE]@.S"VB`MF3:4=II:=PUS=R)`C^ M(SU](,3I8^^PW/CZ>9')PNUCT,SA* M!_\&`LKJH*NO"(`%@`)(`8H`4P!C9&N1!@!;/!U&`]M\O\@EGS#RUU>,1%$^ MR_2-I#[IX650(3A^1$`:!LV3HTIU(?XQ!0DQ\!V"^:"5'$4$9!`0H/"@S#/$ M^B*/4L69(=31SD=N/"(*#F&.ZTI/W\\0C'@B(/6]*P&`\^2/4"KHH[R;D'W2RA6ARY!I" M#BV.>$8OH>;P(LGGVSFJ!=N&N[X*Y2A2:LFI7`/N$N,@];ZJX\^P!;FA+!T^ M$4&%SDJ3@[H2"`D"^)[`!)@.'@V,0A(2W^@$B`*T#2M]:;X:Y.S1=-E5K$B2 M+5^/CD8#) M^=2/+L8GX=GO>DA=3/J%%4<"Y,#'``RQ1$D5_$G:'#&7;3[GY842>BE-#"F& M$,-_*P>U8,K/5G`'H%[N+@>'ACZH8YMOU%@IU#!\'2>!R[ZQ(6?2'=A*U%;6 M`!)^&$$HHN1QZMA=Q!6X+,L")C=Q(PI@!E!N?/(5`6R!9;['GY%R>0Q,(V(O2O`SBF5$&J M!>5_WDB-9?MO,Q"G]`*2`=J+U,4,X_?//0`8^!8>&9.4A[Y$'XH2)A`'X1WV M!$>%\D/](E/@G&@XI",F$9>2SL"6)(B`D`!-W%)2)SQ_7BX_$(F M,.68)@=$(Q+Q(F!"_`>>%$.3VDH]6;WO))`G;!&4#2&63S[%0)T1OA@]L#0F M"A5_C+](I$P046G#?/*!^8:/N,2%9)?2`RDRS%=F$_&'BDM+)I20P0!HM`&0 M`R66RD(TH8J17`FXK!RN#0^/)H=Q(MHO(*@C1/:-`+=]5;&1`^GAT8?D6_9) M_5Z+M4$>9AA3T?>)U`*F&F\%2+^6I/XO:$A*[`2Z)4F.V\MMXA90O$BC-/ZI M^JJ""$A088&0>:B-Q&*2*6F52,PS`/]1%5FE9$1B_W!&Y$#]I'WRR3=<7#*2 M+GF)%G,%#V^*""9RDP88>-O#-BJO&:N M$K6'3,26XJZO#4BQA$8N!+F&>$=GH/.03Q>KQ!YN,ML#%`%.P7?Q,:GJ2UQ> M_DHW#R`N(702T``[;^D7V(0XA=`/"@^ M#S6/&K^80,1R5SF7]`1&(3V.-\K#)8IQXM?+8#K,&CV*[$I+7QY1/QG`9#%J M$KV/YD?!(JR/?CG!/&32"+.93$+AH)IP%)A%Y!D*$HA^UD%>`4A2];<12`&T M`<*99CZRG\1QZV>0R@&<&(D'ML"X'[&/3*G'U$':`==]?<(R99Q11@A@[`HZ M&D"&G,.`8#00D9@'X`?J_1R0J<48I=(/3_D6#&I&$W6:'\S7I&UR"FB+>/*= M&PN`LKX.X0.S23C75&+F%"J7VTP:X5&R';@CL`_8`14`R,TU8$U2GUGL6VV> M$(V.LT27YCOS&4%?O&8R&)9]R,FQI(%QW;?$/`$*-IV8Q;Y$XJP0*``BL!^N M-[N76,OO92>0TV!7+&_:^W:(1/3LD>,"[B#%&!?,BU@%S2^V?[L_]E!"$#KT?6XA;3^6C*U#!@ M#.<$Z4V)IJB/U'<1\"%@^<26ZD/4YN!0P'EWS!`2"VV`STD#8SISG9D"$$-N M-!)GI#X6YAF2QY>&_"1D!``%79-!X1NR)NA2""#\#3F9BLWU()\`;=D8!&$@#3`/L^_I]1%T?V=202U#I,_3)"^5^;`7XPH@3!GEHC%^2 M(GD%]$L\9ZU2G:DM5'%&$V>3D\30P0&P(WFCA`SP"@>-_X(II'NQA!A-Y$?. M%[^2W4I%(R)S>.AIO&T2&K^4T3X&9QS1F'CX2RT>"4Z34,N4GQMS`1AE%$EJ M!L^/0L;NH1L`R=@J'`V,`5P`SC]O($Q3U\?QFS$6)N\?>0'>H82QY"BIQ#_^ M!#Z`AL1;(4,2S;A:Y"]2%7&'6;\Q@>K1Y%?I4P"D^5B41P(VIA9S#`AC3"/^ M(6>4",-^IGC1-8ATN`A&+2\"P,Z^'S7RR?=H.'2&*J.2P3^#I'_P:1#-%&AV M`D,'X)`4308F>S,2C;!&A:!6N;WLZ+93087SP(,_=O0($H#/F] M*!>-3L)X(T:QY/>K1#4R-O."\)!*DE$0SLD]L`VN/1>:.$CS'Z`2ENDC9%T" M(`P$)8CLWSK"6"E9M#?P^H*1S$J3Y8V2==F1H_9E/F67R,HE9!/R":GZZT!* M"QF/Z,R_WT+3!#D\=%(.4DA]7$,"09HP[VCIDT\J$_F,Q;\XIH,19/@.[$2* M)Z&5?6LWA86"1SS<09$UN M#.V-`$`_(J*P--@BX!*H*X6%`$=:9_JRK`D_7`XH(/>+1TQ0YW%1_;X;8OF9DW'"YJ';6;?8(W)QX148@4U%,""U&2(D`[H'J3=5GQ\VW2 M"&`"J<\D0.VR"="$#"[J*\F&#T3(V;B4$U8><1)C"DV/R-&!F`RTQ@Y5Z.!7IZ``'(+LT"2D@):!1`61GZM%&Z M^H"0OT*"00O'!@H!]7SJ0(\`R\HVWT2@3GBDU")*//F$1L+!)7\S"-BOM`I& M`\.*I(??9KF3<%FO3&[>"VU_,L4GHPYP93@9=%9B*@$$=B\O#A> M&D&*#<)_IK?3%DF_/&56(D6!KCY?`;%@S4@S1$7N'M&&PT5O89LP("B-K`G2 M7NI]E<%HIS3P/*ERK`*.^>*<>\B!(Q3T=-A$/(,&_3:A"/\0X!J M#\&([4/ M*W.&B<@4!*^P7#D([00Z%?:6@X@99$Y1XS=D)(<^+S.;V<\'9AN1T$>F)'4Z M&.>@6,&_)D*R^)?6-(?V!9&&W$D.8\#Q;&AS+(0V*-6)-40F9Q14&AK?-!PF M+N.?;$\[:#L&49@$6`-:/>64:D'.9L5PCK@.+8;&-B^%\-"/:"^R%?COS/X9 M%X*%BTLW:$$41U_.M%YPM,P'6$!!',N M^;Q_VDU^Z$'R;6E&%%!V,`>;%]!*Y*X/'NH73#NV&;F%I-#GH:#/"MC[O&'> M"AF/6\00`4:47`@B"!"8$,.*HH5>X@,0'PH"Y$Z")26BGA]';(%"B(LGVB_:%%<\SBT_M9K^S%EK:%!XF M"HV6LD'#8_RP,RI@N#YR1A3R4<%&`*T M))^5EDS/:$U0_M!P."#K\Q-.0>&6$L,IH'SR_U@7S1I* M"T\#;#\RH.V0E]D[&$32*5>)I,^-1DX`!.%20'UZ/GF@U-'\H(^PN/D>A?I! M*C&BWT+P9(GQ@\E./$\:+766!T"G(4-SL5A17/,I)2^<@TK>)RKT"G!2G&>: M'&U_]6-?YQ/0&Q_7-$)Z8_Z!=Z@B73+21_6<@[YL8*HP+L@KG%9N"\^C:L(9(D3I-.1/!H'C!^6/Y4$[U%\J,TP>K#:7(&"&PN4G,1VJ/UOX#E@]!X2 M_=:,Y#\00:ZP$&GAO%HV.36<=T9>J8!S>+CLNW,:2V.B_<8,A1`H.`Q.76@*U$*Z^KB0FT!%),)Q;0@3A7QV.MFE7E*](Z*00*IUM#B> M!N*.(X?'BON0/XD@K&@Z2&V.RD_F)QQ32;K$'#N:&:M_OE)$X/O1'%EV#)$6 M+=.3D\JDY0$P1GF`Q"&4%83K(_+1T$&#)!ZX M`5)][5"'(P-THS%R6#I,#&2E2-#5IR'0/0HP]#Q`_^ M.?^9X$/B)7A1?#]R-=VGRE-=)7*TL;D.A#4B'8L5];Z-*-5@VKA(?/JQ M*KVA/0$\@*HO"#`2Q/N,(;VB)$L>7TNPS+F&?!1X-'R(:TZ;`#S$S=GF&Q'2 M/;^BM<$$(<>SW(D@Y4]&%HN"]T9EHH>SADH&[$6V^5Z/?\F*HDJP(NF,O/^M M%9N,S-"E).O`[6CL3#_&0X^&'E`]Z>'S.BHMW(AJ17N1_,B]H-!R5/J2[$FF M!I&:WC\FGXTT,AIP7'@N(NL$=+^R)=%1)+H-[5\B+'FH?3\$)=NP\V>.86RH8_06RH-+#+^_42D\)!.18DT9ZF>7#*R M2&^3)<>-(2Y:V0V*<^+?9A+E&1#E$GIMW0!%KY&YK> M(560D-#,9-3S?E"-#'=*&^>4/0%Z9?UO"JB^-''J^2J468%B(2`4(0!C3"QJ M`8\$;(!`JJ#S+?DQW4BF/V&S_=Z!9PVWD>/38J.\=_1H"XWT0@/J!!#8#:^6:(6D(]Z!W00YE63"WR M4PN9BL6O)W1PE1K@'!V^,2^9R;XM9/&PK6G;U!OR(HV#W$FBYI:@6:JFI`K6 M0AVHK4OPIU]0TO<^-/0-1J6#FL6-H%E4>$D0I!C&#`6D!%#;:5)PWZC%'/IY M.U<$;42_(.M4,\@U9#-Z"0L%3`%`7YM/@SD`!&@Z.".+'0V18X#0T,<=Z)D> M-$^*`4P4Y.;24V@Z77>&!$NA3[Y;)[U0/@E'/2!:$EV`T,<(80E3B>@[L*?& M"#&%=TBHIY/2Z0`*W4SF-I&`F\WKJ/F4PJC71)16(QP"\,)B)1(2!\KL*P(X M`8@`*3\7@&"5!1"*U/4%(?D?\$)9*;2`#)#RBP'P_%@`@E5\9\^O`9CT._*<#_G_*`5VA3M:AR#TFB"\&I0%)@#\!;C"QR#$V>A=6IZNA2]#CHNQ"P M`-0`Z$M)7\HORI?O,S1Z,K&CU<^=:.,O'VD?6)_:*YN@M$3[::YR1G@+S9`, M#.06LM(I@%_5;;B-;`7B*I$*)DSGGYYRW1BS)"XT_K:EH4E](#U0LPC:'&?F M&4N*2TGE-`F:/$>$8TZ78$;`+-#=6"AN M#I:F<=.)@9M3P#G%#`O\#*&<+,,VP#FQ#=!/!0ET8"J;FL\CI.44%"E5Q:]N M-+2GI$+`@(#A`)CR"[#V`&0`!T@"*[!21#&&O'_(2@NK/0`:@/,/O\JGVT3B M0S<#YE#.0&]Q:[GUK+">"Q^K7U.I*H_O\EEBY842#)20RX&4'X95PSKD,]Y( M#&H`E5-]HQ*RL`IC-0S*6"E^#8D:JX54";E83?GQ6(VB@`(E:Q%TR(J"A+$& M,,FA/#Y'CF4!R`HD]7P65M^@&D'\JNWAVL@%]*^"_%2J*\$>Q7QQS*I8E5\Z M!%VL"H0B:]0RFDI.A;.V`>2L:8!PHAI`4-C="%!@`%>(P$JPB5,3[!!HM?/A M`-R`*3\8@`Y`]<`"4+2J'GY],%;A9J)UT>I[:+126G]]CM6>W](4UGB2=`72 M#B*<"H2KX@O@:WHNK/,U-9E^A%:\H9T/!K#K&[5:"..:O+Y%']=4(M`1M?0% M(&VF2]28(UXRJQG?=)[N3?E\5E#98RP5*KD)Q%96)B>,XT3K(!?4#HA?W4B0 M`U6LQ;X>@*O/>=HH]&B0^JRC]@\]G\_!MHH"[:A:_R2>J,Z/H"N2G$IY;#-N M6CL:K\DEZ*Y/VNJK5/3]"0P/Z%9>:.K3%PH,;0)(+Q&'FD(``Z2-3P%+9D_7QF3CT,/U1V'[ZUIAE@/2?R6Q&$,U>1JZ:UB\:-Z"5* M$*.N!%?,X\$UT((J$A]J](OU(IU#;FZ&L^)R58#XY!O>W#&`3OX7$^>``(N0<>U_X$!E""B M_.JNNSZ#*QU@D9IP7;CF$1NN4DN(*[!RXHKP@X=87#V?&5 M6V6M>CXF@"M1C['9:(P1*GAQ(I6%%-^>5;8:@J@XAE6=(--#?>N>CX.@_H5=FCY.P^^ M4UFK:8#LP6L5QOI[30.L`&(`"]@40,"/FCCH>ZUF#ZRL"-C7Z@-V]QJ8R%]& M6\^LFE`9J,FM`QLK7;`N6T$$(]>M*]0A0\([W+.V^3BOL=>)@TP@^Q<'*<&B M"9FL;E;=:\V5`VOP:U[*6"4&\92;8H&@Z9IC10%`8#.MKCYIZX0P`:L2O+S6 M!>GQ:6\O>`#;K&IIP)I+XH+!#VX`>=&,*:87L#.E860!KV`+B%W;-> M7D\+&,(H;$`C"`N(K<+B6/&P1E@6P`R`Y$KY<\,B'0J4 M816QC%A#7[FA"8FNNJ=A,+*>P#5NZ@3KG!5;1@B\6$V&#Y:NF5AVLC8<_JT;P!7MT=;A3G:^A5!PHI9:VJ M6!&$3T^YGRTP""F)S,868U&0YT0Y[%[!Y&:.)<+V!EJ%@MATQ'ZA'1MZ;;*J M8$<.7(J`:]0/^%IP';O>%2>O+0A$(?+UXY8* M*!:9LE::J'$3*"#QU#KV*56DQU$]86,41;A6F##:'*F(<4IF*_1P8T4+/(S^ M_!J#*;\Z`()P#4@'$#W>8Q%KO,/!H,"/U?KS[`$<8?D*74<.2QJ(;+;5A1E6E;NV_"Y]S,1/:VE4=)A:?8B& M/QVA5]&<;-R4!D!W?!^"54F8Z=7$8$H6'ON#/2U,#'H97-E:K#WVVW@;>$WN M%6"0FTS>8!I508D]',E^,)6.[T.KYR1Q6;"/!4>^9"4V]3Z:9T)TY,`MB1>R M!)B(*<%M'ZK0Y]"8O`.V.'NJ`-*\I/FUD^J=5//A+"&4*-()86FP5/CIL_.] M_()_=53T7X(0D!I@[*"6!+VB-@!@Y0BU!*"&C`SL%_`'5+X4JDV0A3KIZPTP M(J%]B,\4']SSS>DM<`NJ*GT&40%4P@%P"N`".$#R!::J6DY7(/_QG$CI<_4I M`5P`VSZ>YN\@;^@9\`E&^\I\P`(LWS#0YGAF7$3Z!"-^C$X+)M63$?@3[00R M!\M\K4X=YWEV$_C=U/SQZ3"$?#Z\)Z8Q&ZD7Q7O:%ON>HM*'99EOZ8DI7"WF M_]"SR#ZAY@M2U4HFU";B"F*GTD2;X\!/)4C\ZPE.,>4`&$-ZJ*'O).!>O,U^ M#KFB>]3AX+KO,TLAY,SB'RF1@L,*`NVOJ!K\8P-4^MI\BKY.GZ+/T)?R.PH\ M!2NSG$`8Z\H3`0@'J/1)^E*T^,8+P:X//)GFP]$F`9`-$]@E7QW@+UGU$_CU MR\:8MP$,X1:@V[@Y!'+^SET$W`/EOV3EL5!&8 M"GZ&Z%FAID^00ZCY`[)"8D<#=P`K'PQTN"D$%#"H!GR"2;\RWWG6S#CG^TFZ M:?\"ZEF?X,>3T.=@],\^:I-]`UIE)HB6D'"@=0L$$C4,4X#N'[LR\;>[N0'L`2:+X43I(<>TUFF$-52. M,7$D]3ZVH&]V3#A<<`MT#;23>K[G[`Z@V"?NPQF:$F=]+H!E;9I/TV=YU#"< M$SFHR]IE'VU5;_AF#$S:^8@`T%E*X)?V*^H?1=3&9]5\I40WZF0P""A$W`KF M%,*)%TSP)%*!F64`0:T8(W1/JL_O"!."%^!>`(Q@)X6.YO2;#)< M!`Z8@0$"(*(66#I\_?Q%-?.:=CXI`.N@-_GJRR&6%)D`B$8%`"DR#%"JU?PA M'3"$J`0&`\?/1D#G+/8]`:0`3@`FP&M`UZ`]6$VD^N)^3P#=`$+@H[CN&_@Q M!-:'.('7Y!:`!C"LG40:$UD&F\5'JK%/#H"V1`8",W>"XUI6Y*$6[\DDC/M9 M1Q.#3H43P(YR9\`;."=:`2JJ68'TY`%P\!>K5?7!:@NM_4,[-EF M^C9]A,%GK8V4TM>E'0Z,:W6HO;^4[0O`U.(7HV/DK:5-Q2""6U>\,,X:IJ1V-<@!7/?Y"B*0-L<+9ORRS'>O=3*2^?Y^/D$HP^96[V+=4P*R"U78M2;:V9A]+#K6U6=UN?=08389RZ)!R03O&K(J1 M`[<`R\C-:;,V;:N6I)9.^J*U!\`]X@75TW=.1`*X`%H`0P``;IHO5+M4P#-" M!J^UX=I20MZ0R5>N1<^>:U<.LK[)K'#0'/GRNVWV;8>1O#YC9-ROY7FX1716 M!TF2UEF;HJ%C_@B4.`KN$'.1N3[E[0&0>=NJ1!`^;WV1G;]=P?2V62CP8T%< M;]M\V=MJ8V*Q>SOO?$B&;XFTZ,-='P1VUW>^I<^V&>61N[[W[7[PTRB_O8-2 M_-*AO=JB0_V6!1NLO0&8;5$`2P#I*PCBM;@0H$=V1O^KY\$G0'%VJPGYXP*, M<:5\O0Q$81T`#=`\*(+!ES"Z[C69A!G5&^>#"2U<5JI*_>0(FK!K#XR M$!Y_I-L/(,&V4CMKZ`@<;A.UC4>FK1W1ZL=%O-Q:`/5I?,$F@_.P*/#FY-`6 M)TV%JCZ_P+*@?1J(/-&B`'(`F4NMP*X/P^IBG-5."`<4,X"%@&W@98L"4`)< M,VD%HCJP_.=!P.?H$Y`'U'0][#XS#>6 M.+6`@T2P9YY`B*E592O@&92(O\L41'*R,/HSK*5F$PNHLD>HJ:V2YCA%O1T^ M"SD#?44*((FTL6+QRVB"(-!\2#Z.'P.TAD+011,J6-,/R4K2JZTUR*?,M!E0 M7K.&YD?@*!H4:5@%I!+KNFH.O4/.CB#16Z2->B8(`BR'IC/?0% M`7ZAP5!&8>-3F3D;Z)AR6QF/\E?R*-NOPODUS>BZ0IV(+%;,K+)P#$$BY3JX M"$^M3TU.+DK7&EL4-.K:6*.CF%-FI76T(X?IT]0:$\^E:SXHISGR4HD7?(5R M9'>M%,&BKMXGNB!H/68B=(T`2UT9*#E7M*"-A>BJ/IV0?T$M7T77H_B)O`JN M08T.'HV-!&41J6O23>@B2HN">MVGKJ`T"D`!??GYYZ:S=CZ?`0R7^^J_#8MV M)Q6#-TZ\;DB5F%G63>KZ=1>ZC3Z<@`VVI:L#)91Z&&X#;MLWY\4O&FCU;!*H M!6>$7-W@Z%@9+'/)/I1[-?&2&V/DMUB&DG7K*O4U>T6 M"*2[LDO.[J,T_/H7+&\J,[6O3T+&KIZO^^>9Y>[B4N6&SU`"9GO5*WIH%:&B M(5FS9DZAJ]#U%F',[1/`(=V<'#\G0!"@"5`$\".>9;FLI+Y6Z8G`L-#Y?.BV M=#V[MD4^HS(S",!=3*_:2*6G/5%I8A_4E9HD!;;&-U=_>\Z?+D?WE,HH#*^P M(*:&*;^DY0O`:>C\F\H6^I!\<3^/8`LWJ'LBS3.,,0FZD#Z.WQ0@"^`$8/XA M%(R198&)08#B)*C,/+&:66FA"U1;Z*]2;6!@U?%RU?2JH5<$[^_U&YMO52!8 M75FO:T?7JRMPX;J1X+1Z73V.9E9FX!(UQPL2&.N61U&L0\H`JW^W=0D]Y*-B M>8-]Q\,6*PH@T)K1)+3Z4UVR8]ZLR=30QXOFG;.N!*RBC-4U;P"TS2OF5&8> M93L@8D2_K@04"E`%\#`P_):5N+XC0!6`O^L$H`+@>*N\>`$>[YL3VKK"I:GV M'?>*W-WQ*WJWBKO;-*]5ME@WRIT(6DB34_*%=>`M=24X7+!>9CC14@<14E]_%:YZ":3N9O9U,NN M1".\&DL7($/5+OFN'!IF+5>ORM03Y9A7];")S+JZ&D%]LT5K[];2V5=V9#"P M)L:6TL6`:`O"SEA!7:CV,S6,]%=L*+27/GK"O%+$>`\,(,LA@!0`,4`%2`(\ M`9P`T][8U.(3*2AXW5="+3>)"T&+J"PR68IO9*Z^>L6Q55R_8=P/VDJ@U8GJ M"GVFO]Z$NRYI-/"'`U!#$& M3L^(ZLJJ+##U`QK;73D4+GN,YE3DI2.TCLC$W/5.>[<'\%*;H(740J:E'RHJGWLZ0:1ZV M*D^L#0?^)L7:9",N&<[L:#[X)2ETHB3/,*=8NA:6>.RDRUK[B7!M`-Q/D! M$>V\!$M78[E5G&G3)*C.$G>#KD2BH;H7]J<-E0*V(VFB:T`_((H7`4AGS?,F M(M6P/=X"+.:/*]M+M$AB%?V&:5T>G^/A4=""N)XX=#F?;MV6Z[(R4?#F'/E) M"P^GZAGZ[21& M'+&*5=_;P/;OLHMTO1],2*6_@MT-@^RWP:#,M/V**H6Z6%JX[^ZW^>O[[8CJ M3Z>%=$LW0`M@^.L38*8:?\6?64KL;*Z:').6%,`J;UD`5EE=O0R2 M']NB)U#U+^XW]#=+;"OZ+.&*N=6@)>S3?(EC1&B2&T6)^,,V(A82P_DO5??. M*W.&=]_L5@K3S,=1Q2>:7(&5N%#)X@A6;`+[Y:O*?G&7\TGV99.W#=`#N)TR M44.MC54=@!N`YW=A%7V6_:2\>=X8@!*8"2PY;1^B,/$'3=BI,9R%AHCQ;?2?E._V3^,W!IX0FH6,/!>2"6ZO-^ZZYR@L=GJE0"3 M!,B!%@%>P<'0:0A3Y0G:1.63%,,XL(R2:VCT$RY> M0@E]@BV!48&<537OSY%<^?RV!;3X8@(WV*VI_K*C2 M@HN245`=[38R"`@+AM!.(1V.W]+D*W>2>?C)=!ZN=S&$J-Q@+GQW#6F=&##M M!(^Y*U2_KGZ7OSOF)?`&>*N_`UXNZ^L7"=G2I;?&=*^.B-T#0\M0,O@T/?P. M'6>]IU+M8H77$#A1&`T:.;^_+=Y;;X=WO_#A!3UR_(8`&EVOKA+U:>@:'#"] M>V>\-=Z!0!+@T6ODE?3R5X%]/UX%*G#R!FS?+4&H67.A/.!CI>?3'"P,7?+B M6P.8JE=$9(OU[,HT=>J:>4NM7-[`[DFXOUKGY;>F M_-*LE5TWKY^W(9OX]*"^>P>]A=XP*J*WU+?H]:LZ>L>\@-Y1WXJ7TNM!A:3" M7WF"XE>Q;WHWL]GIQ73F=+V9A,D'IP+1E`K6-6&.>2.]J=Y2ZZK7TOL4E@I7 M<5>A%D\6KZT7&'P'O!="@,>\J%Y@;ZA5V/M,K007>\.^CUED;^-WV2O'3/>. M*!&1T5Y>(`;Q\V#QX_@]`0J]A%Y#K\KRUZ?H9?0>A=7`;`C;[ILSV%L4I&?: MA:66\%^0Y]`QV>OIXPNC>[V7ZE[`,+O77PG@]>OJ+:8`\EYZK[T7WZL&3@GM M>S5^I$M_;V+4G)K_'/CJ&`N^FB M?\6M7L+YI@\+-KYC7-ZP69A5V$N&*A`%.X=C4/_C@#`6C M`>^AJE'LZF,4*UMWU!_2)4&]WM>NHAK8*,$[U(IB5>&?GF!SX.;7)CSHN[F2?R^K1,\'YU5PV?K(O+(* M%<>^=<,$<'\QV>GT*VT.)TNZ-``-JAH8'E%,S0W#>@VAF&$3*$N4?.LOE?R. M*&F]@6$!H/CWK&N"5/BR6/NM'>*>W]RWR9>EW-1R*+6.%!&V(D<)*82$PS-*GF8,&I',PGZD+2:VH8['.^ M-H6*@0$?Y=;2IBM"5`.[,J>&Q%[,L!K0=5J"#%0:?F>%J=+-*8=AQHDISA/? M,-T`?%D[WRP8IWHK[!,")2C$DN+/;_'VS:DHUAL>"2B#/<\\,>8/"\0'/!@B MBF6+:U/\(,Z M)/N3%\9-8,BT`7D'S'UF?+G`)5(*@L!6_5GU]4K4^T2,OKYV\!.3-*H(AM". M`HW$'&)A<9K8D#E7W8A>_XZJS%!-+\(WUCND7"-&>`6W7T@FV[07*)]NYUM*J+MPW8%SB56:"34V_!.*DH5/!GA@I5O,* MB]^^!./(HP+1%.E#4$QF?&4"O<24'P44"Q`,7?3V;_5\D=5681-``MHSMK?N M7]V.#-X(\5NR99SRVQ:CB9N_2\:=0AX6`N6**I`N11:PL%F:2 MOK[!UMC&@M>X)"`K]>R.,6&YW(-W[Q1@<^#RD\T<#)4<>U:V#2G%H,EA;NBED%E+^T[S3V]"MY)6^N M3;FMA^-.X(Q0&$PSY/*V8ZR-I+[",?`W5`H-#!PK7S6)ELQ6PIS`&;QN_')B M"-^RTN#6+%D`0^!,6$MB@V^"^=W];AJ8:>H-1I1:?WNG4(=",.RWI3LDE>I& M9-/!GESYX7?SUHH"5HL:^DB=K\*3HJ&/XNO5K4AR>-/"J&!4:=KTA@K"].LB MA&V\"^$Q+XRWQWOFG5.23%N!K,L68=*U&IP1#KWVCHND^U@F[X#18OAB)0G; M(G:\9%::L!ITRUL2QLC-A"'"==ZRP:Z@U)HEY'O^A)T`B6&A,&.X*-SH90BW M+D]A]T:F,*LWL8GIC0H3>P^A56&7*H67/"KJQ3B:5$FGZ%)3KRT""U2@U!_3 M`<3"QE6G\%VQ1_SL%0U/A%FJ[50PJ=NQC4GAA2DN"[VG/M7^8&H1K!A>20H# M>WN`7F%?X5Y",LSQ.PVGAI5_JV&X<%KVWD@7=0#;33?'V]*$L4.U[AA!#OAB MA7.,HL-,KUF882SL@_OVAI69#M\&+\07,DS./0G;`UNA75U-)8>WTZM2E'AV M?-T`$%@W`(8ULCIBW2L21-NF4$CJZ2EP>?P^Y`[K2$^N0E@:J!UV=FDV5O49 M_HS`D.&C0^2T^DLB58QP#UBN45WTL8UT66HT!+RFBH-]I4SIL5C3[;BT%#AJ M>)6(I]1_,:9S`SPEQOR%2MBIPT"JZ'85GJ@6=.!2)NN\0E%#HI"S-)A&E-HN M+&.JV,,Q83*5/HH(A><^&A9_><@HJ/24$XA+K/KJ>TE]_4HHYZ53\`N\;/M* MB#>_'F2Y[XIW%UP3_1DS%*F$DL<$Z52`90SW328K$#:G6U@SI#*S?VLOE*EN MC$&1302MAY0?%#&WR MC9NE`&!@91PD%SJ7:)0:@B^GP,BGJW4T$5PX]A':',FC4N1'LH!1&;@VG!%; M-H')XV1OYDWTC4@5M7Z*@J>IH6/:HQ.37VI7=$?"#R/'TD3*<8!X-+`(YF,B M@^&&T$'!\;2W2ARZ9!2/CD>"I6,R)*F/0*":?>^JCNT']@/V;7V7)JA"A1UO M&&3'W6``K^TX')P['CGLCCV?P5.X[L0O'9PT'JOJ/IW"F.!><11T1#Q!]F:2 M026^*=`E99;1L&A[=!Y??/?!$@,GJ9GO!_CBM4XD2+CE^ MC^VJF4_X,9)W=AE6?D)ZA,NI1DYRJEB@58A9.$"Z!4VLW>.,(F\`#W`3MA#S><&VF>7)KFC9-!`1OBPC M#UO+2%TW[XKWS^NQ^!\'D`^]`V3'L`$9+)Q`7J*N<+O)E]XJH`-9-SRGU$-* MD.^?M='T<`79=\L59NW*$.'"MM_V[[-M'1@Z5(8H83=`" M1;V9ALS;L'EWBHPCEO7RAI>H#=^*XL.7CZE1Q`,SF$F]QT04L\;7;WG@//GB M`M.'.U_8(8YP#2I>?CEZE!^<%.:5;\_7!3DSC!X;#M&>Y,Z09-47YRL>;(K& M;[6.S\9Y:-/78]JG'&D&":2^/,/D)M8W<_`)U-@F$9?'K%]'8=.UL;RE/#R^45>I`<7EKF3Y""CDO/[M-NV6`UA2Y"B0'FC*O4V& M$^^%GF;U85IXZ#@COIYX#`F^+>6D(9^95:S-3"EV_(9_6$56)YV1Z0@?5`L. M/%&-@\0E`/#CJ M":6,$T:]GRC2",C1[`1V'`$$C5"L8D20;+DS!"F"-H.`KH,_)JAYV[SR-2T/ M@#>14$>,EM&NK[375#P3)?6! MC&.H"F9-L_(2;Y@%?%7Z$/V\1E@F<6@WH]G+$!(WG/&&%&>2)>V7GFEHIO.] M>9U_'#^*H7X9'DPQ3#*3(MV.?.9-*K_RS7#^]0N:BL>Z2#YZ)@>0;PCY*S!_ MG/6\0(DGWY[U(_#F=/X%0)]\R]IXBH=PO7QS#K0VG#_.6]?A`%`BXWSB-0+H MG)O.3&=.[M#Y-`!;UO/&G`=]CU1EY@K`^Y>%73H/6EFMM-^4H"93SW=P/FNR MF.,#5;_B\1LSYZQ>ICH7G7_.6.<4H];9:QM;OOSQ#D^JUF+^8A*5FMCZ2SMO M,:&9VL!+,0?X<%DVC"]_30G,NP+'\9./4%P6WI9:/SL%ZURIZ9(Q#_`7&`VL M!,8`),YEL6:7CPH$-C4?`46ELV:_H2&14GSI-$>:G,.?$=%9,^EO7+MG10!_ M`9/!LV;:;YK/!=O\5?T!`UFA1M90,^J9I-QD5N[*GK'.2^)A8*'0W8P74#Y^ M!MR*W5Y;JT`6M[HSC5:J:Q&<2$.L;^JT\1@A7.SFDMO"#-'V8J.90OB7!#6: M^9R'@,;YJ1>8G\PV'@.W&P/*+V>"LFOGZ[ON%0V'+^5^+RJ3#FQD'IQ& M>0>9N5>.ZGZX\1>E#/A%E(=\H@EA9OZAP+L[+L6^=5F?RLR.\@A14F@Y9D-@ M`&?-=T61$)[_WSU^JRY-/5C=&ZS#[`LD+8@+P]5A.G MEH&\)V0@J/BXH5L^GEWJ0*G0C^6C[&V`ATR[_"TOAA.]!.3'<)47R-H%GO0: MEYO"K.`&"-3&=W5X#L0QY33N>_0>N M1Y>8FNB.1@25,.CK4U7&)=^/IM+1'WE8FTD>K213#+F[)N*%,=F7PV=%7J*Z MGDV@5%%SZDN9RUN"F!KF/OVWI>>"[518:CG&E#WID]?&_F04`$"Y[BM01C'+ M!/:Z;\Z#\I-P2GR?[`$N`!O*@<6'LHOT_SPPP/L@)":*XV`RYA2Z`LI1GD'' M,O<2+-#;I^UXG`0R&T/4H?FJ-=["+C`SA,PY]KL^.!V[?5,GICRU+`I!['UR M#]:QO,-6*4':MWE[]>PJI!=][>53:);RX-GXZU0F%IG'DS]!(A"*9P_:=/R#[O.N9I_*;*6`[#P3":T-QBHS MH3>13FC<,0%:"CW81?#R/=/!X%W)95F8O&M)#B-_"#F\J$X^X7$W#ZS^S**> MH66\:N@;K_;8K\P]1@F+)P_+_0=J+8FP9?GP_IBVGA(N[FN4KKU]ZU[<86$Q?DWN]H>7*-)@713@FP"WG MA/N\/6?W\6EYMCR:CNW26=D33FIV3\V!DZ6T8=9CA=4NO M=@.DV.5>8F=:+(R(;O7N%3G);.&"*484X1A,W?R:=BG%2L;Y[\J!^^N&#OHV M$567AJD``>3FZ5"X,6P4-B`;DA>?YVFC1!ZQ\"R?/<(8S-+RU#`P'I5/,]5KYY6$S2TF@IA?N M%0_4Z4/IJ>!96@CPG5"3-(>$D%'^)AG4^LD)UE8VA_7)EFC5<"9:`IQ/OC<& M`9("RCXIJ&S8Q=P)S'_>F&>.*UX-YA07?-AMGDB[1[G$R.)+LL+1NCPK7&#. M_V+`73_QIA>5`6IX2%-WI$FO&%C#(O>+CKH8!>=F"W%$' MI;NBO<1FLZJP(OU*]?O! M?<_3!-J)@*%:M'"H9(+>"FV.KH-R85':%\PK%A*G>/$`A-4\`&%5#P"V#2?: M@.>_RM])L>O9D&B3U<,J8BG+MT)QL6`3*`$(+@NX?FG58VE]M!(2@'P`5F9> MJG]:FTBDX+JO1NV>[/NNIZG#G)&Z%>>-,RBRG[KBS3Q+%F&M*D&Y,[^:9.FOYL+. MO!;69UUZ)H,!7_#FS$B>+"/.5TR(]Y!%B3+!G.H$AB-<4ZT1Q#[;OB MC=_30&KY,Q)Q#$HS[!1/&"G2PF%?IVP74.D.51;C1P72*E*.'WZR5>@=]@N# M+[N:_+_%\=E00/!D#!3>%?NN]$B*(;DZ^F>>EGBZGK_)]%_V0A79!PJU7D^K M@>'%]T;]9,IP=DA$17D*C7>DL,9!"H64]5N'CH`.2B6ZI48Z)U:T*0UQ/6': M/XZZ:&AP-+[WGWQ^)D>GGV_&H-U[HSK:2W@*W2LJE-_1"@*WX[GXU$F/+D%@ M_GR;BX4HM+9:](K@[4?;JN6,]V!&M"6XT\E)5BMW@Y4/-6(SJ'6:"GI*]3_# M?<6\2V))'P5WBGM311Q'1)'2(T-_]#&8Z)@,SD&C_G;0I&.8H%+9"!M"M10X ME>.[^L9+7XJ1*XV&W@8/E,728&-@Y>V8P)N/=NL>09>5)`$K-`H@+:WF$^]> M8$&BR\`])WQZ%*S/'&.:H?7)U^,U]%XZT-N>EA]K M"P'32-8I`5);P^9Q*RD9@TS M37==B\\G]81U5GDIM*KR+GO4S]-TX)5Z6&SGTU*7B"72&D7\I/Z4E]EB7$K: M'J.N)%,NKS.UP()%ZHX&G3$@G)(RUR"D1!E1:/F>03^P%Z_%:Y)>*9ER,;T=BD!PV+B1W M))JQ8.4H-A(`?1EBI@"/G_O)^-XTWSC:#'RW]B3CL?36"&4;Z=^:H1RXAC/+ MHSV"A>N^)8$`CU6\UNZZKWG8_6)EX?9Z1>POCJ[_3#.3LUBMM-9D'UH]E9+G#/$%ME+(:!ZB2Q3QD"O M$F'69=!I\\O1=+B*5F-GGJNOATJBM6`3.EWH%/Z3@R5/BIW.9/*7E&C0Q!:OMKY/!2*+7G7L>,E-.VX"0T.?D)W ME2_*HE?,I?Q!>B'L&7:>UP+I0BK4W6\ZVO8;_L:@`F_1KV"(K>!K4(2E>KPT9.7V(7B#W'?FNRV4ZMCE4E_V$HDM]I:&:1FRU?!)O M,LFWC,?=)'_VTRO7WL?^1L74_6U-XI#2'#F]O&PW/DM]M>)3J-FYE`S:'BH? M)N/5V-`"+'Z6@/C'U"3**3':5U$$(>OR)H`1EC628@>[W5C&-I+1#0C?]`X/ M'\^5(((:9GG7;1C]Q1Q+M-F5SL/"7SZA/KE73&XSERN.Y5E8Z,S8^6QI5HQB M"@VY8L%[[;=:4%P[-6N./Q_&($]1XO))<=!,Z;S,<"7K?J>BBF*W>@R=DQ+CE2 M`;Z%FDU$(6``+&L?D)C>$Y^%BM!-("-T*K#_)*:.<3>1I071P&;6-'`YG&"` M>P&2@]_W*,5QVP?:G,W<[]I.0.1A\WBM>;9&W!>KLX^9WS3T5)-"FJ]^JC.#BX*2ZP5T-!0TW MDA.W"6R@BJ-O@C`0=?1*NX6-2TYZJHX=7#* MHE>9Z,N[(H&TC'@@Q0-(D[VJ=>`P'_:0=DR2R%RWF.&2ST*F(F%0]FC7?`"? M)0D.G.21*`F8)2UG!+^Z#86#RFVMYJ:[VH@!S%5#,V&&$=ZNIAVYV^QV%'$3 M_6Z4?(()=%G3"*SY+7-_!IB#B>!A\BWXC0CP>U2/L7^H?>=58FIVX1V#OO9) M!U<"?>8HYZJ[:C`59$8_N2G='42!M=N/M'"6;=E:_,(`$>=J-#X69GEU7M8> MN/>L\PB8H,\;T@?T5@G:/Q@++>^B]PCQGDTP2`E1=H'>Y$;!@TXVZKW,7M:V M^?:L,H%Y[-(;R8?U_@LV)'!&A6B@\]1[6*GT+CB3O5>8?T%@Q-`[[8T"\'IS M8:-E4&_JZMN;A5E;*@F@O1>?R,0VY/&AWM>PM$4Z&JZTO8P;)JX7TZG=G&#" M_.Q_$L[@H>@0"(IYQ<>.8A>LC>7HZ4(PC0CSOA9OJ=&5PMEK]TX3'8WXHWES M",G<3V[];S0SYDOKUF8K=Y/1"TJ=:=:61)BOM&=>N5';=&7$XL:T$FFT5FHF M;S"`#`%=;=]RZ+P.-@#:!:G%(<)MWPR2+)']\>;]'Q53#6? M#:^"/0*A7UWUSBSZ-N.@H^F9AUHOH;";S,UT1OZ='<"VN[_$MWGSA6MXQ@+. M?_7+CNM%\%71<-B-37@OJ#6C>;/'V*G'FSN:G#:^*$Y2KQBFH>WBC^''L)N^[%LW45?B@>YB3* M&-.:3$0Q95RQUQTW]?*>C'T,_,4T'U?PTGC`9`**(FNFL.]M=-86CV@*A1^N MM\&2`4&QY`BP55A$)2>F)@\K%,8&[;+`RP#_J ME[6:",ZY*I-;&5U^U4`+?#1$-6I0,*Q<`VK#(-GJXW7VUVW(->(XDHRZ5H*J MKH6M=^,VI.+S*XEM5`8WEV'72&79=3X[-2L-I@+XB:<"<@`3(%M``2!6$`Z\ 6`)X&9@6K`EHAJ\`6<"N4"/8"3P"(!SX[ ` end From owner-mpi-ptop@CS.UTK.EDU Fri Feb 12 10:09:41 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA09820; Fri, 12 Feb 93 10:09:41 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA12070; Fri, 12 Feb 93 10:08:59 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Fri, 12 Feb 1993 10:08:58 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from super.super.org by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA12062; Fri, 12 Feb 93 10:08:52 -0500 Received: from b125 (b125.super.org) by super (4.1/SMI-4.1) id AA19655; Fri, 12 Feb 93 10:08:50 EST Received: by b125 (4.1/SMI-4.1) id AA11977; Fri, 12 Feb 93 10:08:49 EST Date: Fri, 12 Feb 93 10:08:49 EST From: lederman@b125.super.org (Steve Huss-Lederman) Message-Id: <9302121508.AA11977@b125> To: mpi-ptop@cs.utk.edu Subject: Net responses so far (2) Well, I haven't been bowled over with responses. Here are the two so far. If more come I'll forward them. See you all soon in Dallas. Steve ---------------------------------------------------------------------- Return-Path: Date: Wed, 10 Feb 93 18:09:12 -0500 From: johnn@eiffel.LCS.MIT.EDU (John Nguyen) To: lederman@super.org Subject: topologies Some of us here at the Computer Architecture Group at MIT have been working on topologies that span 3 dimensions, but have lower number of neighbors than the 6-neighbor cartesian mesh. Our most notable result so far is one based on the 4-neighbor diamond (or tetrahedral) lattice. I can send you a draft of a working paper if you're interested. John Nguyen johnn@lcs.mit.edu ---------------------------------------------------------------------- Return-Path: From: suhler%xanthus.usc.edu@usc.edu (Paul A. Suhler) Subject: Tree Structures To: lederman@super.org Date: Wed, 10 Feb 93 13:29:35 PST X-Mailer: ELM [version 2.3 PL5] You might want to look into the X-Tree work by Despain, Patterson, and Sequin at UC Berkeley in the late 70s. The initial paper was in the 1978 Intl. Symposium on Computer Architecture, pg. 144+. Despain and I published some traffic studies in the 1979 Summer Computer Simulation Conference. I graduated in '78 and lost track of the work after that. I know that no complete system was ever built. Good luck, Paul Suhler ---------------------------------------------------------------------- From owner-mpi-ptop@CS.UTK.EDU Tue Feb 16 08:48:29 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA18685; Tue, 16 Feb 93 08:48:29 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA20450; Tue, 16 Feb 93 08:48:02 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Tue, 16 Feb 1993 08:48:00 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from super.super.org by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA20436; Tue, 16 Feb 93 08:47:58 -0500 Received: from b125 (b125.super.org) by super.super.org (4.1/SMI-4.1) id AA21547; Tue, 16 Feb 93 08:47:55 EST Received: by b125 (4.1/SMI-4.1) id AA12885; Tue, 16 Feb 93 08:47:55 EST Date: Tue, 16 Feb 93 08:47:55 EST From: lederman@b125.super.org (Steve Huss-Lederman) Message-Id: <9302161347.AA12885@b125> To: mpi-ptop@cs.utk.edu Subject: 2 more net responses Return-Path: <@kakwa.ucs.ualberta.ca:myrias.ab.ca!bhl@myrias> Date: Fri, 12 Feb 1993 16:11:37 -0700 From: bhl@myrias.ab.ca (Brian Lake) To: lederman@super.org Subject: mpp topology request Steven, I haven't read the info about mpi that you suggested, but I thought I'd give you some comments in case I forget to later. If this response is way off, please accept my apologies. At Myrias, we describe interconnect in a fairly simple fashion, and it revolves around a hierarchy of different fanouts. Our last generation of hardware had 4 processors on a board with their own interconnect. 16 boards would get collected in a box (cage). The cages were interconnected with user switchable cables, so one could connect say 8 cages into two clusters of 4, or whatever. It is important that our OS understand the interconnect topology, so we can optimize the usage of the network and do caching at the right levels, etc. A regular interconnect like this is quite easy for a computer to deal with, and the 'cost' of communications between two nodes is a function of the depth of the tree you traverse. Let me know if you wish to hear more. Brian ----- Brian Lake Myrias Computer Technologies Inc. 8522 Davies Road, Edmonton, Alberta, Canada. T6E 4Y5 Tel 403-463-1337 Fax 403-465-0130 Email bhl@myrias.ab.ca Return-Path: Date: Sat, 13 Feb 1993 13:09:36 GMT From: Damon To: b125!lederman@super.super.org Subject: Re: topologies and MPI Newsgroups: comp.parallel In-Reply-To: <1993Feb10.123714.24760@hubcap.clemson.edu> Organization: ExNet Systems Ltd Public Access News, London, UK Cc: pgram@exnet.co.uk In article <1993Feb10.123714.24760@hubcap.clemson.edu> you write: >The Message Passing Interface (MPI) working group is trying to create >a standard for message passing. In other postings to various >newsgroups, people have been invited to become involved and the Would you like to write a letter to the Editor or Parallelogram (me!) about this? I think that a lot of the potential paticipants in this effort do not read USENET, or at least not much. If you want to direct a version of this to me I'll put it on our letters page. If you can do this in the next few days I'll get it in this month's issue, space permitting. I think you might just catch that slice of people who read more paper than USENET. B^> Damon, Editor PgramEd@exnet.co.uk From owner-mpi-ptop@CS.UTK.EDU Tue Mar 9 05:57:27 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA16057; Tue, 9 Mar 93 05:57:27 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA23618; Tue, 9 Mar 93 05:55:57 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Tue, 9 Mar 1993 05:55:56 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA23610; Tue, 9 Mar 93 05:55:53 -0500 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA22780 (5.65c/IDA-1.4.4 for ); Tue, 9 Mar 1993 11:54:23 +0100 Received: by f1neuman.gmd.de id AA17533; Tue, 9 Mar 1993 11:55:45 GMT Date: Tue, 9 Mar 1993 11:55:45 GMT From: Rolf.Hempel@gmd.de Message-Id: <9303091155.AA17533@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: forwarded mail from Lyndon Clarke Cc: gmap10@f1neuman.gmd.de Dear PTOP Subcommittee members, I found a mail from Lyndon Clarke in my mailbox. Since he is not a member of our subcommittee, he asked me to forward it to you. Best Regards, Rolf ----------------------------- forwarded mail --------------------------- Dear process topology subcommittee members It seems to me that from a practical point of view process topologies can only provide support for a small number of special topologies. The subcommittee is currently planning one special topology, cartesian structures. This is a good choice, because lots of programs can make use of it, and lots of programmers can undserstand it also. I also seems to me that process topologies should provide some general purpose capability, so that the user with some other topology can hope to get some help from MPI. The subcommittee is currently planning a general purpose topology, graph structures. These are specified by neighbour lists. This is also a good choice, and I can immediately think of a few topologies which this might get used for. The problem I have is that there appears to be a striking inconsistency, in the working document. I shall explain. Please send you comments to me personally as well, as I am not in this subcommittee. I am lyndon@epcc.edinburgh.ac.uk for email purposes. Now I try to explain. In the first place as discussed at the meeting the capability of process topologies in MPI is that it can map the coordinate space of a topology into an existing set of processes which are already placed on processors. The intention is that the system is made able to implelment mapping optimisations, if this should be possible, from which the user will benefit. In the general GRAPH structure the user constructs a neighbour list. One can immediately imagine that this is constructed by a rule - such as a butterfly topology. Then is it apparent that the user has defined computable (I did type arithmetic first time!) relations between indices in the index space of the topology. For example, the third neighbour of process 9 is process 13 in a hypercube, which is calculated as 9 ^ (3 - 1). The point is that the used defines the relations which exist within the index space. The system maps the index space onto processes. Given the index of a process, there is a well defined way of determining the indices of neighbour processes. The user then find from the system it's position in the index space, and immediately knows the indices of its neighbours. In the case of special topologies which are expressible by the user to MPI in a more compact form, consistency with general topologies requires that MPI *defines* the neighbour relations between indices within the index space and positions of indices within the topology. For example, in a 2D (M,N) grid the process topology section of MPI should define that the coordinate of index I within the grid is given by (and I'm using C style indexes starting at zero here to make the expression simpler) (I%M, I/M) (or it may be the other way around :-). In the description of the CARTesian topology this is not defined. I expect that you mean it to be, perhaps you thought it was obviously the case. Please note that I do not mean to say that the user *must* explicitly calculate neighbours in this way, and of course there is much sense in providing (special topology specific?) procedures or macros which do the calculations. I am just saying that for consistency MPI *should* define the neighbour relations within the index space. For example, in a cartesion structure to calculate the neighbour of a process at a given displcament (dx,dy,dz) is a calculation which is easily defined but if the user spreads it over her code and makes just one typo or error in cut-and-paste style programming (typical of some fortran programmers I know) then s/he will have difficulty, so for his/her convenience the process topology section of MPI provides a routine to do the calculation. Please, if you agree with this, or disagree, then I am most interested to hear your opinion, so CC: you email to lyndon@epcc.edinburgh.ac.uk Best Wishes Lyndon ----------------------------------------------------------------------- /--------------------------------------------------------\ e||) | Lyndon J Clarke Edinburgh Parallel Computing Centre | e||) c||c | Tel: 031 650 5021 Email: lyndon@epcc.edinburgh.ac.uk | c||c \--------------------------------------------------------/ From owner-mpi-ptop@CS.UTK.EDU Tue Mar 9 09:23:00 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA05803; Tue, 9 Mar 93 09:23:00 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA06876; Tue, 9 Mar 93 09:22:24 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Tue, 9 Mar 1993 09:22:22 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from sampson.ccsf.caltech.edu by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA06868; Tue, 9 Mar 93 09:22:20 -0500 Received: from elephant by sampson.ccsf.caltech.edu with SMTP id AA26746 (5.65c/IDA-1.4.4 for mpi-ptop@cs.utk.edu); Tue, 9 Mar 1993 06:21:46 -0800 Received: from lion.parasoft by elephant (4.1/SMI-4.1) id AA10427; Tue, 9 Mar 93 06:14:31 PST Received: by lion.parasoft (4.1/SMI-4.1) id AA21439; Tue, 9 Mar 93 06:14:36 PST Date: Tue, 9 Mar 93 06:14:36 PST From: jwf@lion.Parasoft.COM (Jon Flower) Message-Id: <9303091414.AA21439@lion.parasoft> To: lyndon@epcc.edinburgh.ac.uk, mpi-ptop@cs.utk.edu Subject: Topologies & neighbors. I've now read Lyndon's mail twice and although it's very early in the morning here I find that I can't really understand the dificulty. My impression is that the MPI topology computations "don't" specify the neighbor computations in any way but provide functions that return to the user the (opaque) processor identifiers used for neighboring nodes........ for the cartesian case. Ah.... maybe I see??? Is Lyndon's argument that this is different in the graph case, cos I suppose it is. In that case the user gives the ranking in their arbitrarily numbered tree ... but then what do they use for opaque id's? I guess the tree enquiry function can still return opaque id's for the nodes but is it obvious that relationships like parent and child(ren) don't get lost on the translation..... I suppose you could arrange that they don't. Hmm. Still not sure I understand. However, my feeling as we head down the slippery track towards opaque id's is that the topology stuff becomes clearer becauses to get anywhere at all with opaque beasts you're going to need someone to tell you what they are for the processors who you wish to communicate with .... and the only thing you're realistically going to know about your target is its rank in some real or imagined group. Then you ask the topology call to tell you who that node is. If I've missed the point entirely, Lyndon, I apologise and please try to clarify for me!! Jon From owner-mpi-ptop@CS.UTK.EDU Tue Mar 9 09:39:15 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA06237; Tue, 9 Mar 93 09:39:15 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA07677; Tue, 9 Mar 93 09:38:39 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Tue, 9 Mar 1993 09:38:38 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA07669; Tue, 9 Mar 93 09:38:35 -0500 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA12565 (5.65c/IDA-1.4.4 for ); Tue, 9 Mar 1993 15:37:06 +0100 Received: by f1neuman.gmd.de id AA16778; Tue, 9 Mar 1993 15:38:29 GMT Date: Tue, 9 Mar 1993 15:38:29 GMT From: Rolf.Hempel@gmd.de Message-Id: <9303091538.AA16778@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: Lyndon's proposal Cc: gmap10@f1neuman.gmd.de, lyndon@epcc.edinburgh.ac.uk earlier today I forwarded you a note from Lyndon Clarke in which he discusses an inconsistency in the treatment of simple geometries like grids as compared to the more general graph structure. I assume that Lyndon does not have problems with the general graph definitions. There you have an enumeration which he calls the index space. The user can do arithmetic on the indices if he likes. In my view this enumeration of all processes in the graph is simply there because there is no other obvious naming of positions in the graph. I do not see, however, why the user wants to have a similar indexing in the cases for which we supply the special functions (rings, grids, tori). There the most natural way of indexing is the multidimensional coodinates, for example (i,j,k) in a three-dimensional grid. What does the user gain if he knows that this process is number ((k*ny+j)*nx+i) in a linear ordering? All the lookup-functions which are provided together with grids/tori don't use linear ordering, but they refer to the grid coordinates. For example, if a process is (i,j,k), then his upper x-neighbor is (i+1,j,k). I would tend to see more room for additions in the general graph case. One reasonable addition seems to me a lookup function which returns the index list of neighbors of the calling process. Such a thing would be useful for library writers who work on graph structures which have been set up by the user before calling the library. Rolf From owner-mpi-ptop@CS.UTK.EDU Wed Mar 10 07:51:58 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA28067; Wed, 10 Mar 93 07:51:58 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA10346; Wed, 10 Mar 93 07:51:14 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Wed, 10 Mar 1993 07:51:13 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA10311; Wed, 10 Mar 93 07:50:48 -0500 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA08118 (5.65c/IDA-1.4.4 for ); Wed, 10 Mar 1993 08:36:34 +0100 Received: by f1neuman.gmd.de id AA15791; Wed, 10 Mar 1993 08:37:57 GMT Date: Wed, 10 Mar 1993 08:37:57 GMT From: Rolf.Hempel@gmd.de Message-Id: <9303100837.AA15791@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: mail from Lyndon Clarke Cc: gmap10@f1neuman.gmd.de With this note I send you another mail from Lyndon Clarke which I received yesterday evening. I really think he should join our mailing list. Rolf ------------------------- included message below ----------------------- From root Tue Mar 9 16:32:00 1993 Received: by f1neuman.gmd.de id AA12221; Tue, 9 Mar 1993 16:31:57 GMT Received: from sun2.nsfnet-relay.ac.uk by gmdzi.gmd.de with SMTP id AA15110 (5.65c/IDA-1.4.4 for ); Tue, 9 Mar 1993 16:29:56 +0100 Via: uk.ac.edinburgh.castle; Tue, 9 Mar 1993 15:17:08 +0000 Received: from epcc.ed.ac.uk by castle.ed.ac.uk id aa12023; 9 Mar 93 15:16 WET Received: from subnode.epcc.ed.ac.uk (onyx.epcc.ed.ac.uk) by epcc.ed.ac.uk; Tue, 9 Mar 93 15:16:20 GMT Date: Tue, 9 Mar 93 15:16:18 GMT Message-Id: <2102.9303091516@subnode.epcc.ed.ac.uk> From: L J Clarke Subject: Re: Topologies & neighbors. To: Jon Flower , rolf.hempel@gmd.de In-Reply-To: Jon Flower's message of Tue, 9 Mar 93 06:14:36 PST Reply-To: lyndon@epcc.edinburgh.ac.uk Status: RO Hi John, Rolf Perhaps there is a little more subtlety to this point than I had at first supposed. So I shall have to go slowly (I thought I had). ---------------------------------------------------------------------- Hi there This a reply to John's and Rolf's replies. My thanks to Rolf for fowarding it to the subcommittee mail list. The first point is that perhaps I should join this subcommittee. Please can either of you - suggest Rolf - forward this to the subcommittee on my behalf again. The second point is that I think it is unwise for process topologies to talk of returning a opaque task (or process) identifier from the enquiry functions, as this presupposes the decision of the point-to-point subcommittee which is clearly not yet made. I seems more prudent for the process topologies to say that those procedures which return task (or process) identification to the user will return them "in the form appropraie for use in the point-to-point communication routines". The process topologies committee can globally substitute the actual form when the point-to-point committee have made a firm and binding decision. [Personally, I strongly suspect that point-to-point will yet make that decision in favour of rank within group, rather than opaque identifier.] The third point is that I think it most prudent that process topologies include a general graph mapping facility, as indeed you are sensibly doing. The specification in terms of an adjacency list make sense. The fourth point is that it seems to me that there are one or more special topologies for which there is a well understood pattern of usage and there is advantage in process topologies dealing with these in addition to the general graph. One such topology is the cartesian structures which you are including into process topologies, which I applaude. The fifth point is that from the point of view of formal specification I believe there is advantage in being able to describe the special topologies in terms of the general graph structure. [Of course, in an implementation we would expect that simpler and better mapping algorithms known for cartesian structures would be used, but that is not the point here and is irrelevant to this point.] The sixth point is that we should expect the general graph to be used for structures other than trees! (For example: Butterfly, shuffle-exchange, de Brujn ...) The seventh point is that the general graph being specified by the user in terms of process ranks within the group of processes which are vertices in the graph, seems to presuppose that the user can compute the ranks of neighbours of a process given its rank. This seems an entirely reasonable presupposition. [The "rank within group" is a convenient abstract labelling of processes. One could have chosen textual names, but the rank allows computations.] The eigth point is that in the cases of special topologies for which MPI provides special creation procedures it is reasonable for MPI to provide, and convenient for the user to use, procedures which will return the neighbour identifier "in the form appropraie for use in the point-to-point communication routines". The ninth point is that consistency between the definition of process structures in the general graph case and in the cases of special topologies, and morevoer the ability to formally define the special topologies in terms of the general graph, suggests that the process topologies section of MPI should define the computation which given the rank of the process returns its neighbours. [Of course, we would expect the user to observe good programming practice and make use of the procedures provided for such computations, but that is not the point here and is irrelevant to this point.] I hope this is clearer than before. Best Wishes Lyndon ---------------------------------------------------------------------- /--------------------------------------------------------\ e||) | Lyndon J Clarke Edinburgh Parallel Computing Centre | e||) c||c | Tel: 031 650 5021 Email: lyndon@epcc.edinburgh.ac.uk | c||c \--------------------------------------------------------/ From owner-mpi-ptop@CS.UTK.EDU Wed Mar 10 08:12:06 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA28499; Wed, 10 Mar 93 08:12:06 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA10954; Wed, 10 Mar 93 08:10:15 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Wed, 10 Mar 1993 08:10:14 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from super.super.org by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA10941; Wed, 10 Mar 93 08:10:12 -0500 Received: from b125.super.org by super.super.org (4.1/SMI-4.1) id AA20459; Wed, 10 Mar 93 08:10:10 EST Received: by b125.super.org (4.1/SMI-4.1) id AA01738; Wed, 10 Mar 93 08:10:08 EST Date: Wed, 10 Mar 93 08:10:08 EST From: lederman@b125.super.org (Steve Huss-Lederman) Message-Id: <9303101310.AA01738@b125.super.org> To: mpi-ptop@cs.utk.edu Well, once again I have not posted for several weeks after the MPI meeting. Let's have a vote: Should we: 1) Do research 2) Do MPI standard 3) Do administrivia 4) Postpone the vote :-) Seriously, I have several items from the last meeting that we should probably decide: -1- Should we allow the user to specify a random mapping when a topology is set up? The main argument for this is that you may want a topology for your algorithm but the communications patterns are not those that the standard MPI placement algorithm minimizes for. As an example, MPI would now implicitly optimize for communicating with your +/-1 neighbor in each direction of a mesh. However, your communication pattern might be for +/-2 in each direction. On some machines, the optimization for +/-1 might make the +/-2 communications worse then if a random mapping was used. The one argument I see against this is that MPI may not directly support these other communications patterns so why should the user specify a topology. I don't think this is a strong argument since one should still be able to specify the topology even if the mapping is random. My current feeling is that the user should be able to specify a random mapping for the topology of interest. Thoughts? Now: what do we mean by random? -2- When Marc Snir raised this issue the main point I believe he was trying to make is that you are usually doing a permutation between nodes when you are communicating and what you care about is minimizing the time of that communication. We have implicitly assumed that the permutation is a shift of one node in the grid. Is this what we want to do? I think allowing the user to specify the permutation would be great but messy. For example, the user may wish to minimize contention when doing a transpose on a 2D mesh. Any thoughts? -3- I think we need to pin down how we are going to talk between topologies. As I now understand it, a topology is its own group. So to talk between nodes/processes in two groups requires talking between two groups. Is this correct? I also seem to recall that we are not going to try and optimize the relationship between two topologies. Is this also correct? If so, does the user have to search the second group to figure out what node/process to send information to (i.e., find the rank of the process of interest in the second group). Is that what we intend? I think we should make it clear and decide if we want to provide a routine to do this. -4- The question was raised of what happens to extra nodes when a new topology is created. For example, if you have 10 nodes and create a 3x3 grid, then one node gets left out. As I recall, it now simply gets a different return code. The question is should the node(s) go into a second group so they could be manipulated in an easier fashion? -5- The MPI group agreed to zero up numbering. However, the question was raised if the grids would be row or column major. Another thorny issue :-( One is FORTRAN oriented the other C. Some felt it should depend on the language but this would be tough when they are mixed. I personally liked the suggestion that the opaque structure (that great idea that is not carefully defined yet :-) hold a flag that tells which choice was used and allow the user to specify at group creation. What do others think? Steve From owner-mpi-ptop@CS.UTK.EDU Wed Mar 10 09:38:10 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA01079; Wed, 10 Mar 93 09:38:10 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA15108; Wed, 10 Mar 93 09:37:30 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Wed, 10 Mar 1993 09:37:27 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA15078; Wed, 10 Mar 93 09:37:19 -0500 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA26003 (5.65c/IDA-1.4.4 for ); Wed, 10 Mar 1993 15:35:46 +0100 Received: by f1neuman.gmd.de id AA16743; Wed, 10 Mar 1993 15:37:09 GMT Date: Wed, 10 Mar 1993 15:37:09 GMT From: Rolf.Hempel@gmd.de Message-Id: <9303101537.AA16743@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: comment on Steve's note Cc: gmap10@f1neuman.gmd.de My feeling is that at the moment the whole MPI activity gets more and more diffuse, and that it's about time to refocus on the short-term goals of MPI-1. This will also influence the work in our subcommittee, and we have to restrict ourselves on well-proven concepts. I appreciate Steve's note since he gives a good summary of the points which are open at the moment. I would like to comment on them shortly: 1. I'm not convinced that random ordering really helps much. If the optimization is done such that +1/-1 communications are most efficient, of course there is no guarantee that +2/-2 communication is nearly as good. However, I do not see any application where the topology is a grid, but the most common communication pattern is, say, with distance 2 hops. A more ralistic asumption seems to me that there is a mix of different distances, where the nearest neighbor communication plays some (perhaps not dominant) role. The multigrid example of Oliver McBryan at the Dallas meeting falls into that category. Do people expect that random mapping gives you any advantage in that case? If the majority thinks this is a useful feature, I suggest we add a parameter to the topology definition functions for selecting a "don't care" mapping. In this case I think it is no big deal which random number generator one uses. 2. This permutation kind of specifying communication could be a good idea for the Collective Communications subcommittee. However, not all communication going on in an application is of this type. Sometimes only some nodes in a grid send messages to their neighbors, or the communication is not well enough synchronized over the whole grid. In such cases collective communication functions are not applicable, or they would lead to unnecessary waiting times due to synchronization. So, the specification of a permutation function cannot replace the topology definition functions which just order the processes in a logical way for easy and efficient use of pt2pt functions. 3. The inter-group communication is very tricky, but this question has to be dealt with in other subcommittees as well. Is it at all possible to communicate between processes which are not in the same group? Can group ids be sent in messages? I think that however this point will be resolved in the whole committe, the topology functions can deal with it. The lookup functions for process ids already now contain the group id as the selection criterion for the toplogy. The generalization would then be that one can use the lookup functions for groups of which the calling process is not a member. 4. Another related point which was raised at the last meeting is what happens if a group has 10 procs, but the (3,3) grid only uses 9 of them. In the current proposal a topology is assigned to an existing group. The alternative would be that a topology always creates two new groups, one which exactly comprises the processes of the grid/graph, and another group with the remaining procs. This second group could be empty, of course. Another call to a topology def function could then be used to apply a topology to the rest group. What do people think? 5. I don't have any strong feelings about the row/column major ordering of processes. Probably the best solution would be to add a selection parameter to the definition functions. I think this would be completely sufficient. Rolf From owner-mpi-ptop@CS.UTK.EDU Wed Mar 10 10:51:39 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA03065; Wed, 10 Mar 93 10:51:39 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA19284; Wed, 10 Mar 93 10:50:56 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Wed, 10 Mar 1993 10:50:55 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from fslg8.fsl.noaa.gov by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA19275; Wed, 10 Mar 93 10:50:53 -0500 Received: by fslg8.fsl.noaa.gov (5.57/Ultrix3.0-C) id AA26712; Wed, 10 Mar 93 15:50:49 GMT Received: by macaw.fsl.noaa.gov (4.1/SMI-4.1) id AA07653; Wed, 10 Mar 93 08:49:35 MST Date: Wed, 10 Mar 93 08:49:35 MST From: hender@macaw.fsl.noaa.gov (Tom Henderson) Message-Id: <9303101549.AA07653@macaw.fsl.noaa.gov> To: mpi-ptop@cs.utk.edu Subject: Re: mail from Lyndon Clarke > With this note I send you another mail from Lyndon Clarke which > I received yesterday evening. I really think he should join our > mailing list. > > Rolf I agree. Tom From owner-mpi-ptop@CS.UTK.EDU Fri Mar 12 07:30:35 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA18907; Fri, 12 Mar 93 07:30:35 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA28091; Fri, 12 Mar 93 07:29:52 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Fri, 12 Mar 1993 07:29:49 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA28083; Fri, 12 Mar 93 07:29:34 -0500 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA02607 (5.65c/IDA-1.4.4 for ); Fri, 12 Mar 1993 13:27:55 +0100 Received: by f1neuman.gmd.de id AA16470; Fri, 12 Mar 1993 13:29:19 GMT Date: Fri, 12 Mar 1993 13:29:19 GMT From: Rolf.Hempel@gmd.de Message-Id: <9303121329.AA16470@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: IBM paper Cc: gmap10@f1neuman.gmd.de Yesterday I got a paper from Howard Ho (IBM Almaden Center) about a library for the management of grids on top of their EUI library. He asked me to forward it to this subcommittee. In the following I send you a copy of my reply to him followed by the PostScript file of his text. I processed it with compress and uuencode. Rolf ------------------------------------------------------------------------ From gmap10 Fri Mar 12 13:17:54 1993 Received: by f1neuman.gmd.de id AA16182; Fri, 12 Mar 1993 13:17:54 GMT Date: Fri, 12 Mar 1993 13:17:54 GMT From: gmap10 Message-Id: <9303121317.AA16182@f1neuman.gmd.de> To: ho@almaden.ibm.com Subject: process topologies Cc: gmap10 Status: R Dear Howard, thanks for sending me your paper on "common group structures". I've printed it without any difficulty, and today I will forward it to the MPI subcommittee. Here are a few comments to your paper which reflect my personal impression and which are not meant to question the value of your work. 1. Your approach seems to me quite reasonable for the IBM parallel machine. Since the optimization of process/processor mapping does not gain you much on that system, you can do without it. This is different on other machines where a bad mapping may kill all the performance. Of course this makes it much easier to implement the CGS routines. In a standard, on the other hand, I would not recommend this approach because it is too much system dependent. 2. There are different routines for different spatial dimensions. I know the pros and cons since we did the same thing in early versions of PARMACS. In PARMACS we gave it up because there will always be someone who wants a higer dimensional version. If the interface is not dimension-specific, then you just extend the implementation. Otherwise you have to add a number of new routines. You have seen our proposal for the MPI handling of grids/tori. Since that is not dimension-specific, it is even possible to subsume hypercubes. After all, a hypercube is a grid with two processes per coordinate direction, as you correctly state in your paper, also. 3. In your proposal the creation of a topology assigns each process the group ids of its row and column groups. I'm not sure that this is sufficient for all applications. Sometimes a process wants to communicate with a process which is not a neighbor in one of the four cardinal directions. You gave an example for the nine-point stencil yourself, and already there it is a little tricky to get the ids of the corner-neighbors. Therefore, I prefer general look-up routines wich for arbitrary mesh coordinates return the identifier of the corresponding process. You can see an interface like this in the MPI proposal. I would like to invite you to join our subcommittee. We certainly need people with fresh ideas about process topologies. Best regards, Rolf ------------------------------------------------------------------------ begin 664 ho_howard.ps.Z M'YV0)4)`F=(B")DW8LJTD.$"AH(2)8;(*1.&SALY.D"0L9,&SAP0-5S0N`$" MA90I+VS`6`DB1L,4((:\@9-'3IHS:.BTS('#1HL<,4!("4,FS9@P;$`4R5,& MQ)0W9NC<"3/Q80DJ:>BP*9/QC$TR+C:FL0HES)DR.VS)DT;J#(>3-F2ADZ&>F4P>," MSF0%+ZB4P4+$J$X9-?(6':.33!DS($23-@TBX6,W(%XXR>T:-ILW1#6^SBUD MCQC(9(2;Z;/[Q93@4?`[>._<6/8\=])'2&XJ4G+3D1/&S1PV%9M63]X\ MS?F/9I#.:=KIK&R-LRC@'AMS[+%%7D&U$%1>#'8QQAMN'$5' M'VG`=H-?4J#U!AMUT)$&A!JE8<=??EF1X1P;=O@A;D6-Z(9C(,QQU%8*G)BB MAR":J"&'..)F'WY-O7@&"&W4D11X1%9D$QX@C%&''!,IUH:2:3!)1AUP:(0E M"&R=18<"D\65'%O1H0$"'!UJF>656=8`@I<@A"D72""4>6::\]$QY7A5,I<; M$%O1H9D<>[SP8WXM.<2==R\`6AX9,H8!1QF%#OC1>'70E]NA32D(PJ)_AD%# MH9PFZH(-@0$UPPPXU(##I]V%.D.`AMZ':`PAU5###$&!VNA69)Q%:*U`MF0# MK(P"ZAA2I-JJWPS(_CFE&W4@948999`1X!QT5%3''*OI],*TU;)Q;;8@8-H4 MFA,F^R`<::!5Z`COQOL1%K&^8$03%6VQH%X@>-J@$Z`9(800!P(\<&YNX,9? M;DD4`7!SK,DQ,6]FM-!M&FP4VC`(KX;[*0@?VP99P1#2044>DX(`;7-&I,PO MGTR:H9@<3,*<\L%O,&FP=[MQ:Q-=N8D1QGP@8`$"5>3E`9H06?'K4=+$U<$Q M&4.@0=4>65,EA-5LN"9''R._0-B#/Y///(C6L;@B9VQ/)USU"'&FY>1S:CS:)1Q4T[-&P]] MMVI25_WU;&6_O9?>%]\"'E`YWM>KSI\O/9O4TW4]M.[CGD[<%[\7."\/]0M0 M_J(WO:WXSTQ^"2"<^D(2`H+@>^;#0_F>QT#^.=!Z9@J*!+MW0=`X+PU3.LL& M]3<].K"L*2\0&F2L]T*242IQ<#+@>)JTPZ`0A3T46@ZCDJ<7!31'#G>XV`OD M,`8EGB%+#S,@%,MV!"6JKFQ=D\/7KE8&0LTA#'9H"K2",ADV)&4*"N#?"XR& M-#CY!3F;$:!3B!,U27T$3L_)X0+1IS3<94F'%GM8_ICD$.?1SPRP.6#]FF2\ M\(UO@-ICY/P6Z3SD*<],/Y0D`A%)-SH<13K9*@-'QE`&!50R>=<#G_APHA-U M^4MA`4N0"X)BR$4&4)&`>98IS6$-<4*+12:2+R(4 M:@Q-5!K_7-BR&/+)?S5\T1ZZ$$3=%>UH38$3-'DH!P4<[F,O@!H=I';'R\PQ M<<\A75"TR3\/KB^$,8HD_]XHESCFT9V02Z.([C2A"A$4<3M,G`]Y&:T7)*%Y M35QH4(:+CF$K5''*2R?"K8N4<@X?O0,<@@F"C,IAHSV% M@T=!"H*1EO2D49WJ2H.X4R!PBRITP.A\GDH;:X)5JR&UJDGW\%4YT`&M+=41 MBGBTHM+9B*X@4MI^6"*Z%Q#!"DE(8>G6AAH;*$!O?-L-@D"`FAKL86W/$>$U MA\2_O$`+CDPZ'.!PPRZRQ6ULS?DC4DLZ.."P!S12F-DK5R_\PEB?Y'6MCWLX0Q?#&-LPM,"738X!@I`4E!$.*.F M'#=.N>4;@)N2VHKL00BQ+6@*RV!,9'I%F4;M@WT/G."F(.G!"GCP@C&\%28) MEV^1VDI+=KQAHU! MC:YA`Q\!)H\@)HYOTY7VG8@U_N#!HU[,'.^5K#'N2 MUBFC.I:E.&,W7VVX7<"@=8=\G;33?BDUH]X@:D_=00'YBH.H@Z+JW!`J2ZG. M5X&R1.I\A35+;*GU"S28E^=T&:=?SE<>'LW8.&TH*7`X;$7?4"#GQ#.F#>UI M@7Y*$8M-P=@H9J9WVC88N4B&,I;YDFQ*<]WH)HHE,X`W2[S:+;?FQ@C28TL, M8@"MQO%@)3`P0A$`'O"!`SP&!E\)P@F^<`4^ M9O/SD,@"*$I><@"$%_.@S<)'6! MQT#C?P>XP&TP\%I_?-L[3OFR+WSM/2?[Y35?^X@Z'L1EKZ2T_,\ M\*$/^-E?+G&3%W\(QT\^V)&__.+C(.T93SS0@=_QE_]^]QH_N]FWK_WNJ[[W MKEV\OWBV10;_-,/4AC/S^]O\X#6"P_Y,G_O4` MV'\KH1*[1P0XL']%X`,T0`-1QP-+IW%OEW("MQ)$=P,&&`3[X0-)\&\`IQ($ M!W$P`((0MW41=W@*9X(@)W<.D8($%V\$=P-:'"'EWUR1WPGYW!4R(,*(.'YX+;2('52(B]"`-RYX&'5XZ,R'^+V(AFZ(S\]G;&(N<6`2-YXFAV(+["'`JYX^E2(H$MXF]IXD&F8^=V'T)F9"?:'7] MZ)`""9`1.88Y*'.-]W;,&`.H:`06B'%]X0,W`"U21X_L:'`)5WP3QX)G]XDL MN8V'YX4N"9$@%Y,X1Y$V69,XN8G_F'#9UY#91Y`8B8(:EX_<5Y3=9Y0-"8I= MEY0R.9'^"(Y.>9,#F8.K1Y`HYXY$X)%&X`,GEWL<^'7H>((X:((NZ"J8B(LJ MN(UR1W6SJ(,*1XD1!Y=[I9)S^87\N)0FF(\61Y5_YXQ\"((/R'5]V75_1Y4X M&94J27!>*'=[:'6,"9>-F8[AF(2328,TF'!":()%<')V.00J\8B)1W,Y&`-! M$'L=:02DR94BR8&H*7>L$HACB),BUW*".),LN(4RJ'*Y>7*[:8P%N836"'3` MF7U%L(>`-P3[=YS[!X,CMYQ[AYR+B`-#,(Q)V(XC,00YP"MNF`,WT(`UN9>Q M&9Z4*80)5YXU>(+H68*`1P2U.`-X5P0\!Y\G9P1;R)&]IXJ])X/V67`H1WG^ M68(B-P2S.86]*'(^%XO[L8?O:'2XHH$<&'9B.83I*'')!:(=D^IL> MNHDZZ8\KV(8K(7\GAXK3R)LD%W$ORBNU)Z-2"8ADV9"K8H(WT):4*9Z02)3` M>9XZ>JE_>7`WN*D^FIB>:IJ.66*IZ:JJ0V*.8FJH$QZD(>8,JR"LJ M^(X#Z8,V\(A)B70RP!.[JG$)**.%5X-H*)851XA@QZI1NH0'VG$DQZP]1W2U MF:2<&:W4BJ$?1Z*,9XT>QZ<4*`,)-W(\NJHD&*18B'-?J'%"2(`]R(X,&`-$ MH(:&ZJ#>-X*`.$5X.)1YU4YZS*FJ9&6IM'>J%* MBJ&\.82XZ9@AN!\R6)Q)J*9'Z'%K6HM-6(L?EXT*1P,*FI5F*J^U:'9J]W&4 MZG4LV[(5YX+RMQ(TL(4S*[,R6(2L6JE:9XL\VWBLLG^LPIGB>(2K\J-" MX([8.7"_^F\).8'V&K4_BH5ZJ:^5")>2VJ_,.:3\UY@#-X>0N&\F&7LFN7]I M6HP*QYXIV8QD";:.*82L`K>!ZIPP^+!U>X1WRYQ[9:!*B'"-=X8;1[(-2K%> M:9M_IY82JY)?*(2,>ZHZ:ZG`Z7%_R[/I^866ZYIRU[CB^:B0Z+6>V[D#J(AV M69ID:G0.-P0.ZG-P!W`R$+#VUY8CL7\V8+8V>XA/:+/A:7`CD;N&:)+E&H*- M9[#`*;R^RY8(^X4'6IM8N+#HZ7-WV["8@:&IZW6HTN M^(6QIY%4F)VHD;H?)W.M&V_ZUYVRRW\JIQ*[>X3Q6[>^.[RL6K_.FKS+B[`+ M.[/9&Z@.YZ41&Z@@1[%6%V\N>*!\4:BI*:,L6H(\&&]MF;_YVZS2:\$Z5Z$K ME\'6R',]>;X0F7TJ)\+A2<+`*<%KB;,K^,#6.;[:*:_;:G"M&WM_>@.HN+Z+ MY[XKM[L+![82UWE`_'G.^GO0BH5$?)N\F;T.>\+(-Z1^.H;_>HF7FH_A"KF_ M>\7?AX?N>G(P>J8QW*>WFH@F1[,FIW(C88PU/*%5G+-L/*[U.K7Y>K6.291D MZ(,1YZ(NS&]/IY&%&W!$!X`JL1]_?*TGJQ(FR(QO-X8R9W'96G606+'9AWH4 M:'`'&)=[&,@U1Z@=R2LS`,-KB;B*B[GAZ9M5*9[FF;,[FY[&F[B7ZYAKZIS. M^;!%^[<)U\3C*P0!)Z_`"'(N.'`QZW/^"\P2O+Y86(1?.)NI>LP@1XC.BY[8 M6IW7NA\^9[&S2,U/N(29'_V,)9&:]] MD1IXT^''0:)=7=W#D[*YY;`,P+,]5E[FM28'V>'*1*;-< M-YI<5]1A28L2.I=]RG7I3(N-N5]UX9JR!2`,>*)UM'=<*YYI& M:[0G.M4F"!1I>'&&K'-^. ME]_T+=_\_=_^'>#X+>#VW=\#?N#[+>`)CN#W_8`3Z.`=-P,*6J8>9WLTD`.% MR\B"E\@S>;A)&':CR'L32(TC_G(EC@,36)BC6)A(ZHQ[J9<$28W^"'$/S'+R MW8(0-ZH].HL'JL2S2*(F77'SB9$DN8Q$[N`0B)%)3N2X,K(,BKHR"K4;]WWH M6GGD7-=8OHR)9WW$Q^5;ON5:'G@R3J,X_H(F>-]_=Y(<_I0W2)#Z[<<#J=(% M!Z'D_9/C;><<7N>]%X35%X16*'T;1W/O;8M0VX8*6@1/;K)MG..,_H(0)^5I M+M\)YX40NG&5OGL["Z$X"*%V?:P0FG*[''8_CLP@/H9,:^JK1W,1J.127NE@ M.>..'NNP/NM5F7$1:.A?E^@;>'M/37I!28]6R.%]&7D9*01G/IK(SG"C:70O MV9?&'7,DQYL=QYDY'N/6'JO6QYDO-XI[>>+>3G1#Z7K4F)LBU\TE>-]$3J-K M.::G6>&(VJ^HJHGUW75^.95>Y^&$>84M*.7[#H(TZN]E^'8K6X:\UW7(W77V M5^_M6>\JJN\ZB.]@&.DMN(+\GHX5[X%KSHL%;ZEY.)!*7I$??W!K>840-O)M MVXS?.I`I+W@)]X@;[_(_";4?'X'O_7(*@!JPQW$8*`-!X`,V\(KK2LGR'7>0 MFL@(A[S,")Z.O.`UCLA069-OKG)C7KU6B9M6.>J"*>U9SY]>1W,`S_7T_O45 MS_(L._:]]_6N+J&O+O)3OO53[Y9O[]%^R8-S[X\H2(-C?KH"#JIY+IB_SG46 M!W&QE_-$L//B;`,DT7>LKH*LGLCY6/`:KLBTOO%COO$-#I1#2I`0NO&OOO'& M*,+6+OGK*OIDWO;KRIO7;OKJKOI:WO8PCOK[#ONGS_K@R?FU'_JT#_K[/H:" M/NN0)_JJ_D;**K4_`.%_6C?GLHYT&_OG"H M.%``@UK*R-G1-/L'I`B=:PH_4R_\U#@6I>]^S]Q366_O2%T][M>JUE[W\W?G M2^S)P*YSD6:@US%[8$\3\:2@5.F47!!RC\DH?#PO6N;@C!L,?VR)] M0S#SR3`83H`A\@FS0\<1!?&?YWB`FHH-P4/C9P3:(Y+Y/']R# M%PC=X!P$A'AP#0K"16@(5U@>;(20L`X>PO,%_\I@&,1(Q>E"/#\,=/.@'`]0 MR3ZW9/G&H=6C1!LO-^4IU:?["ISH$W&-[Q8P-88##CU;SX-O30E5]*15*H!$(H&R`0`Z(6&HBN#O%)H80X$*<> MG_M!\J]5#40T1WZD$$4\F!D6=V.#& M&B]J<$21&=U$#Q5VTEW,8W6\2,G%FWK(;RJ@P,M(J^[[#+LIAX(R'\2Y:AGO MJG4\MX;53-!5DWA[K1=)N:NV>MI:^9,[;#'RL46)QQ:Y7EWD<'6QXRF`NJCA MZF*DJXMM$3#R'L&XW7C1,P*,:0XPTL3P`_,4X])1C)'O$>&[1Y01*2.+HHR? MR`/I)4KHYC@CU_DX:(@JR@!/2/]0$(WBAX(')>Z^_?;X>-(A!,(.,8OEH^!('(_C@_.- M+#`0YD8B>`2UW2M,1;)QSTT^UTC\WEQO!'SPSQNB(/G#!?M"$%``TP\JRD.H M5?#:HW$+@MA/]8$XG60&VUXZ1%>$D/?<1_GH#V%=.JR/\/$,ND=%=@R1G(3: MAMXP0<:@;K@@%>0S=)"/CT$^2(^'&ID?`!*-/&T#L;I'Q^I$4>_[?>\Q_VG' MXH>9=)*)=(;[L?AEGVJH(E$D?72&5Z@!.J,!2?_^'H&,;Q'R[$U($$D*9:2C M>H\_L@V:/HC7`&E.D"R$0S)&`DD.Z.$,(']RDHFL^L4W+(CB,!(X2F2A<1.: M1ZX4AP+AM#1T\A'J$4I`J=6E.P<'[IA< MR/.!1@YJ8<%OE?$*GL^R=F=N"0;*DU0$".4"?),.*<])IP?TM59/R]%#:*CP M09B_XZ#^(&Y,?EH//.%&ZUB39!!XVD*XJ.9P-?`TS580J=R`I1!5*CU5R8Q8 M)5R,?5[G%,)*@//Z8J7"0X6S$CP:)ER9"V,!2 M#*)&9G(CVK*6S;`M84O6IRW='%!3=MGR6QX<<+0M MRZ6W#)?HDEN*R\^&!O6?N027ZC)=ODNV%R^7466+8"Y.P3VB*T3C4MGLJW?' M\`@2I-9U\[2D#)!^&C(\KIX[QP]IWEAT:P\S2)&\D;?R'(_!(67@IP-YO+:F MR"[4NHIV^^Z/'<&40^\T#F4,.(+.W^4AT[B,#-!20HU#[][AP)GI[VI@V;.! M8>]FZLP;B#,KVW$D>WXOW?7#390EE\[.ZWDX`,.-I,?8C*8DNC*"Y$SD0!Y[ M27T11@D=.2CF)(^H@T`D!VQ:320HZ)!@E7L]9W--]B7R=\F')S'=XDU8VH[R)-?OFC?-[?/-O^LV]*?_4')5LFG4JP:F= M][@W>X\5$CR/4SO:([.9&AW/<@R;E7-R8D[)&1\G9^1T;Z!S2=XTA)-RFEW" MRTCSZF?E_)TCTMS1 M.A=XTR18]L%X'P?LX$@=2:B0CAJ2FTX'4:%%E+>4_-U-*W.`SWI"')^E/:NG M]\R>WU,.7L\')`1XTI:+/*A.>F6=]'G3RF3&V4;O"[9SJDSB? MIN-$?B@N&%5)^+?X]%WXT9]F+H0.S]$)YBP>(3-RA\RXR4VX*3=['K!R1JK. M%L&GBO,0R12=JXB.\Q/YI1U:D7JHO1MTR/`*\=`?ZCB)J(V#G$9TB"+1)7J? M?"@3/:)%E(D^(*"6D20M!X>SCUZ(^*G4#Z MCT:.*AI2"2_!'[OK,V_19\2Y]\!6A[0D=E&]*E*>Q(K'9VNM''"4HCU2FNI++6E5K.5RL]8 M-$OKI]]1/0+O>JF>!,HZH9_K-%^ECT8>OZE%V9KIAKM(C#.7(JAH.DT;9S65 MIM0TFUK3%4GS9*<06`LZCU?`@/1X-ROI)LVEG709#L'`^4G;*3H]I[$HOL73 MFM.31IX2(J6_2\SQMT@5>$80:FISKL[9'4%)=N/F6Z`+E$NO:NY->Q1/[:DI M_6#@)Q120HD:<5;G`L5'>B?Q&<.,B4VWJ4?MJ"#UFHK4]XA/0ZHV-:D?=:0V MSFX*M;YI,2U-MX>H%4$>.J`"YT_E9X.(96*4HLJ M43VJ)W6:EE2DFE*3JE-MJA]UG8)4EAH*J^K"O*H'TWD^TG@C4W?@VFQ&=HW0 M[1^8-.!@*;6[I6@UE_;2M8I+>6GDX9YIU:W&U5BJ5G'IK0.FOQ0)8B`6.@1- MEIS$H%?'[=3/"-K-6!291*1,,+$.J?&V6'GGUU&LILZP!B%$BC_!7$-MGUFL MT#TJ%NKN-FKAW'"'DP\]U+-J5(6JXNR'@4.GZ-:?1YBRIGS23:E<@Z#>W:W?EKAMN*+TWG*95XV8K@%*S1DWF`Q^O-U5U*5^5J79VQNC3'PE*X>F-[K(W] ML356A+%2*4=\""B!U'`%+N&1G+WZ/"EEPL0^_K!*0MDI5T+_D0:%>L-(4.D_ MW)BMM*5J/+%!I]RA.B)`,M?:UL1[UPO%3291JF8#):LHC,U("!@CV"1X<)D% MJD@BQV7"NK<7=OCLYE1*"6FCX:BV9()0@_.\.'!(G.T;I6D,)=!$Y8I%=.MP MNUA59LF<9+1YIO'B4*6'6?K2*562@E:GQI6AZRG4W-Q>A*+UR5&ESQ-FZX9< M*'2U$2=I(3Y$BW#28_6QJA)/_HVC)4J#P@]F,D>H;NL$5DJK_ZRF#Q)S\HX& M7:2>$(/.6;,UC4RS6`++W_I;ZQFKHE0L\V1M/8#$WFIB14*Q<8XRH@IHR_!6 M),/YIBQ6;OXJYV8$F@);B#UIS1CB2>M745\/>G.I/L`A2!UEQ'^ZHAF:1DBH M+9$CLJC_$)C(>T3.=D95(R1$QW)0PM5!>$\R!;)7>;($KF2*-_^V*_);5#1^ M;D`0R`%;D)SY@"[IV?2?9T,^TZ@UT:<<`)LX7$O*F#.`Y:(>''1T`&[)G47V M3:Y)LZ^6`U#1F4)X>(G>!=T0=9>DF-HC01SJ*'4HI1MAZ5G25;J'"2$!O*9[ MQ1Q7T26Z5U=$`3[3V'BRK3/]NF"79?7!(8U%JJI/MH=S#;F'#JY!N M\?:]I5QGA$%_J&64._PV<9JTN^34())2JSIY=^48)E%:]NR8X#F9>V[@+5V[ M)W:]KJ@%.$96,`7:/C5YRY!4_8P>BN#,)[Z4=#4455)$HM8UW2`_VH+TTJ.C M=31O*](>=W1QBA-/PP&O0O'1TR\JD9Q1-E)!H/?4GB,6U98^45V28:1WJ4F< M?=IMXQ*(4FHO1QDQ*=@;@]36"W*^P3?K-J4'^W2I[C^B8S%,*2DEI?;#<:I2&BM+ M82P8Q*6`ECB"3L)'Y[`[1->:')MWBEQL1]N&W6:J=6`6S3K"?\J?:2Z,>P31#9-?S2R'";"BY98NSZ.*1C!*^C$W[S1R MJ(YG@CUQ:0;PL!2D<:"PFG7$\$MK&=VLI8'KJXZZ9,"L`Z*C?SJTBM/'^5.: M[7M9UY`EOA`=K^AY"\T['<\[]K!H&$0KE7!.%'(J\20R/IF+@.&MLS29ZK`FU4@ZS10C'0W$;H<$#FH) M;2C>J@0(%([VSU<#N-)/L_W@".:X\)3"\90F2@;%KSXJCBY:ECI"K^P($3(E MM[DPD>8:3P12FK%-[J3'X%`?$\(M$.@*)G\W>;W0,AQ628T,\1]0M9`_+T/. M89H7Z4+DM[2!C:\_Y7?;Z_U&JMO+A!)76+)+GXRI.6&_Q5&'*092.DS'7?F\ M5T3>BL`&?,DG"R;+Y)A,DVU%1 M/LI&.2D;%J7,E)&R4V[)&PJ?2<+!KFS>ZP./T\?M6O+F&2F MG)+CW5::*C)YG`S(TB2=ICI)[+@$G4GDU<2(E5B*>1;GU/8%TF1W@`(-R*A@ M<$VF+6V[<,A.*GO0(RU"8Z$'/;E.FH6&T!,:_#"[B!-[J"QJ6F[[H7").F_& MC2J6+K+$FP_LJ.CT";`&4,6:RQX'-]$ABE6;.YF&5-!6D^P\,U[,O/27C^9? M0'I_">D1G(Y*SM[D6?7H):7GEJ:>4>.`/FUMV2H*J(CSKE#.1:-FV+F196?K MG*734X&ET6IV*Z%7J?-W+MD2VC]["#2>K"IE7VE4V.1)0^MYG2SU!8`T3O]I ME6`:`LDK9<2G.6;/DDG#M$MS:>VLG;$9QHK!B!ICA2#1)7>2UFD#N;R1&I,U M(X!/4D/G0[D">1VEEMB>6GK5YZN)I68\];K:NC;ZZNU;]Z5P=K7WT][^0=W3?O"CLA MGYYWX;Q3'=23*;>;N;?@&`/A''/D3R#NRVH]H!D\46:@W*>^.L):)7I5B=SK M+^I;AZ?ZJ!XQ>&KIY"_DC$B05[RKK+2L;_./W$0QY^90QS[;-ONU,:H\'E0H M@6LD"?$R;(V;U=>NQ4705NCMI":*PYW[.E.>K[!Y7Z&@W'Q7B*Y>2\]]]/A$ MD5`*E&3)V/FCAF1X'+-N8TFC6`8G,W/$E<71V-_)6NKF,^I.SW:MU!7>AV"F&V9MM<9;P"OAG[&M*># MWE[57JT3+UDN.#U5L,.1PT.>F-<%_L*+O/S@96W\@!-O-NI+16@8,^(\@[&5 M613*3E$8AK,PQI$!AX\E^\+UB$A9M!,4GNYM%"79AWJM"%WA%I"&>[0FSMK' M!DF@\E28!/#OE4D1EDIY;Q!\C3;[P1TR,ZHC=S;G+CC&"'-;[L/=^@SKDI/; MZE&4VNWB5`%+8R$2@\!+1\X`,FLO._>WLJUB)PG"IW']]K"/(TI[SU5O$D1_ M-_:*-_$^WCDS!JE,$#1KU=ZHM4I8[3.^N8G-]BI=BRO'PDL,AJ&;.@0[DNM- MCP&3;\NZWQ<;42]LW(ZTLO+Q25J'%`5<*%6*37#5TL'X#99(Z6)EV^_[?O-! M^IV_Y3?\=M^5F>FM[U6KOM%W\-1T,U)YAE<;N?QJ]NMQG6JH.(FSA.D+&]^* M[GQLY]*MVI\XA,@=$EI[.;$7A7`_A.IV^G:-Z7P4]9 MI>/@'V[E6>\2GHYB>!4?.#)\TGV=%.[`53@*?^$O7(:+<5E'QENC-5S>9SR- M*V]-I#!_87X+DT9<,#?0K\0TBQKLAM4@AW9WVE9WNX'GXLQO05QO!G*.![T+ M2MRX^@S$SDC+X-<#Y*[NB7N^S3?(']U`5K6Z;<>>+Y7K?7&4WA\ M>FIOP.1Z-&/=]MZ["HG38TK.RB]X*V?1KGQ%_VM83LLQ>"U_Y6`I^:$FQQ/! M-1)((F>*#P112:%)(,(<#$5:4P\?8!S9PGM12ONG$^A:)?-;2"+-,5LDQ/O`.-.+S4H[/51$6Q^?4**7O-YU^N'1ZP?'I/A&?+R.?CA:7,K?> MERV/C..JI+ZV**.MV^BEK*:O!7>DA5/Y>H.9]!B_M>]7!P(?S\@[Y[,3\ZDB MH(3I'-TDA]HK7!FR\"[.UM@A,>"8/S<&=OUV_7]U+ MI+O$_*L"(`F<"57Y4X?<`91]"MD]!<0''`*_X;HQ\D#V[R@=RS<"W]8R;B0R M3R1'1?FX&B?CP].0BVVQ/6:QV.RK4[;C=DA?FKRF*5/N-A+;UE=32=JMT0V5=C:*?Z7.^ M0G>N3N/B]@O/@71.S`*_*Z[NMK&=_(T:G:AM)"/>G;ADAG,]9!>@.>QSI<+* M&!A:@&KVLXG#'5K<\Y(7U][T*!AI]#77QCGM?`U%9.GACIP`'\!9A0]7BORI M03^@3\Q(D0YA5\FP/5/18]5KP\TXQ[3AP[.3=VH,7Y@UO!"N='JOPV?X$2_B M2[P+]/#GVTV2^!._XCD\B[?5D(JAYC=":,>8D65DHP)3X8':4U[5Q5FK<-8W MU>72WG$'-7'G35NCNYJ^*C("*N;8J!N-FC;NZBRR?%UU7$\-`'5AUL@];&Q] MO?ZPW>$53N!F_'$HA;N2MTJ+_6B M-.K.Z_F>A.>7D9VO\Y0[SD.?,+EXE+QO$Y6'CD$1@=O\0!^GN+.M4OYP7VX3 MJO8"MN_LUO\=M-^TENZ-7F+F>X36R:9W,47W,/V[$\VP'#;5.]%EN]_@#GG[ M-2WZ M>806IV@.55FA!W]6T`J:02_HS)MY'_O2ZVPM)(V@#[O$C01R![8FP1>K/_GHV4]_1^VX?0CIWO\3WC_?7VVN>`^%RL$?*LS/4=HJQD/SG]3, MX_PSK^9Y/IS[.X'/\?"%QW.^OER)_4G/IZ$J?:2_]$4LTW_Z3C_J'WVI;^($ M#A9,AE@0V024 MS*?Y[TKWYWZ:)TK`YY=:S'B:2S&GW*M.&H?0P36.1.AN*L8AG&*'-^8<"W9] M/(]"K?JGT_<<>SV4Z_YI1AW/,0_MF_W-_U1!JE3E_)H_FJY]&(!NU9"OFG[% M45/BH\'$-!$.#`K:%W\YL2?JH_7R?NV7/+D3]Z\XW8\YI>KNSZ6_?\YC3N%O M-8E_WC=WQS_W`W^2(^B,?Y+;W57_SD%_*8N&Q%="&XVHWUKO;LHS]*%KB=U, M1W_9-OVI3_Z['-4OH>:__*/_]=]0@ZFRA/Z^C>?*Z]ML(LUFM*ZLE56R2E;( M^G4<*\?A___?8J7_45;11V7%XH`^0=XY)KY<'*>8[9$P>3AO1V"2YOQUJ=.< MMUQM<8!>GL?G[7D:8`LF4YK8`^G]VG`@J!45T1V!>D'8`8 MK\#3C&G_QN"Q)UUIOX@4$A`-(_`1EW-TT#Q!7R%D6*%HQQ/QP3"=,^J/)B.- MM6XWER@3RMPDU!+9)26=+U=2EA5%05!OH!MX1/4A+,,9`=O8L44+]C-'T#:&R_1.:0]E)Y$P):H(S":0,"4)C`H2AI5[$LX2 M"(CD0\)4S-,,(E%BR"BBQ84=W!Y[`_11)9(4U:;-,T(_$3WQ#3I8A((,V/<'#*/$A*U_HA,YHX/PD>/`$;MVE)2+[:K<9,55T=WU!5BM1XX%9J1!D-6!\(TL5I M*5V`B$[#`'96`1H6*`HQ?+4)N.?><'HBG$@DL"DX4\\^:.Y@/(1(3-319$&! MU>'FP?PZ91>O@`RV(B_%I*8&W#P@05#@;Q!@'!4^5H\-A4)A49X$/%O1ZB"Q#B_EA!+0N ME1(C%D$-'0$@QV'#D&%H&CJRAX2(?\K0@HH8(@,--B,B%B28#H/BABAB#Y0I>4?,U<=$)LH(^_74&$:?FLM6<'UP0*%AY-,- M9ZY9D\@DUFI,767BN)!C"U?)Y:D)=?*9YM61#8G?%#YB@1!#C^DE0IUMV')J,I(03?B4G MDTFWM&V*70?BY2ERBHX(J"C^Z$UMFU`W*H:*"QR1D_R M>(=@D(?@'[F!](D9V,J<@8H+.#(RGHPB8\K(RJB,)./(.(KP;$#*CAB/ M:%&M"EO"JF1DH!<_4Y-0;E%>T6F-8 MLLL,)\;=Q:28&6.KRMN6XJ`B,Q_@N('P/TT3L5/XO6>0CF#B.C(B9?RU:D=D..*U?)@Q3$*5X?2-A*-*PP$*9E8C1Q)5C(_:HVHTBWB MO0R)^J);@HB3W%[%V<72Z[!?J^&9LSS. MC](BBJ/"1&=II&@B@Y![+DS>L8'T05R4<1>=W9'\RAVY*M*1HV%',H;Y/)^A M[%A'+25420!CF#Q>B"3+=/AQ'8LDP=%(OB57C_[E@I&*"DH.4DD6CAT**D5P M:)(=65)R<]`B3TO--I=@;:.D($DTS2C/8_^(DR@CG\BW\F)IL2 M-,CXQ)1RHPHR_\R4+:54XE)ZC6L)V!%4WB0-F$F9N4@D863/6/.<)!*=$6>C MR5L9W2^)DR`CD@U*B52Z2RI(,+E[994VBT\)B=55"B1A MI40"07I>&MF&PE8:-8?)6!572B7:I+9UT,B!KE[(,S>=DTX65,G?_9)HS,)5 M2,TH<\E=*95HE?[CAH)8!I-A93CI/YZ5W-;V5J.8E35)6@E97I8U":5B?1V2 MG"7_Y5G&(`!EIN7``2+=FQ96:R$J$T@JF9I9*HECXF(X4I3"2:J2D`0R$(HD=*4H]D MAV8'^'+*S#84B=;'K2$<'&4M"0AR4.G@-`E.SBB^I)T231Z3Y&5.`EZ>E^6E MRL$;H9=3CM`T7IJ7Z:54^5[.E_+EE<1>TI?BI7RI7]HI<"!W27PTE5I8P02^ M46\XB3=98&J5!F:"B6`NF#9)8'E@=I,,YH,I88)WV`WKY9&,4_/(O:>D+8K3 MUQ74.WJ861>(.8!UF!NFB$EBAI@?YN$!A6E0.E3*9GV<7U^(>;A*4B;Q1DBI M/"9Z7=Z`N7H0F.W3FGC0H)+2)%R)D^R296)-4M28,6BE6`E:VC%E3"EBD?@C MV8DH`A30.!@(C=,PVB2:#TXI]3R3'*8RF5G^*_NEA)E@!B/WFQ5T>LB"JYO\ M`;YQ1T%F`_E`7H\2Y)OI9L:9.J4#26?"F4]-`8:3[):[Y12"=\V85E9@J;H( M*AV26;:ZN8/S#=P!X+0WFT@\V:=`K2<'[.F8`4UQI;IC'`21IY??9\B-8'85!`=U5[2.)TE M:/E9(B:;Y=1Y55*=42=BXO#82.@'7YE<64@VSA`Y,?Z"7":8N73^670J' MZ-$G\5DA:BM:5_-9-#J?=@KT MJ7]%G\_GUEH0/'574?I;QU.J5.V*11:6W^ MF"BE8UDFLI\X26$I),:?A&7]:57>GYGE^SE_PI_\Y_[I?ZZ?_V>9"+[T(."( M?*9B1BW\EM##M>DK?Q?)LQD:<6&BO/7O8$>DX*AH>TDE.2.^B4,V):*72(9M MHBQ-C;;YH10U>:.]X_+8C38)O\5S[""?V%TRFFQA>-\6=@097A86[S6CL"BII/`4 MA29.DULAQZ87N4Y=D10+/_!_'5[IYD*!2 M?9?)@4HI-6[H3`*'=<=BG+DH=/)'NHS>4Z55+GU1+F=B9E" M5JD8HK>EX;74%#6C41!ZMT6+ZN!\0C\95@.8?*:%[DWK8ABJ9)HY?=?(88?V M5`-9X#5<2IW\"E,RVE6;C.35HT;"CA'6A"/^)"7%II88BW2%'HK\.21NDX#6 M_Y9ZW5469G$2@:8[61_=9=):`BF0CZF_@@Z$L(I8+J'Z)*Y%EV MC";U9IUE#;U]^=[TI8URF`2H#NJ-#EW2%S>ZC9JBHI87-TOVHD,H9U?K"2'T MW`Y3=L0D,0EC641)E6B'/I)UB6MFZ-K2?;@VBR+AN8G]F9N(G%?+="(IFXJB MVQ12JVAALO@X>!6D1Q))06TYV4R&D^ED$2E%:I%6I!CI3I9R1'14"`82=.2> M<@C_D2):,G%)B$@B#HLHZ?XA!+0*P")CY2':'_*)ZA*"+"@EBX:$UBUE3EE3 MII/FI#SI3NJ3.F4F'51V(0!M'ND,B!897-5E,D:N+&:#X]OFP^$C90>`)50. M)V%+<.*%;F&0!\4B?>P'K`+BTYB\F,U3$MB*<"?>"1X7.H*.GV-3BI9"(B$' M%>/-%"141[)RF6V'KR%7HW),I7BI39*7FA?:X0H6=)0KYPP/,HUH'>Q.P@>( M:%&W2Q-S]$@?6TA^ACWRES!G9$H4>C0>#:O9"M).8B",`@BBC;K6%LE`>F1< MI&>J0+(R\LD`,C==*-U)]:B:1HQ+9&L:,+JF]0EL"C#VBPT++X:;H*84RFF* M=>PE34[E=YC.(Z'A*@-@28R0X1%B2'8PMV9]$AYYGO77!9:XU(/^([AYA)@A M]->^L0!JIB8+BU+0T%R,V'1J%"*%X&E4MX1]IT3A45B>AJ?DJ00(@(2'P`L% MEG4\-&G9,!.CL2!3J4MIGV*9^"E]*I;(4+H9+P,Y)C6`3X`:>A([\(VX4W;0 M(/$,$D*6N(ETFCJB,?$@Q\L'9(N,=Q>'SZ7+J#QJZ>"85!IF&>I9NJ-<3ST5 M#[*9@%CERG$B6X:#!DN*FHO!82PJ/=FBGJ@PZL$2HSHKR=!CI*NP6.+AZ2A# M[5+^$AZIJFAPDPJ<^";&B4*JG(BLA!\8B,QQ^*$NHJ>- MBWU(GEZIYBF6"IX2*/1&E#S^TJ3,N%RJUQJ!YJFIJ65C;&&(># MXXE9$,UYHHQ(HS(H6_J!B"'C#>/#*E`Q6\?U8IZ8)^2);**+T9-2'NM!?!2F M#,JY-YQXCD;JD.JH-JJ0:@V"F!&IE*IOUE+M(>PI:;1<34[.56H&J>`CB\=$ M]I/D+*.J:U:_6&:6RJDZEZ*J*\BJZJIB9K#J*:.JPJJLJF4VJ=8OFV!=Y\O8 MA%\>ACB/1'P%1\T"B9!8)FJYZ:(>JZWJBHJLRJC,JHKZHC:KRBJ,FJO*>0M, M.S@<;FU:&7K2L/1HP5BE%7,TG.!JS>E_W2GD:J`:J!)I1!KSD"R7"&IB:K:+JN'*HQH*RBRY@<`@=G@YK$'"['R0*55JJ1*LYJJ\K_H((]2M!/ME#OLB4JPB@8=&HFH MY[/**Y%)U@HONF;X4WR3LPZI265P=KN,K4L(V4J]%#!F!QOR`^4`24L$*J/^ M4LLJ5/JL.JMQ:[0*K9X'"X.WDTX'))!7N"SXM8HH(CK M[F5\^285915*^;DRFDH&A(-Y?P++!])%8F!D2&22=L"1&^&Y]W88'<*42M.D M'"GH&@NB-\:0R(PBPD81CZ/(M=1O)9;\1^09M)RM*0CY49=,)R')P<&>''P+ M&AF6B20EPLO!0I1,7)ODGGB]((.Q1Z54$_U&9XG) M@N9[EJ8A0E'2F"J)H=:W5!XNAS_4Y56M7(S+FC"E7(#.Y0%L):[^HP(*/68L M8(CC*K^.2Y/K?/9PH"#@7^6*N:*&`R14X[\6KRD8X_-85J-;4-:%*VB$YIZ# M(@99+Y^+_1J%Q:&!ZLQ1KBPN1^5^T*:T*40'G(*-0#*WCB:S$;YV2D'!D!2P M3O''=/,6!HG69F32+=J.1Z)@N;5>)K32URK#OK!3(I)H)=Z6)@@-DB7ZB8I( M::)T9"5;F`+RV(R)MV@`EI9P3#I=`(,D(B_FXJ#FQ&YI4&R=J'"`(Q^CNR@E MQHLF%YF(P@()_=U<#1=;0A?TG3)'Z:1_"$TCF&FT;BY M209:BFSG>=1D7K,(.Y*HM''DS,T6?VED))NMEJ(D8-L'H8.RQ#QDU[D#M`4B M?=4M*7LA/_`A;-HPRJ0KQ[.CK7RKK2QZ0CX&@^7F4R2,'NP2(')4N)R4RG24L;(]=\N/D M(#AEG/;(]K'8)_?9R'FS?BP&ELV.LYZ*QW1PGCD2RG\:<-)+.0B<*B1EA"TB M\S$>RAPJ"!;GV3RS*\R,PL/]+D0:NV3-DK,1%L#HR(:S_VHXJWURLZIBSWEW M$IE[ITHB:):=7.8A"LV2?CK7I^K7*#@QC!J"EPIS2M/773?M\ZE\Z+?6) MM=DI<6=0JZW(UPZ1`7<4!2:DC2TO,`+)A?"BD'0E'"*.1EBVPH M5V0@:]6>8RK((AFG`!Q>+2WK:=8NH"99.]:Z+UGE5ZE8JK5IK?^A6!9BAXEU M=7N\M6J98KEXVK46"UZ[6^::_BA?&]>&C4Z(WLE5CIV,B$T4DAFVX`?`Y>R% M+\?L-Z5@@26T"3Y;EHB0$@OF2=),CIIG9NMYAB'UW-ZLDNV8^SYEQ8BSE8C@[YV-5=C(-**3"->21]T0IFHON?)(GS6B-8([6*- M8*J_RGU*G]PLIRA8!E%`D%2ZZH9UL'Q"/_?&"X&-41R!RK38C]RJRT+UI*O)$X_CDDRN"C MANP;M0=9VG$.\:=]@?C$EW) M:QIMUGAUG"XHC(I"@9AH`TA`%($TJASN&$9YE8BR&?Y$05P2W_"VDNZ3MMR<([\5XX"SN2I&&S'"KU"RZRK!\ND8,J,OICF#'Z:@[ MY!@=@`ZM"^SBNND)OBCL'KO!KJBVZM8B>L?S<=^"./A9H**,9;J`B,D%OE(N!*Z" M*Z2M,CZ'L:OM?KM!VK:+[=Z3X*Z0MKP(:.-NB5KU54E$C?IA=_0%V^5D=8)@ MNKMKV>'IAKHE2*FKL.R[G:ZHR^_>NZ;NZH'J-D_"$%#P&]*,-4B=>]!5(Q"C M99LO4H2ETL48\7:[F"?$B.AZD`EOQJ@\3KDDKI3FWS);#)V`>^[>D_OCNJ,O,_+O^;K\;^5*^ER\U:V7))G.3R;B[$;PU+[34-O&[>\6!A:5QG!`7P:K8I8!^:PUKXN%6XDW!J.^F[&@INI+T:*^^\;CZJ_8KU`,\QN5<2P$Z'L2JTB_2LDZDQY*(F((ZJ@E MRF=M:\W;M`B"!0DO)GP@(X6(%*:IY#>Q+]M(BF98=);1^M0D=0EB5J@[M45$ MC]P$QY*XNB]8(JTL+\W+Z>O_JKY/R5(#@["2KF_MVH/$OIFN[MJAVK[_JY\X MA`BPC(]34TIU(+C*#4+S4KG36%JH+[0!?P$VUIW\A*!I")RSX6PD\,]47.E# M6>5U&-&.2_SW@H966%.W,[A%LW! MC71C%URGJ!PBVQ^B^KU$]F_H6Y5I5#9?SS>+#H%L254 M(;8BV<%8QY^7!V>`/X<>W.%\,'1PX`9=J:\#W<_6Z\Y-6I1FE2 M1A])U*G7835<`@G_`S*1KJD*G[26N+M/;S%D]050\4L."'.LHM!'0AK3N,&H M\,ZG"O=\0E/9UJ)PO"[*'L,6D&E&@`%BUS!;.,BT,^W@*9>2!/)X7$J(5)H" M[4UI'LQ=0S<@U)B%LI>CBTZR@:TTL57-0>3;'RE'5 ME#F;T3UZN;8Q2B.26"HFQ3]H(3)+AB7\XQAI$G62OXZ[5A,Z4C-QZ@)*EED" MU5]"2E51)TGLV!3W762QZUA%12[."=SC;8J_+^\]TGU-'!/G0NS*IC(AF4[\ MD1$B&9D&*=1HO`I'1@S]Z"!OR&HBO]V\+LBN!8<=,2T;CZ/$/%AM(VI(>^&" MP8CW:`388%$>>\C*=KL%L4*<&7/%G'%ULNF2OCX./HC;Y+,(55]<#P'&S-Y& MIP_CS1: M+1=ODU*3N%<'*%PS09(\P9#;!Q@WA?\8K-L03L>F:'7,,J&B!`=V+*$T8#*)%6H[L5* M7J)G1P:T.?XF0DE"M>Z^-Q'02P5[U!X_#^;7QC$C0:>!653*3P?-5`K\.50K M98/Y(#=.#3)1Z2`;5D7G^!0P*9@R26JJ4D+(`BQ-B5/6E#+E3,M3FLCVE(-) M!S/(UN0;N`_9(O20P;0,!UL1#\#%J@0L,^1L MXC&M5"1#DL'E0&E0+LFCJ))\?3DE%,OHE:MPHX0O)IF09(YPI1C"J(EDNI=3 M["?;R5$7G6SN-+2(5\ MITC+V;*U+"YWR\RRMUPN-Z]4++B<+%=(FU=/RM/S*P:PPY[0,,T[K,$LH"_/# M[-.6,$44TF>-R!S3AQ9*=.:(+"_5HI4>:2&)3`G$,P9?VC$JFC)7 M9N>A(ZF[$"17*,RLNQPT$L=(L`G/S#=SN/EZU,R?B$>INWPB9,NJJ;O4S$2S M%BJ,U,QA*-)\-&NA[";2G)I*,TQS6*B%>J+YL1Q:-2,#XV>YN]PY2LR8`)8EE@TIW35R$%==YJGG-$=)?;Z7:*!F.=7F)$3C9#F/HH$MZ.8\W+HD>T50 M8M\HOCDHY:5`34!+\.,YDE`CVDPM0KRV+IHN97+3Z8@:TZ2CKKR_[]=_>C"W MC=:.Z+D#+4#+TGP5P!TEW_-C>;?$)K.*^GPFOL\+6$!:B[0*,??MOF*5%R5I: M^R*L2#1P,JDFO/C:^QB"@)!'[0DS*D&\$*]Q:$4#84IT@Q:5;M$78WSBFFHC M("1R0G5$:WB(`H5%(<*I,6:BQ$J/N.EMPJ&0(1'S3AM'QR:4,#KRA3R4YXAT M>M<(.`-:4XCH<%SN*8:JCRXA1(D1;4@OP$YJ%*+Z[C*=&;ZX<_ABOMAY@J]Y M*+"L,"9)7]*[[`%T`&DT_>FTE%>6T+FG61_/1 MT,G(=+:R.O^<#_$2PI.P)=9[/K4=8E^L>ZD4.I#J$@WF=KF\V?=!)AM- M%JR9.H3P.X%6.[VF;M1)]-HH-0U8L8BSJM$8OW/J8UFGAB!V29XJ>/2IG<@) M5YR9JQI-*[FBXJK`<@02^!@Z5Y0I'9Y9=$,T\]12%ZE`]9B;LX:M035$;52? M,M[40/TX1KT32X#TTQ&1'@WX!QS6JXNB<")32]&PJC@-8#D$]8M6G;ATU1+E M5@U69]5A]5<-7^U=8_5DYE4+6EBU,)891G_57U-8H?(\NLRTYEPY!,T*5V/" MF#!`;S_M5__3UQ9@?>.VA@&U;R-7]URG-))$I+`UBK$WLJ>F*HU+X^)+3M8` M"65-S%;6`$ED39TH,Y)++K:X4";XL]QT6`,J=;6=XT2#?[FM9.9?V28[2QBI M1:\<#AH7_5K/UB><%\U:?]'#ZO!RAZE@8XY&W5$CTJ+DHQ)^H2+23Y.:E@D< MFDT.^,M"@I1T4?U0#VH"KNU7DP21]>*,3O7^#61VH3L MUU2AK`F5QC+122Y5Q+5>84U/?80\.`%KLB)8_[SV85^]8.?3R6HUC.A:>;,FB3(D=H>PI-Z_`_($4-N4`JE(KCS:J#;*^V MM-/X22(<*UANU#LUMPP`[@) ME.I:RZ$,D+4F:-J*M6'/:2=*1A.^P2)NVD.0C,4F>=JG]:9O:0%*.M&J_ MVJIVK)UJOT''7E6,D:0B1'"MK7HX2RC9QX4A]@7_V*M&QBEYM1;3HI@%KBOG M\S,"'I%;!^$BUZV!GUJC0ACH`)2$B?44!%O/<*(3QI^))UWOKX-K_SK,DA%, M$T)@__-ZU')\4VE(X5-N9Z=SC^"\_-"#$L\\M?H(>:@/I]3QW#Y_CJ(;8`%, MYI;7^'&_.QLJ4H+&,7(S]_0$=,<@:0^[E\X& M5%);"C+)N4#X-F2-GX%P.IP'%YDQJNE3+;95;SZS#F9RV/F7UQX$=J:>/2N+ M\29PF\#(6XW$#W%VW7;B00]U)SK/U3VT\:#-UF3S>,W<7#?/],TB3BP+(V4K MW4I:'LL26EE#/[<]UGGW'OOVCN)U,]$`C^J2?YG>L1(E1#A;/9UEM%6J=-Z> MDL-SPHP]D0<(XNW]V5(+"AP/I=MM7)[4>_N<\U7I%7Q_3.(*O^-/$4H"DWDV M255=9)^,19(P3`3J&!1]%R<7(G<"_4PC!YHI.[25GNP-),?W-'``=-84N`!V MGRW%U`TFVK!0-(3N+$.=Z>B==JI,VL[]%7\3DI#6('D1^HM/Y@V$>:/;9<_^ M#3--W?WWK@1-06?\M_Y-@-_?'%W+XGI3(DC6(-EYRSW6DQ6B,LG<2.$RQ"?= M($:V*`FC?,M*OL*H>4DR>`B"J@5>XLBZ'=X@D#=R3OX;8($12MM$KZKX5Q!X@[)3&">1I"SLD0R MB?1C_X,D,%?X2!FF@E@JTJ5Q`G`JHE:240//W+^A!A>$) MCA1T:!,ZHHDN-\318VJXO;6K?(EZR**'*OQC4XYJN[(&03F>JI-YE9[$7&ZU MYDQ/:XZL5/E4/;PP)=4);E^C-Y_',BGBUI`5?!\BE!W3!+WA$.#_M\=#B?O? MF#@!KNI`T:WM?P9$W%$5'=&\XXV&"\S1?T"EKJ%HW$C>DHXRK M.,AY#+NC.?;T#A:9(U+X]4X-DZ-1^/->#;>C6_CU_@W'HY/X]:X M\]TAG>,^D*YB?2MGJ4E,:XS#25SG!YZ*Y\"RM[_],F5;GW<^7H_KX[L3\7V/ MVTKX^#XND`?D!+D>RX\#Y`CY/UZYD#Q.T)G3D+\J.3A+LMU*J&INC$21$Y5" MU$5^1%TYJ90]/A8A4&3?[JT?"G%Y(^S!G10^>HC3`=6.)!L.VL-U)MH7X:F= M<]L^$Y;KDW/??%!W*[Z-.7V>`+>D]-3.GFD M$]I2X,U>$<0+S3='DMSG=6LX>A$;Y\!Y/"#(P12,F^`'&F"0@H-\JR))WI!# M+"]X"I6#^^0_$KB3,1KF*9897N)DC!\UC/-1)^;;[6(.F1_F83G3M&A32X+Y M,M/502%Z-[-T@2?D1KE"'IJ7H@,Y:!XRF>;^^&A^FD_0![EHGBK*Y?J0>L2Z M7`@GN7(F::9'#1QQM7NFBKUW4,3O2$$9SV6.[3@JCE58KI@SYJGYOJV8?^95 M>41.>A;GRGED3IE'XH.Y+$3H5#N9SQ(.,_+>RU#DXW,#0NBW(-[5V5&)&P%Z M4V'GT+@]+LJ=YZOY>E[XM>?R#Z!$>:/?C!1S$N0]+,`YN"><<3*B4RCUOK)5.F.*RX^N0?H9S()>[6RBJ7,G!Q M)$E#'/4V&Y>7A'LK'.EY:>-1X_/@?6F_?PN<6J<"+X66N*"M*UU:B^-^XT_% MZ`KIU<1=N;QDSG;;H<.-$/H_D]5Q4;N,Y>VNW.5ZR&,KV/$^?7CM&=D)2)(= M-Y@&ZN!<><\MI5+HC!WDGIX.J!<<\QQQ1%:^@*Z./!4G,5%$UN;&KD).-AC MJ9=(F'HDSAMQZI4Z*`]G.TMR#C]/K:S^V M?`C((Z*C4_P:VD($M=R'=N0T%BEP#1(;5TYM3,*ZK,M^'^L.DC.4J<-"!XH3 M/"(M+J0#@8"I]\=?67H:9ALWIYMK$1_*6TOTZX$+`U+QM*XSBN9 M2K>2N:ZN_T.9H:UNA\A&FS>=0W?%>UC1!V2OCT*43^R4?I//5"`L%(O7-_]Z M4-@+\>L".WRSKQ_L^#J__B?=ZPO.*:3V_GOE,\AAAP!VN)5U?CC%ZAKQK&ZR MA-X!^NAC_/Q9]%3+'>697^P0"]2=F.P[.BP)ZUAPZ]&.[HHOW\JZZ$/'=.?F MX-9&/\G=K$SZ9VPGWPI'TG*MDRY6+NE#\!4L#X_T ME&@^4/L"!Z]-[^!AHS6X^5)0OWQQVB8/\>%OCSMHS="QE>PD+-'2$3=HP MGA+0H=RJ.)B>NX]"4+<3="4UZ:50[\Z[`^^]$/&.R!COG?H\YP0E[^.Y3]V\ M!A7OT_N3*;P'[\B[\[Z\:^\G]UTE^<4=F)"S,Z_ZE\W@\DH&R4V' M^]WAGHS,K$_JGD+Y.8GZE!8"&2`UL>#A,CGL?%X*20J1Z4"3PZZ[,T\.>V7F MO^_O_WLD1W<5\/P[`'_`CT)^.RJ4P!OPC,__C@GA=Q/\@A4(CTJ#OJB9R4HX>' MZMS:I`ZJ:^K2.W5^J8?JF+I]U:F[W)0ZV,/?[4"E.M`DWBKQ]PFU1)=GVT6\ MRM.UT4P[T]]-\?CI@CH;7S(5;W=Z^-9XR+=`/`?/8>^8C]'I$?'5PFQ3P&&` MP$-MSR18;ZLNS?;YK&TO\M>VM(UM+_(!7%8GR:<[OLW/7N%;MD]M=K(<2C!=TG/>OWZ34U=E.MA M4XD?/WR/`4)%Z9+6S2&![,N^U=^U6^S)_#CW6 M[`_L\\Q+\[.5P@9,@[P[9H#3Y,Q:00#Z;FMEWHN1U`3?B/,1/`%OSCOP"'PZ MK^XL1^H\.G_.#_#N?#P/SU\YY;P\O\#;\_5\J:<\S2JMUNCCA4SLGVW%/IQ? M[/N&!J_'>X1Y)3E?;(=6Z=#KO`Y9\;,2>KZ06.%G$-(UXF"#(H4)MME0#U1 M?=MX%?2]X%U5[)$3!_:QR3U)0YSYI-SZ]J6T&+/<=`EV@X>?08P5:!T',Z^8 MB-_Q$,%J2+#HV-K,K,_054^"+!UCS!AUBY]E=M10`M9_]3!3KF5SLKV^$]7& M!-$[^MOK[.J*SH46:E)W7!Q!`$5'P7E&`;-3M*8KVA]Z[4.-WLD8ZK$^>2T= MU7$E;GD7ZXV<)#5I(^`W=Z8RD(T]9A(TE'+YH$1?]S4]Q:4Y3ASTW>65P$XI MT[%G\+0D3;]F9MX?83J(T*]+T>!#S\5']/>2#CA0G6<4O>R&T?.@&WV#U-&# M41_]RZ3''4$FO7S3VT_B!GAP7Q_F7RW]-FAP+X3(/>,=`D/SFU8P1O:PWR>< M&G^\D:-HTE1.`\G`K`L%,=,OP6DAN0$1$`9DP+GQ&YP!C<%MD&Y@"59!N^4M M3`09@0IP5Z@((((+$F&,!NR&3A!=E!MDP'D?;4`$9<%9D!%,&`J`1<%KW"HM M`70L/<0,*$`0D`+('[4!"G!NP`$I@!R``E`;2$$*<`:@`)]%"H`&H`"&P4WA M!LP7*,`1$"9`^&D`"C`%8`JLP9.`%J0`(//-0\+F`2@`9`#BAP$:/H?_;)0! MK$$:D`*\`2B`'9`"N`$H0!F0`I`!-OYA\..C`'&!45`1K`@U/A.0!H@!Y($< M(%\T)EZ(/\8W&`%`OA)0!GP42,6.C^33^&,`"B`$X!9C@)>_!K``(+X4@!"D M`'%`D.]6U/A#`,N@7:#Y%SX4X.6'`6+`;V#A*P`H0!'`(105Y#RHP!N MP'J?[O]T/Z2_\'(+#7^/?_FV^;#'J=_HD/[XO(OS]JS^OX_JO_"V_ M&Z#DXPA>_H>/X7C].;_O\/"K`9"_UD`'A/[!OQSP6;#\Y+^^7[X/_4?>F_!! M^-I\703#GWF@QH<\2/M]^#I]7@&X@$>@QJ?\Z_3!_MI\I(0YP/V/ZC?B2_]U M85Q_];^='[TO_U<1&/E-!-(#%$"YP-!O6M`1,!*@`&E]$4`LP<.OQ&?QLP!. M!MA\*C^TP'Q@#K#U6P9Y_9(`'[X0C=`".K\5WX8ONV`4\/*- M_S!\ZSZ.058`!=CIN_R5^*@+7H$PP&%`Q\?CHPN,`"U^,;^TPP_-U2@`3K^E'S2PZ4?J>_I)"`8%VCZJG]5/Z@?IR_I)_^9_NK_]GP_P]7<#Y/%Q M!+Q_%SZGPL.OT[<)K/%=_LP`*K]4GTGP[_?R$P@VA8X%,#[W7V*`E!`&M/@1 M^1"!Y(%%H"/PPB?TN_`M`?=]9H0PH-"/UM<2%!#D_X1\M#YO()(/'*@`[`;J M`MM\KH&A7QF`"Q!QL`&,`>(%$8)Z8/)/!6@!C/U=!+H%QP&#W_Z/%)@"8`/P M^3Z"#3X%GT"0\``M@/&1`\E^,#^V@-?/XD?K(P+6^-I^;,")@-//[N?TXP7R M^60"][YM(,IOP]?A:_]=`E-]G;[$7SGPPB?C\_6U#;(%8<`,'Y'O\K?AR_P) M^10`&[X:H$40*5`3U#>L,S`<,+[/W][O]4<50`IL!4I_/3_4GV:`Y?<3O`%B M!:&!F<#;'UNPT[?[0_,I`"Y\OC]('RI0^%?C*_Z=`8Y_H$`0'_-O*X`8>`)& M_Q9\08'LWQ"`"9`7#/#E`R$$W(*L0(=@Q4<4E`:V!&P8.$'6X)N/!C@&+`1: M_TI\10`\P*"@[5?ZZ_15`<8*D#X@X#RPBQ`?X`G^!E$`6P`:0!=@Z)?AV_#Y M!2]\GL&>X!B0UB?Z3XAWX8/"O`$\/)=`8H`4H`"(>D/TC<%D`VJ_FA]<,"YGPU@Z#<%@`+$ M`.B`WS_B M_/A^_[Y_(%4P&@@DY/`-`9"$G;Z"8":P`N@%O/!1_-R!8\`ZGVTP'Z@#A/F9 M'%8;FXZZ8+D/_V<,1/:L:IZ$U<`SH`30.PB(T&RX^N"`4(`C`)*0R*7_(`\!?DXPHB';Z"SS\HX$$0#H`$/`OR M!R\"(S_!GX=O[;<61`&4`A&`O[]WX'7PWS<5?`RZ`=!\P,`3A:O/YY<0)`H6 M`8E^DP&'X-%`YW<&,`=*_)Y^,P307XG/QW?N0PID"B!]&#_!W]`/0L#O(PE> M!>F$&L!=H"ZPT^<:%!""^#B#54"]GQP@]#0SQOT'<%[/J9"4%\C<(-G];`RW?RXQ0J M#!V&V,)?80P0PF?#\"C5_F!\F\(GH+R04-%%J!1Z_>"`CP$R`)*P+ZC_&_JE M#.>!*T.-Q(H/[6?]@P!V``N`LL(+8,CP8+@\*`%:`,$*1H&.``HP1+@+[!E> M![N$YS]N8&B0"NCF4P`V"1D+A@Y77Z+0#=@H-`)F"D^"U4*5(*2O)6CYL_X5 M`(\"2X6T8%9`21@:+`\^_9"`.+X*01Y@AI`<[/2Q`-$$Z`%B81A0:&@NM`-R M`"6`#,)0!J#Y*_&]"):`-L!K8=^P$SCB^Q6B`L5_0S\R0!Z@[=<&4/*9 M$>0+5D(58,D0]M`J(`CB^)X$]X&.(8HPS*<;=!5*`9&&/0*Z812P"<@1@`N< M!\2"FD-"!7+PPQI<@G`U*`9=^6T'E(680_S6`Q,"]V&W4*>(:>P>:@!/!C2^&:";0`^W^8`/1`C#`-: M!-:!E[]<8#-3X=XO\/,Y@KE)L,`S6!/D09'_[/,R!*^!!\"U"$F#$G@P+?AZQ/.+!@+-0"KS)7OPJ?@(Q"6 M"^N$&SZU7YKP-=!%6/JE`.Z##T1?(0508'@T6`-L_SY\&[[\8*00Y@=P8"Q< M".A^Y;XNH#$P2Q!5(`L$$]0"C(4`7^[`)KA*C"9^_:`%A\-;8K80B=@!U!+& M"RJT#.&$3<-/'[E0=*@#-`CV M#G-_1X/[P'%0>,CGHPCD#QF)#\-L87. M"RF#OS_FGUG0K:!"A/R="J6'?#X4`6(`SC<$@/1U#=^`!\'\8=`O%C@:C/HQ M#KV#PPI77]5/7(CF"RF0!`6)-``\P!CQ.A@AJ`AD!#$+R4$WX:(O*H`F'`LB M_Z@"U<(T`)2@OR!*2`$T"Y^`$#]^(:J0R"?^:QDR!GN*%[Z,(5"15RCJ2QB: M#@N&PT)IH!<"#4$)Q"2>_B($=,`+8H/O0:@7G/")"\>&7KZ+`*0@(R@N;!BJ M_2Y\)`5TX6'Q/K@*;`5R"N&`J$++WTS`+#C>RP,@"2F#;8!28!Z@ M$!(#M\`*7^-/;Z`D#/.E^C**$<7R@#$1GXA0'!HR"'<5#3Y)W^3P[]@S_`1R M"U6`&L,#(P*0B9!@1`!2#V^+*CX&(?Y/P1?FNQ:`#H-\>(#^(=VP9(@J"`_" M$,>$3+[.X,&/*F`O!0%#CVP(,"7E\&<$;P'[03:@]:!U& M_>2+F,,180&P":`##":T`.Q\/+Z`(MWPPB M^!X%5`'D7X#$D@@T`5F8SR(O`2B`AMQ8$A=B!&R M_)B+*(!I@5E@UVA$Q"7R^5"+]T,?8HV/J``W9/=I%J^'_,0ZHT6@5#C>LQ=( M_Q@0&<0G@7>P)Q`>_#"*&BM\8#^O7^5/UA@'?#3&%6F((D#*G^@PH_`K3/7= M!V%_)L'D@?!P1!@Z'#*V#H>"5$6.X=]/)[A!K/')^%*"2\0BXXBP9!@8\`M$ M":F#S<71X?IP`MCI&SA6`K^+5<$2'Y.1`_@))`IB$D>!?+Y7H9OO%!C\0Q!B M&5V!(+Y-(!Y0?^@'+`_@F)#'4 M%!.'L<(!X8@04QAO?/CA$%>.OT(9'W*1##@=O!JJ,\*#U3\A7YAO(C`E@",` M&$6#140,XY2Q(OA8C!Y6_=Q_^8&O(ALQ;F!&D!QJ`AF$UYB)81`PB^AWK"YR M!,H`7T99(;BPX$@B1#@:!.N.=,90(1U1W\"JZ1!^^/!_EH$_/]9%;=_O[_>X\;1XX@+)#XR M^0J`-`"SHZ6/&-A?/`:"!)\`/CX@7Q?A\+@$A#TT^+:$,T:#HTG`\9A:;%?4 M`,9A5[XN8-H1YC<^+/*M"FE](KY!X8;/Q6A9A/3)&&F,=<#Q8Q[P5_C_,Q;B M`FN!P,4T8ZRQB(@8;"%&#;^$'DC'8&EP>0*TB78,VP+#AEW$&: M94R.&+Z+`,NOWB<`O$$*^?!_H<:J88^`J*C\DQ-V'2%](4@6(Y\/4:@H/$3: M&*^0;LA%8C[P!,A#O"0B_^Z&+L!6)-A1K]B&:/7!^""!%T9QX@YPZ["$4B?(BHRW!@&,":2%RN(:<6`8:=O:ZB"M/^Y!"6*[P6P8B?QXD=@Q#ZF M#U>/Y<9L82#I,'+8`'HXI/Y+>-A+W< M$95_\!:OWRC053B/K#(&^0*2]TA8([J1R"=>[/&A_V2/"D;B8`D1JHA"="!: M_["$W0*Y0+*10@A7/$72(6V(MD!,CYL(,7QZ1>1K/&A M&_N"NL`-GQQ1$BDP!#_&!*&&'4.K(*TPA@@A0#[^(K%_C,=F9,(Q7X@BH$6R M7(:(Y\`[I!M01:`6M!?<%$%\^T.-X[[/K4!N'`ZV"H^"0;Y'7P$P!N!;;`X& M%TM_$;^;X_B@BU"6C$%"""I\$;]L8S\Q" M&4F/6X%!7TR0?R@V4`B>%LN))\9LY+DP%0GB*RI,'\>.&SY\H;-P7YB;/!9^ M)1.+;4G^X5L28$A<9`O*%H&([Q;6B>N/QH?_*PK($,0`VT&1WUB`&-DPL"!^ M(0>+8T#!I$$Q(:B3U$52$3<$D()AXL1/>Y#G$QV:^;R%:82M8_(/2G!`]!W* M^]!\LL7]G\*/C0@8?!)4^/R#ND!:'P/2(RGDFQW.+":,[,0U(5'Q_S=Q_!@> M)WE\;X#DY%G2)OD5X$5."E$`&D0A'VQQ2J`D3"=B"V"-4+=<.%:3#4$UXK!P#H`D##CB$1V`;L#Q9%TQ*I@B:`/8 M'N&.B[[\XU=`VZ=`K#2V%:6'!D`3Y2;2!!E4[!5*$>.*V8*)('802;@'S"-" M)V\>%\CPH([2P.B`I/_Y'2F'`<.,(@FQ-,F+!`D&`1Y^74HOX"XQ>Z!EF"JN M#;^.8$FX0-A`.-AS!`[V%^``9@`C@4LP0)GKJT\&);N#,T+<(R*RALB;]$QJ M'RD"FH$*)&-!)0")-"T2^="-`<"QHCLQ_"<7P$8^$+^/>[^?9!W`1]EP=!L. M#..+AD9\H&F1G`A)C%2R&\V"E4HFHB+1VMA,#(VX_F0`*0!JXDS`FNC?(KU\=P"BXI;P4UFA7/")(2F,04J4HG50E_B:9`T` M$D>$.LDG@6BQ6@BD;!@6!"U\\$/6)(_O"9!:B.@<`:0`20`B0*R2YM`%W/IU M^C:0]D#X($,1*@EQU!2F`V>2,4-TH:<2_"@>M!?"%Z&'<,"7X,J1`;CH$RK& M^-B(P\&)(^(P*[`BR!88(:&,BD3K'ZU/O(`&4%0F%6^1\TI^'SH1!8`%8/QY M$Y5_R0&/(@H@"[#@>SYL#TJ*80`\`#YR'?B+/%8.+*.!,,4XXHA2$DE[!/ZQ M!2]\#,B)XTM0Z:=!YP)#&#[D#')*VO?-@&:`'P M$4V(G$.#91VPQ)=$?$?J)(&'F$#O(':Q0[C_>T[.`W6!W<&OXYLP-TE2B"R. M%]U\Z`%2(3;R)GD=E%26"E.5Q4*T'Y_2,/@TW%0N`,F#,[Y=X-820BFA_%JB M)4N3FDN02U\1UB@?`%TZ_N*1",/NX,0/#-@V'/J5+@N/?S\Y0.H27@!UA%2V M+D^5E,JT@;;0'*BO7%IV*M63-L7I(N^R:^F69`]Z!W4%D,CXHO1OM4']6TA^ M"]!\,LDMY8I29VECQ/&)#SJ&[L/=Y80/Q/C/&I(NN.>CX,8'K@JR@[9$"H!."%U$)J)("QB+BT+$=J M`AF8=47*X`/S]Q?!;"-J!I!_'\>$):?P03GAZUTN)RF4R8.K89G1U3<%0"%P M#*@"(\!&H2QP^N@UU`4J`"R3U\?,Y"H244EN1#>")MF+Z$)`8\=!Y0`S6UFWQ%"F M%F8W>4MQ)7FOI#BFW"]>'UN"30&-9<52)U!2U/9-+#F6J(':P,52!8@-I#ET M^K0`$+^+92`3&,AO=/4Q++^7G3X<@-,OT&AS1/XA'9Q^\LLXHFTP!Y@M7$L25:F]?#?"!#^3-\I1'S"PY.CJTV4. M_7B9[\@+'](!TH<86$QB"-U\ST5R1@:QXRB8[$>>,4N'D+Z*Y)_RG*F1'&). M,#."-47Y@I%1A0GI"R-"#YX*)$EAY$ERF[ET;"/B`E62(LHIXQP1'PB3_/`Y M!?5^;0#D(R*SB\A=9&0V_Z:*C47\HVURVU>2O%9N^#*3.<6MHMYPOACD\R'> M$[.9Z,)QH4`0Y-"-%#IV",:&Y`'1'UQ`+C`SE/U%#RJ"#4VE((BO>WBV;%^N M%#N0U\HUII$`^8?0O`=N+Z^92(4*@:A0?Y`P_`U@'/.9A,*Y`R1RD%DMM&)F M+G&7<,A[90R34"A\#!]J&3X*=0".8IK2G7F;=$H1`$*,)N8O\4GIG#Q("@&0"WB M)XE^+TW*("[PKSGT._JM^+:1(SC.Y0I0L@GB@R-Z#NV7&D"\8"%PKYFVC!N: M`6"'?+Y%'TZ`+6G5?&'V_+0'$\FY(D(PA/FD)$W"%\H##<.IOFR'2DS3`%L_4I\\D'CHDM18&B\K"*2`4!_O\W\X7W0M`G`G")> M^*2/U<*SP(M@]&>%)#_:$.]]\L5LY12@"B`$R&2F_#B9BDJ-Q-[!U=>N!!N" M!8R5JLRYI8F/OFG?##RN$S&7V,&-IBM3]@?L8_!!8TR.>T*3(K!O'*;?5$-Z M$UV84,QZY:=Q7RG:1%X&$TF3HTM9X?.RKRF]K`2.&:>!?`%X7^80W)?TT6S0 M_"*-"7(. M`T.*;\`)$;928RGR2YI$-UR<\#X^(;@/5X`$2G%F(%$`O,@LX@82MNA4,/VU M&W.&-("=89!1NLGS\W%R((&<4,XKGY33+/-VH7%>.2N`$TN1GXH"Q4GSHSG@ M_V*!FP/WGR,PDAGGA/D5"@M\J4RG@1'@SNE%A#\R!/A\_$V^ M#56$B$N7I>_PK4F+;`I!)7.7=`'`*OF2T<52H(2OB\!SW'5" M)->``\GKGW6P4;@A9&-=.3>01+[(Y&4G`C&=*&7>.#LD>)TT2V$E42!CR M,)&%K M%5F=DDD-'R73R\>M;`(,`>B;10`[8$FQLIFH5$2F-^&9G3XX8*%3&?G![`/6 M%?>3_T/3'Y#`@8B\C%,N"J^7:('7I?92(-CZ<_4]$&&+<4==Y&32J:G'+"CB M+.N*0;]`ICG1-@D;/M\+.F6],[4PKVS?,+I\R:. M*0U^QDB4'[QSV2D"=!=6K:Y\,$^[91/`6PFN%%=F"4J9Y\X`9]#3DMD$P&06 M/9><7<`C),UAZ5GOI`'@+9^>1\\N8$8P.E`&X#[F$S%_]P2;9:GT6QNF+#=\)(5%HM*/OF`U\"4> M+66!OT?,Z.(!Z`E.#U-`^6`&F!'T\#8]F3A#A"_!4B!=B>I$"'I8K"U0?&9%BZ M!/67;X#4Y+#B+JG[NSPF_W@)!$M$)X%R^DCD&U-F,'N!KD=)YFI0,VA@)`4V M),%_?T@'/DOBI]9Q8/A3%&*" M*GM^X\][HR\R6=F&&'8Z'IU^GLSXIH"SDEGO='U^/A^-6<_F8VG3C,E&1#?" M/Z5]5,7[@.!2CQDPA$-N^&"5!TK&P@P`!7?E"_/I^&BB/T$0C0)A/SEF&,"R$0`.%],NA)5JQ:`GAXTUF M#G.&],)Q82[Q1^D&0!$6`/5^8D6DX@("2CBQU/'!!DJ*0`5'9\_/^`?F/`J@ M%66/68'R9LO1-=F&M$@F$/.%F0+38RBSTZ?C8WUV`2V*G3YC8-EPH\@)K'3. M_Y"<7,&50$=1FBA8;%_J$VV4UDMI9AL30@"BW&M>/L\NPDL8G_N@8!CMW``6 M#3N7^TL,H-31@]DBV!2V`!X%+8#](41Q7%CB6TB^"@N*5$8X(K?P'&D32$<" M'EN5LTX8WYU15KG?*P'T][@"7`(N`:[R+Z!O,/"5^VJ==T9O"N`A4`L> M^>0+@DEBY[L%&*CIU$#2*]F.D4YR8S!2^_G[HTPB^6JA!$LQYX@QXSA>Q(3J M#8F""ZF%-L[2GVJ15T`)I$]N%RV&RK^T9,U38_GA2X!F M0]V`H$JV(H#1$DHU/'FZ-HE^2,S:8T^R6BB3(AJKC'VZ(I><(-&(Z.>3(VH1+5=J M+E^),#XG0`93,%FM]$1O<.D*!HS:*CQI$N"`'X>\V\Y_8OM$BU;!C2+HT,-@'-HQ)S$4EP=!;0->>A<41IY#R0]P@]?!&P M)(MT%L^])APQDOGT4P16"R^$A]"$)W`2>*DE7(=*&3&A]$@?G$$21H>5A,S3*-S0-'F/C!A:!PM]*`+X96$1:^D7#!V* M%I$"9TBYGX^(B$C(G`]H%WU]1H'O)6@2Q*B'W'G>/]^1HD"39R:TYXCSNT+2 M^@X"0E$U9HP/]5@/5"`._8P&4()X@1P@8MC@>P]V$6A_Z,[1PE`S#*@7W68& M0F&!V8.=(HGQB)D?=2LN&<,'N$"HYLC3/+@:_(J"^$*1/5`3W_BSO9F(Y!;R M,`N*(DLGP0FS#IK@H_T5%B&7R,&5II7O>`E;S`\6&J>'`$(\7_`/+=G[G/%I M`$6':8`))@-0W3CVS'UR`G63%$H)8);3.IC8;/"Y1A>'*<2A'P-Q1AH&1!XD M!1F"=4JJ7V(@`7C"I$4>GUB+Q<4X)N/R*MDCS1^^-K^,/[^K@?^R`T@SO'W6 M2`F8B`&](\DS`LKWW`FN!L&8^T5L)(V/UF"2K_&`$TU4(6H3W@+CLVC&)M>!'\=YI]"3Z*FW#'UV`1A(I_*"+TPUE$Q`/$#/N<8DF$(1RRQ)<'B!FZ`)"$AJR7H4D121C8-#G0 M,,F*[#Y?)3UQ'$@1J$W*'$&8<,;S)ZL3H4C,_/V1#6V,-CI$,)H-,/#C@#,$)J`(U_V\XN:0/377K2_!@N M21&`L$$%(C1!L4AN?"`*!&,N=\1>)5`S6YJ;'!BB2ZNDG5&QP<`4!4`#0!)^ M'=>E5,B6(:VO=-D"8)C^0VV28$`**?70,JHDE9%"3)6"1$4DJ!"TX- M+V6F`L$%Q/H/7#JG%)<*3"F-^$`X8`V@QMB&I#GL!3L$+0"H0`O@]@?DFVD* M0+N`"`Z+`[>-,>>LT_D917Q+/#[\TKJ+XF#'DJPHJ,3LKA*6$B> M#E^FH<'U(Z3/O_B/_&\*MQI\5P`N(@<1W:E_Y`9:_I"?A$+8BX-QN&E@Q/]Y M2I.B14&U8M\16%G>;';*3$M\-0#'I5FP,FA3?#(B_W(`\M/P'_U45)CZBQ!P M#(Z?/TR&9/"OO,DHC'@&_Q2#_07FHZ1P_\E!1$J>3]&>A<()X^X1=YI9C/&9 M$;X%XX$E7U#R$"BM[(NR$2$$E-(AJ=``O8H!3I`AR6'*F$^4+%H.2HMF1Y0(OA(XF M^E:@>,]N@S)A[?`6*:5`&KZ1?0/(":M)02 M28.`2$E]I3]T>P1'B__/O! M&IF:+,^F9IK\CPZ+S>+D"[?25 M,&<$JU+R@)*PE&`$\/JY`:!\1SZF9@(TRSE8A$YB#8%M<=2\(-H2Q\<'[6?: M3C>?FTPTZF4`!W!)?9FV41&C4$I2@B52.VG3[!:L`RV*F]%`8_+4R&`.#*0Z M3PFI-3XX()704OF1))DB%9$.\,*B@%G@A9F6I&!Z."6.R]3I)MNO>?I$):0: M3.^'?S_Y87%37PD9(%AN-B.8*K\*7^;PILD4Y!C\+A&:ODG,H.=R[`@@%!U6 M^&:'`[)NJ"3U97E&9?MA4H%\3$TS*",UR+BT;%K2^)Z64 M&5F*<-`^!=0SP,B!;/`A#P)]P$KI'[IS;LD>I)3"+SV6#LS>)%BS:(I(32?* M%EJ6F5'!X>.)Z3C-S%[N&G.*I]23I/N4-#F[$8YN^*R>94WDGQ4T26AE9`EV M0B-__$=CH<(S\BD$U03>^92$7D@T!(9CJH,:E&^F5/.`"]4TZN-OC>HWA`.Z M`:2*?T&@J@+4"'@K)6?B_O0'&<&EZE,A>@`L!0Q@*24?`*[)7256^62#_$XKXOKSKQPP.B M`9Q^)`6WPJA/J`I+C/NM(2D";``50*!1K7I5C1[@)_E\6E7]`1RP9%H'G$LX M$_F*,+ZM*"F45HD*=1.P+W,'A-5)X(&O"[A6=2"X#%$`/8`:7Z"1L.K5@/?U M`]N0WL3_GPP@6/`56`ZLN M5SM]J=4+W[$/&+A:-*XB5RT#;\P'XC5S8,@ZE8_*/>FHRDQXH+30`NI3W:O2 M`;2K`J'-9:#0,:`PFK#4;7:AM1?NA-3&I&!0B1X<;&:K<@J_I$W:K: M4@N`Y=+[(CES#L!630Q,_=JK`].E5EQUZ/=;)9:R!="7KKXA`'+5!8!>[8,B M1B^'*%0R:B>R?BE+#4I^-/6HL-0@W[`0OWH5_6/65_6AJLC4ZE@@;QH'/?;Q M38V5&(Y;A853P+E?E'>6%`VLB%#ZHH(U#,!@94PZ6!6&H\@`HX3UO$EA;:LR M`/V6$\*`IX;U###UZ[#6`;&8#DCSZG&UQ'K4Y&?V_/RJH=:>"P4QL*J'O+/2''BH/E9OXG\5R"FSX+,& M$`NL@-8RH:"5T)J8!!8>6J.LE-5%:Y6UT8KWA+-"6N&J7-:A'Z6U#AA:[44( M6"<"@<$/GU`U[8`:@$I.`8ZJ-40_JY24#L!*;;`2%6^K4S_=*A>1MPI5V0KDW("Z_O:GC@&?Z5,5L'I6#?VE5<.-P]9;X=K/NYILO?#] M5IU^PE5@YK#0Z9=BS;):&D^M@%5/Z[0U?P:5/`$F5A>K"=9SJV.5?ZEN[18$ M";NJ&`Z-1%<"QL=H'9NB5E6K0-;-U8LO6UEDU:?*5FFKGT`B'[@UJLH6I'Q2 M57NK*(!_J[(U9F@RM0G*$CE\(U;E:K&0UM=<10$\5R.1T57]H5BQ#'E=1;+B M5],.QKY#X99UZ!=>];D2"N\:F-;S:MO@J!G';*]^6EFL==0WP!]2OLKRHZ\6 M7&FLM4A(I-6UOUKBF[5F7!,/`B&@YX`5-HAK56H&6I>L359?ZTUS"!ELG;!6 M6/6M&SX,JSF0WYH1]+*&3TF)4=*T8:2HMEU)V!8);G"5H^LN5:@XJYUXMIKA?(= M6H&*-H`I:_0QWZH$Q+*Z5;>L75::*[MUJ,I_Z$9V7"%]J=4+Z%9TVMJAN2-& M5LFI0,6E:1UPP]=T';W:4AF*KH+`*XEUZGKANS..4VVIE\.L`(FA_=DO/;B6 M1RV76-?X*N.UZRJ?9+6B6GFOK=4ZZZFU\HIG9:U6?VRMQ\RVZX&U3/A[M:4& M/$FFJE>V7\55N@I4/!(J6E>`K%='*\+5!`E[G;DF#SRL+C[,*LY5ZEII=+[6 M4NFOTE?UXKMQY^E]O;K"5_^4_$)2JY;4KTICO8'F!#VM%[[P:D)U;RIK#;FN M!?8/Z%>TI/HUR2I_=;+B48&*=5>SJ@GVIDG2O`%(6*^9'==PJ[&5_QII3;9. M6F6O>=:U0"`T6PEM?1($`6DEKLI`H:1QL\K?NR9>1:^BJU!=Y8&/QW=G1`;J M_AJ9S<$C7V&3$>C&[&D23@VG:,]=!1.TI/@@[$#^!]F!@E&"*Q.0ZJ?,O"1J M!0NGC4QO(:WOH]#P7),J.SV:!DTZI941M-C\B\):!F&-H"5"A0B.V^/KT>D*NOS$(C@1":. M0.&`<("J*JTOYSE<&+="^NRF_86I(@Z6_,<`=5[V87N8C<1QXKD1X/DG#?T5 M8QN4WL$+`061CU@\_"-R8@^7DS^GJ8@2R(?#5$`J!L:A6%/S9,"4??@$/1P> M^>*5KU39(Q)3@,DJI%<^4HNCPDE]H=BSGKHD54"6_D*9.$XL@:M4Y-<%-)O2 M^G24NR@^"[F.V=A8X#96-QKCD_7A M#.6)0M!Z:KSR"7H&K:"2/:>I2\!Y+(0O95D`[,'*`=2'%E.V0+#/XJFP7+L* M+'616S_\7Q'`1RKL#"=^(JN4VE00ZK*OGDJ.5$!R/DN*AD"8+&-4!'F\-`J" M,".R(DR?(EI@4_AN`9JQIM+CI1E672HC:!M,',.R1\5UI!Z2,JHO]6!> M/N5^%TC^I@NU.1LV5JH>@6@CEHP@P8_>C44V1!?-S`]D%=!PF M1"VO1E-ZK#<19N@LF/JQ5?N'3C]H+!O`-/M_K0-Z)FVR"T!AHLWQ$GL&6*%2 M'4F3_4ML;$(QA#@HW$;.'2B(-T/P*$KV*8M&_00B9(N;'T/OIK)05TIP_!RZ M3H&N:=*I+);`Z9>'C9Z:-\FPOD=AYB=2)KOV.TWJ`+B:B=DK'QK5,'JS!$>#GZ,3J_GZ2QYD`JNH1<3U[/W5?DKJJZ2. M`V6D!EEX+%"QB*B:]1;*3'F@\\=]`_-3US@W9WW*M.$[;5AGB6ZF<8W:2F`/:L=E!+:8./&?F9'<9N M6PNSF\PN(!Q07YH1A!>46[F'K=EO88E6'YJ[1-&&9(.D]-B2;*^/13N?9=)* M`>^JE,LH;2%P2KOB(S%::>,`U=+3]ZI1XT3NO&K&8Z##.U9%`3:SVV/BOR\]'. M!?$A@-/CXZT3#`F%W3\N`^F&$$(/HAOP#;OY[/0E3C&.CM)4GT`0,""W](5R M"@NE7\J3:GW6?GD8)1#.!QI^HL-K85:PQ->)M0FT%N.9EMB*HB1RM@JJW(T. M__A\[]#EH/,/_=B1#3G^#KF2SLU6P6JR1Q"410$X^42`5$YAJ?UUF0D7X"B^ M7U.OP]=\H!\0Q_>>?&2>3H^&E$3<*'>6)KNJ-`OV,Z<$VM00)8ER+.A#9?OE M4%\%'TL>*JPR]/E#M1,&^+R)Z5!4;59@B%J&E03+!/#,F6'\>&$-+JZ[\4-WK-M!_. M4&.DBKH$#`$/U*G1V35H27#)*9WN@]=3?:DE4B+K@0W="/F^8TLO=Z@C0,_AM[2_" M`&:V&L6:;46P7&HN-46N;;>#!,K29H_2DZJ@"`]B%=^O8<;A*N`3^9=1.)NN M_2Z$#E.;Z=TT83JXG1Z^:3D$/DHL+81O5RM<;8E62D>"%,-[H5EQ/IDCC1%V M&%FE`<45:GLQ3AB^%*!:&6>'9XBPJF506%"!35\(6*^M13_WIAH@MKJN1;DN M4E>N>$^@*I^/W!I[O:O>8*.TK%76B>G6.1@<[03*9T6%<$`5``!V+="+[*XF M6O."XT//JQ=3#(FR[=`""9"Q"(80+8@O&1NE-5`F30Z%_<4:I:$0C?H5,(S* M))NN3]6AGPJ@F:DDK".*8^RW!=/*Z@(W6UDA;`6D7]:OWU;`>V\50TZON6`VD=A`-RWTD3OX`43@.<>VQMP&7]VVH>KU:[K* M4C-_Q4*2YF(RY^BH)?(]52NMTXANY,W6C+O-#?NE7N M<1NZ7C]F+"X4*9E[=;IN,V&XMEQ(+D8WA5MDO38>.+F29$*@(@&WCHL"T.@Z M8QVIZ=SR[7#5-PLWM=_N(?*YD]Q/X`.7G"FNU9J^<6L"RL6L*AX``*O.6$V6 M2U&YN=P"+2\7Q)>ZY6IF8VRZG=MM[A?WM;J_5>:F.Z.?(D(`+LS5I4OYA.D" M;J=^[%P[:>04J"A_J%$:#K.58=SY(^T4/KL6^"Q><:.MM$C]ZBLTP.@"B(42 M:^NPQMI.K2-0)DF=3$I^9O&V1SA>?2#,VVB-:VKG6S8[G_7"F" M!.N.P\&C0.LV"M@7S"_V3G.)4-A0854PB@C&;.T2"F\5\,)XH%AQQ?=\C"1* M"'RZKD&-080R@'C.?`B:3!^'_T]^)<>@85BYG*EB)#>[\EH[+/0/#]NA5%/V M1MN(4E!"K;[A0J#0/5[6#ZL%$L7YH/C`$5AX??KY"!N7EH0II$=V2EIG!&/* M7N&"U+_6)-ROP0;?>`T-(?Z>Z1']O?_2+L%&>(6\LAY(KU@.FCI%*R M*/N[<=))IVXR.ZJB`,Z*`P&S0U2HUCO7K`M.1.OJ%Y"*+M`6KDF1;)O*=17J M="VY(S_Q`4#W;#O0A>,&==>!!]TOZ[LECD7,9>@V>9VY-]J=;CHQAPJ]]6X0 MSM%%V M=/>Y%[X)[I@7TK<"V/XA%:4RS,]RZ2MWSXO,->FN<2\#B=HU@'$6J.BHA0"N MM& M=9N\2MUC92_7R4N0=%$*8,NE&5RI[LMS[0K11=DR<^&`8%Y'ICD7-8O._>I6 M((VZ_45X;I.7G!OFA?+V M6N4+=<2R+%F7R8L:F.KJ-3VX)<5O'V40SGM1I+>2`NN\EM4@+(SO!E`2K286 M86N5,K\D;"M4BJND[2\R=/>\V-([K9_W+YA#[?1-<->]&[Y";Z>OJ5OF_?1> M=!F]W=RJ+CAWVQKIG?0N::F]P][Y+%!17KB?Y?.M="NM#TNRKD@7MOCMPQ64 M>ON!IUZ=7ZI7K\A=A<_&&W"ZNMR;Z5*7UL?PQ>-"0`FF2%WV9:-7C1O.91^. MQD0FER":52W\KCK3>N6T?@G7+1F& M0LM]/+ZMJ!-V"C`655M60:N""I:66:SL"Q<1(69&\8[K`=S3LJ=U"L" M!?J-P801H/UQ%&LX1%9R$$&V?U_H*F/T$XC_''?Z/C^D3\7Y`+TVQA@KY:M& M(M>G^5`083>6OBI7S/`I'2&1[%`$(SIU0_I\K#06$:F2:T.VH"[6_B=1_:+R M]1_^!,$*`]/VH^Z/#BA_G/OF/2>35\H$Q=Y6\!G:W/W: M9G6XX+>V;LPLW`1IBBUM2+,2R M_*I6X,B&(1;@$6KVI/1^*1^>RC_\7Q:``,RLA`/2>-T`-<8P*4?U)JE6+&!: M__*_+@-NJ`_2C$`W3#G6"7>6.$KE)79V_PA_C-AJ+2>VW5B89;S@H1K0G4DN M;6V*&&""*(S/?F@EA/M>0,65]T'UY?D2N@@XG0)D`9P`3X"!0!+@T:B1"+(" M8@NLV-;8HZB4Z4DJ%5?"!*>X'EMWKWURJKKI9>G*7,^WP]4B\%^@?GN\G-[* M6Z&J'\4`JN3VY5IN];?>6WNWM%&H;_N7PQ<$8`(P`9(`3@#*_3@`J`!&X64/5$+!&<'4%^@:K@Y*WV)B"=/P^8\.>X,:!8<626U@`3F$2 M'/V+3498HW(4%&O-C&.^@/=_@.!$@4J`^*E6K012@HVC#<-+,%?T`BD)[+DH M%4=\/4!3#05!;GD(C%?:-:RZ<,!3L$PR%0PDO1Q6+CFU'H+A+"9R_DG(U`4C M:%^LC=>"IY>4GSD*EG#`"P?`(4MFY8IU`1P.9B-*@&FIK,L5+W3V.5@$1@)I M`=6NBELO!3(80UFC5'G2*TV+"%K*H$E6Z]JI[=)Z'^63X&"1I\)6U!0[UOB(P<+A,_!8U)Q[D$8Y'G^5`C? M2;N1#>$2WT/X7[BGW,5:.Y46X?1KCTPB[`#C"\LP.,$A804JQW`7/ M=\FT?4:4\%P@/>!\E")":%_"V\F8\(:PTP@X);-,`88`X$HH`!4@"?`$J(AZ M@KD36UC[[]#SZEDJ'8'.-0FS-\^AXE+S+3QTM"@/W7DR M"Q6(]E"Z0`M@`'P2CAZNA34#_=#.KZA0(/JC'5NF'?6M(\&U(715O3@[%$=, M&/65`>&Z(U*85.GS;>;B8(.BE%&C<"=3>"HF90[O;H&*[%S)[\X3+JR;O#<2 M2<.3Q\K#K%=C-5F8G9#*)U_"NEG'<,KRFHD'="WZ19V&$L6S@-@SCHFUC2\Z M8B'#"T.Z0*#@N#@93GB20=&FYH)CZH%T3-D0EOP^04M\BML=:]23+)BM[-WZ M:J.3T,VY+HZ0#@SDXU8V"#E\?6$B0!)@+]P7W@,+A-8"!+XK9Q)@A#E,?!S& M!9J6XMK58"\152A;A&367]&$*=P(KCXU74@:IJ>*/DM\6X%B)@D0%Q@#>&#W#!=\P[0W55U=2YHY&1!(J6=0$O9C6U,]D4[DH7F/N[%?^Z^@B-Y5\C;(AF/Z#OC>*R M?]V_\%_YKV$R3B%@)8!>,@G#XDH,\/XW4.C_9>4&@,.`R6&NL'38]_G&U`(P M@'^%R^%9I+IV@]D*?@G'@#7`0-)O9#93!>H%A`%:(FV/NF#N[-=Q*ARS')0Z M#:W!^\H8,/DA;CI,-#YJ,L.6)4,X1%%U#>P#!@(+@3W$5]$C\)$U"2PC7`+7 M.YV>>LLG<+]WWDK!;;E2@=T`>@#R;8WO"FP3AKGJ`7"P[>"0Q(01WMJ^C0(7 M6GW%@-5@L;WUKFH&AKD&7(=^Q^)/H+2Q)T#K9`.[@>'`N$=K((4XQ'<'+@+D M@6'%?>!0<`YU%"P(;O4F@VN,S&#C9R*87KD(3O[!@JNO+=5(<"S7%HRI#?)! M(?V@]^`O,;N8MC`#!@;+A+$QH.`:L>_0*4U%%"9KZ8%AQ]Y!?[BZFN+=2(K4UQ%'SI+!@+20_&%`3!0"A8#V`6CC]$ M3N&S$&-#L"KX@7`8A=#",G>?_-)()I]O8QSN)$:B1A&`^$!::3588)R8/4/` M"SO!?.#EA(!U3'P,GOB^BRNKMMAJL'G2.JPS6@@J=MIN M7^^'*=[0K5-8LP&5#!1G!`?%2-,:GZ&X'%S:))JB@^&9.\QJX4)SZX<$F@EC M(+.5\6"*V,4%X`LS,S3$R MA1/"O&"3+P6Q(2PIG@IW027">,R/AD^\"L2K3Q8#'",Y1M# M=`W!C>&^8>`X*.P/]@G*)]'&7T>D<&@6PL<4#!O/4$?!]Z8M(,M2SHGNM!Q' MA(>9SD6%(%880GLZMA=CA#V.7N&,8%@8XUA$=/%6"SG'[T@%P"CX`!$>9!VG MA6W#RL=$Y=1O*7@>=@W^9+N.P6"!$""B5:MVS1:+_#@9.>$:)>27&4R_M!_F M:K>T_6"LL2!S[QO'7!L',OM];F/=[8[3X&E@1'@ZA7D"=T3*<6'1;+H)_5UF M33>A%.'^+BSS8ES>C+Z&C^.9HF/R<6R49KKAJQ/J*W=8+ER77#K.=>W">.$D@%Z8+^P7;M:8Y-3`@F%Y>MK=*`>8,XD]%AY79=V[F6(8ZHK5,IR5'8)^+BF-=MX]&X13P!G!5>.* M;%&B)=A5YP9V-$Q&9AEW$$>@,P#4<&GRLYF'Y<=&=NNA@M(SP&RX!;``-A1# M!J*FV\V`:&X!4-MHY2[:/:&S+6%5[V7C&WPHGBJBC?FSET,',K-2>:PH!KHB M1JFZ4][TL+GP@O@=/JUV+TVU'5.7L/V8BOR)#"(6R%"&F<#V,$<6,4P7X!3J M:4W%DU^6+B-W31IRW`\;8PW"LL(L:-HT@;EWG!Y#"VG$`D[%;3XVA8L&?K=` MB=^<*L!09HG/?TPFWOKQ!(:JH5"38JS5WKO$K56B-12T3UPEK(18"C`NMA`# M(S/$&V(B,JQX3RPBWO]Y0'V?),(3\4U!P3BNE>&RB&-\+N*;+9H024PFIO,I MC)>R)44"8XR89:Q/O7DZ$G?$0[\><4KS;!O5529[;.E]&]@[XX8VC:S/3"B/ M0SV';F1N;A>P[GAB2`]8@*M]Y%5>`?-S8/CXPQ(/4N>.S4L/L#34,#LT%ADC M0IVV8F)\K#?1F7S`E?>:%#NBU&3S;_PA_M`F%JW.==^_\5\.)/VW3ASS;`)8 M/?&_9EYRLC>Q3PP`ED]>`1K)0S^S,:0/;7P2?64>BHF.BV.))-S8DKQ>O0"; M>86!)N%1,@64?JG<=.]V3X-\N,!,<7]X/`K:[12W@/65NF!1L>N/!NP6]@*^ M5#69A]S_)BV7!SS7;14'@:<`0^"#\>Q"#9R0S=X*8VW%J86?\M-35PQ4?/S= M`>:\4N`Q,-M/6DP&5@)651>^5=4+W[7XPN<_-IEJ@7''*$/VK;P56/G9Y!6' M;Z'%S4S&TI"ICO:'O65,%5]<2WX9_PQK@2' MC.V%3F%GJ\D82Y`_1AC'0#V'Z^*B<<-X$&SC!2Z#$.;#E[\HH@OY7KSS3"[3 M@GW&RU&0<<#8I&P6[@5#)9'&&=?6Q769H=PR'@9SEPO!WN7K:PW0S.H*+BY[ MDS@43C=L5WN!?\,G8$RR[\`.SC$?!2N/>;2ML#\T)8P4Q=1AE?(''+5>#>,,_W85RCC!IS MC:?&_6'9XCXX@5PX7B!OC;VQ[.5K\-=X@4E!UO5Q@XO&-&%7WU&Y'/S95"IO M^![)V%=)LCDX@DPVG2JK@^?&*&/$QWSY#A"OW%68AON].F$_LF:/\*Z94^AG]C5G!"N6"F'3!"22T%QLKB&G MA!/-*43W(:.YV=P=MBT/R`+#*`!F,@!Y;VQI'B!+C5.T!F3!\3-5&GPXIECXZZAK]M-6 M1E^\*(!?,Z0/X*QNWH!*4%6`16'1K>LX?GRTEV>5]UY4Z+@O])A-7O\>^(A\0.2\,(>XB`Q-YO\*A@7+>LMOI&%X-_E$!B5' MD9'.>T.?\!7Y@_@EOE>29+?(DH+MGZAP?[K7Q"CC<37`"N,E(,'O\OIZ>"@? ME$.?9V0CK;CU"1IX]@Q'E&L#(U`:@!S9/MS:W,=685W#PUNR'Z/OJ#P`[B.W M`/[(@-F*(&\X[/L;AE\B)=>&Y-51,8SO_Y=(KO&)F>_,%$NDLNGXJ=S]+12S MGL?-!>'@YL!Q8+A#5CKC_Y(`X4GV)<$1BBS&90LL(([#3D/&L+996CKAZQ"X MAQ^V-,M6=DC9+:2F?P@OHK2A;?)W60Z,%H4G+QU]A"K M(LBX7%A$Z)B2\(M;<`.@B"^[8<`5,9*R@SN$'-@>B-/(2\#,J];3R(I=KB?[ MG<6M>=GOI/UA$7#7G$K\`**I#X`;VU1*CRG17*]&3`L]X4!$U&;C6' MAN.@)&B(\H)WXJOVHV5F/@O&4N)Q8:RVVM=(32-[E.F9+T6];]'6*HP3_7I" M;.V5',YLY9A8`5TI?29?5EU__(:G>5V\%XY4.@^N%8N MB0>RE@&GKN4W,&QY#CQ;Q@/K M@6W+1HNE,2FZ8TEI?AIS_?S&Q+_,YEDR]NL[I,_2+MNU<$S3J,>8BIJ&CK'J M@J/+!..58W7918%B1@\$,DG&K(+[PIN":9A`@0NH-=4Z*??_08^A3;8F!C9A=J!?Z:9N@9.A6Z]UT7KL8[?E] M.>6);D^%Y8)3QV<.U4-"F)^ZA^:#,P]1=GD?C9)&([N3M-0J*I#WE\A@CILL MC6_"*V9KM(T2TZQVMD-KG/G!W>8:LS>ZI`R$AB\7^$:J)%=C,-79VMQBQD8/ MG[W&7N8:LV(86@E913P:0=?!^$"DL!EZ4*A;K"\3'N"%M&>M25LT*=R7>;H$C&:.FNV'B,\*:F*ZT5P1S#3/2VW(L^:Z,Y*7`QJA#4@J M9.V8K\ABXCJP.6F??`M>HX.T@LF6M(?4'NUW;DGG9;.2OKZW;W^V'PB(UJ>2 M%7N_7.&>]+7S)0B4;DK_2,F*.\C!!U2R]]O//$JS#']_#&J-)5/ZPBF4-C9+ MI?VBVT)X+7OTU(F5?@Q7_%R*R,=K(%)2,+VO)/*5FG5]Y\_IH>@2N%E;3>UB M#3>N.&D`H#=1U/G#;0BO;OVTR6=CHF#VAQBZ%35'E27(9-!?85RVR^`[8._: MI!>/F6=\HN!P`\J3C?#*I_?3K$TOY$@0]UP=_!.K3LG3.DEW)+-2E9P5A'?N MI@'1YM#7=)-YTDN0'E%[F+,%5\43,M#TW;*KF#`69IF&[%2/KT,Z""W@'!/W MDW^%B\/@J.T3%^B2)8X&;@'>$;;<.08DS=9"PF;0N,M>3I,U)1$7,TQDWR M/@'37-+Z,7E:DPPQ7#H?H5U]&HEQ94Q93J`3D6/H8G0(]-<;T;Y("EN[9:F M'M&*3$,(+7VZ[)?"=4>K_:#5),Q_\@4ZH#SU<^&Z9N63^NGY)']ZC3DX)M&" M$;&07F#HZ;E9Q3M]##V[_M@$,$>0:1E\6%J)?/6S4NS#6NV M09W5^VB67ZGWEWNJ%2A_,0&,5%Q39%16:/A=F`]W00_4H4)R M-HVTYNH9H)/X-AF#YG;2H&>2'65S)OA3K;F7SIHR#7>>7>.?H1`:'SL"E:@V M^`Z['$0O;#@RHSI#Y3YO82VIZ<_TL^U1`&TSC!-&60>V4U\D("`Z+ZN!);*B MJ3VB9-USZ:6Q(_BM+DH?+Z.(R])QM=\Y94VL)C[S/2M\E6"!,G08J*BM9D#H M"L*#E[\P'\<:&NBQ/J2"K)^`(=X3J[R:M4E)Q1&SC='58$\+=!51;QVSSA*K M-<&-6=/HZW'Y4_PEWC&N=MO0.VOQ\TK94^VJ9'[B"I30G-7_7G3#L%!UAD+? ME./$.F5G8DW83FSOQ'<>`X/*O-"AA&,`J8T%Q40I!M"J[";#*&^`IKZ28J\RS%`%_E<62>FKV,@J84[P"EEDR MKMG+^=]%-(D9PE<#/M'^JEE^:6K?]`LSPXORU"NJBKO%=N57L5Y95MQ7OE8: ME$G7GT\\K4J7%#WU6^D"0&'1&`Y9-+/XI"FU1B$(BZ^98P":*PI`?JOE!1EP MBY'1;6!E=+@XP3+K-37ZBQN3KJPB)7&X&VENM%@0C&D;55YS M.%/*DFO&K2=X[G#D[%.#I>?!RN`']CW8)SQ6/DMSFEN&:VFN\6@:OBP)M.?& MI:G-WTBZ=-]8\!A)Y5N'*178\\-X["=RVZQZ1,EF`A74*\R>88-Z!*HQ^)%F M,"O2>U94,F)4I,R;W=1RFX7"N,OX)$?9+RT6;$%^2X_4@^GF,!SP6GKK+!KO M*B`,95JC+8TR83MGIH"&1AN&TDTC*QFTF1S7I5+7A8L`=V&M,Q'9Q/Q^'H#V ME-/7F.ES87^W/'QB/#K7GG?8:^?(\&W6:\PFY"(3K=6/7^3IX5SPCZG`SEX? M.LV7C=\S-G46B>F9'$W2GD^]V5GPZ`N3(:A\7/SE72,$XMH90LK3.MAYOD]G M5(?#*T$%]EPP1.Q-1`H_#R&)9C^GM4GQ=NV0/&/?+/.'B^)+,MHY>*P#]`TV MCWW2^6@-=5#Z#Y@HMA].:\.L,6I,-5FYK8GC54SSLG7$*]8L:#.3FOI15HH: MHB<#Q4P^7QI[>RO,]M%B-\>`L=KDXIUVY$D9?0EO(],0\$+3XM4R%EP"9D/; MH2/7VLG2[@?5^,R1C>PFK&O&LL9&]H7S->U4E&1OC46$-+[9:3(P(*SRY_$JP%2! M.L,`MFA`=!H"6M"A&;(6JUG6%F67]2V9H]SRDUEG&$>>G.J13G9K#Q8V^ MLMF'T%LK)SR;C6V$=BF/,]/$UF22P#?W"6U37@-/H4/7=&(KM.?3="V'CE63 M+>/.1R!E;=1Z&NW(+E_>./.!T4Y0:3E:!"Q6;D-'?YV[,E,Y-/"ZJ6U3Y,YZ MD?'01NKX:KHW0/$\CHKFBV",9L.,O/;TFC5>+ MK;V0<^55,9$O?(U7MC]OHJ^C-YR?:`&JZSLLB8]G7(S]?]"/ZS#?`K:IN M)_/74=K+,A=X?@U@K%_CHOM]OFB[83$Q&-VDO;?"MD5^@UWQ:6LY@`TNCBT7 ML&G+S^@--F]9@8VW\5C"C,VZ?N,?Z",;*2#!MC'S?:.9<>U*\&34()U>WK2" MFW/6"V9A\-QAR$J@'D-'ERNZS^V,-!I;>\T>K&[OIRX]FL;3INVY(/F M3.^EF-WOM,'8$XSL$;"VI!G&+FPH\"F8T$>3YD_'LD&S^<"IM3=Q)PTV]DGG MJ;O9'.H_X-W::UW;7%OJ#)/24,$A=DFQB%U?EC]T(SW4KV,'HF34*KU4*%&' MJVW#WLEV<#N"&)QB;AJGMT^K9&E`MEEZ17O"]C13@U78%FRWM'ZS@]T/E$NK M,V78MU\"!%C\(\]\TCBE\+O&& M#,V.L43-=GL!CAU$'B+WA5'&FY7K[5B;YIG'CC73/VN6CTGF[7AZROO,3H%Z M7`F_=6A5+:'8N!RW'?QN_GJ+3]<+@:OO-5V/9FVV?B7!=@2YYV$/&"P]@QAD27.OG37DAN:)!V0RF?WF'.J[.5+>D<-C/47AV;'58; MJ+O6OKX$-8_3![V4'F:SIC&00VD*-9;00GWAIG9#N#?<'6J#1+!<*HL4/F6"UYD?O MJ-G;]]D%I2]9.TSE=F9[?-7.*U8.)LX=NY))Z+KM(=;6^*!H65;-JU`;1#R^;C4X6S]H3.;5]L^1!?2LW/5&>I#(%N;B-W;+`]H M0^N2:.%5=ZQQ,HI>3HW&J?'-LL8@-4EY@+C"UEE[L"F_*VN8;M3VI+FN3EQ2 M>`W5VT[ZJF&4H6B+9`57"]_4L$9_]EOS7>O>)6AO'C/."^TN0ZIS]KQPQD;R MD`.P'6R.-C?9HQV'=%5WB&'5,XL8J*Q:H_E>M#&JDU/$%4&8-CSY;&N`QF%[ M$S.+P^J\+`03EPA0]DP*E$F:4=T#M&V[;!VM_@`#E*>^UFKL]0@T.ZJ#5?9* M68/$],J-][QZV;TQ6`#^M6G1_U%SX(OZ$EO5Q!($K$6_:\4WJWR44OQ)5CZ# M&[>ES$W-K2@1XCSX[N/^P93VHMCD7KJO9]T1\-)=XJ>TESC%#KG_>O=J>=>73ZP>TIEQ' MD]L0F.NH\W_O[>*QY&HG7T>K:G4`+79`<7TSNP[5OG MM,?0L.^J=7F;YE"NR)M=[TN=J!Y2Y2 MDBG`N\Z>KW7OOQE>%_=P^>":J5X!+SOPQ3O MJ>V']T&S:7"R>>UR5BLK9MG+W$5"-CIYH3SX1FQ/*!7=P3]*M):;LYV);M9\ MMI'`?N4:XOFZIXP"AV,O5$_;$>O[=3X0?JV*F&UKEGNJMFTY^&\[&+UO_%_/ M=9/1Q6UQ\8^/SV?`-AQ>`,MDMT@XW>AGZKF!G8&&D8]PA;Q@T]S$L?N5/8YD`)>(F9OQW@[G'# ML(N\4."*AZ=^YZ9IT4O4I+ M.D6.HLYN,$C7Z7K;U=R2;-^\;.*0[0Q\H=HWG0:>'Q[37&XY]I?;$SR%,($S M/=O@C\:=4$DQ,[U[-CNKN;O>I.0K[3H5B]QV+F'#G2G#X^]R]"I;@HI&[H*S MO-G;FTU_MG#:\6S)=@ECLO>/FNSI)KS6TMT1!&6+H>D")-)2MJ18.(Q(]COW MOK^XOUQ29C#WF+V:1`K+)*V7HD$=>"%4#:O-!H("P2G<0G`1XSO9@*F8WB/J M(MO"0&P@]3L+#G>SEB_]SA4I/@,9WVN ML0G/SO"MJ)%V\(T1?7W+6N&7ZDSAMX=1'OL+5QD.9T/*N-^0\YFY=MN--4JK M"$^(I907S^W!`><&^1)V\6*B'SGHW4AE8VLD_BD?!N8+334?WHHW=; MO:G/S-Q<,H!8H5U(569;DBF>9M[]ZD$QVVTO3$T;!HFLZ.>S.+"[4NHNK-`" M3CO:%>*/=JLZI!U./GNW@V$/8+]R_8[OJ[`$G3]BCU M(,M]_^^(]:0OIH4=]VD3GRO*O[XRZ>*Z89@`;TRN_:BBL&8-*]V@!+4?/#S^*_>1*90;P;`OB)CJ/ ML2';E@5(,:686KW@+(?NLU'BOTL=,(,10%Y7OD2+K[7`KH*T*QK\6GF`II!/ M1"NBL&^G:D8P9>VF[">453.)1@&T@`)7("1EEMYBO@&,^E:8KBYZ?ZWIQ2PT MR5<#4G(NN9:WG+&%3782MY?1!.QFM"$<5MQS(8M_^#+!/1>U=R@\S(>N;(I' MPMW;'FE"I@YX@QV/5KN2K:/+NN\7=\287BFV/G!OA0_!#P33H^,U1MR;CHVG MAO^;_.T7-`?72:Z0=N>^N&E].%P/>2`[#2I'OFS_;OG`X==).9C\X5>1#M'T MEK/,F')".5K\P+U0'IO:/%7DCT=QH+ZRPMVB=7*3AKW@O^D=[E;S8KH%08C' ML2/321-/:"^Y`]VQ6M%GG9G5+8#^&(7(#FF$+9&"@2JW&EK_B MRK?7P$MU+>CW[QAC-4Y+GF7#AM*"*S]4%!L&Y`U/:X/CJ>KH>#)R.B[2?E5[ M@GL"8^2U]SEYH^GV]HX#H+NEX7'E8#SY?OL$9A+KO=/5\L;!-:U/;9VU+@1# M-6F!7W']Y,\TY^TO+_ME*\G6]>]DMK2Z;&AWW@DL@V!\-'."Z6G/FLE?#N%6PFO%)>B#N?GR4?[9O#6# MCO'=@-=0:G.Q8PT8_UCGS/FED5VQ=;B2)9HD9_LMR6/.(<)1^=0/LS`EYVM(I#/+8M6683KQ=KXEYU]? M%6'E-$HQ^9?<,XN3R:3MX*CF#O M3I'<^=EB)9^\)OPGCQR#'!+GH?!!><.P4!ZFS#`CRD$(BO*<=]C\1MXOWY,+ M@U&)`E9M^8J<9&PI%W#7*+VREU&T.%*[86/9/PU*AAO-3W!K0L>=R;11GE^I.%L.?+<#6A1O&XW MT/O%<4-P.8+;1GXYCXU73TW4>&C`)_^QSNU%WD9*&`GF2]&#='NQMZBF[IES M$4?9PVZ+(;XCM^@PNF!2;!V>7I@' M*QGH>.1[J`\S]+L/?4Y3S'%^@^2GZ\#\.3[VUICO^LS>3@#[<[5K5;SV_N%^ M#&W5+6WL8XU/[MW!)6D:H!GE?^I#<,R<6AW[9E(&OCFV'>D)M\0653NSM*(? MH,G6D715L#92WP!R@5L/KL-\-/,?\>/;-,T/SOUITL'B5V1]I=Q<5+XM)WP# MJ@W?Z^IKIIS7I3K\#FH7OT.#67.D]Z85#KD9AD3.:NO;L,M?8MXU9XZ%W'D6 MRG_F;&.[>4OYF:5NQW#J?C.TFBM7-]:^<!ZWW9TCS_G@+NBPM[=8@.T&E)X3PI';;/*1:PD:3MXG?W%S MSR'AYG(\.5_0C]V]YI-#G+&*[&))N?]VCG/(J>@H=92Q[R)_O MT!G&JUT@=ZI\?;XJ/Q*7H*'5.W0O\H&642Y!+_V!P1&`9L?6A1H8Z]Q"]W)S MG=,.CEQBN3C=(4I:599SSC?H+ MR&W9A=H#*;:6??H[-F"^Q)O@.,,7)FB2)A[TBX9N&^>ID>[HY:A1_]F^I'OV M>L7J#W4D>;)\`KB7G45K!H*C46R;H)6/L?Z4;&`\7-X4P5-RH57MG2(FDPW6+HN"6S[$T='VE_ MS,.LWL23]LB3C/25>=V[!!U)?YH+KGFC`N7.*$%Y9L@TIY0# MKLWK+6'>Z"Y=O4Y.)'X#E('IM$CR`[Q0:TZ[A55:I&O=O?(.J/QU5Q'=T*\' M"BW@1^AJ.JQ2;ZX6X&FK,[GIGVL(N>"\)NQ0+Y\9NW%.;Z1-*SWWJ'C M_QS`XFX7^<-S1C[KPZMSQ1_K<6W6XF?RW+@=/D;H"3^@'[8'P]#P5GST<"V_-LL_><;5@!I:G[ MU="!T//V.O MH_'G:E>!.:)>C/]9RMC#Z,C(9?C&6TTA"QVR_TK]ZJG MI/?HL?-B.1&`)?K:M4TSH.6OIUK]N3)4`EL^YG:'.;63MW+ZN:-],6YE]$P2 MNMN^(FD^*&)<([$/QRJ&Q9V&_/(P^G"1QKZ.WA60Q2OLYF)CH#[UAHY&=H=/ ME.GL:G4S:ZJ]B*YK9[6C_#R7@.VSNJU]9]Y)7TW;7"&1*'1>^X\6@!Y81R^3 M&]NL_,P;:@1=U1Y0[P?&M7&2>NM<>T#]T7Y`;&O.!:&;G]^(.>U/:(W9O:R# ME3/KYLO->MO/D_W\)`-H&UF%DNCJXFC]0KSR!/B5-S&B*5&1:)#=&`C[)K87 M&'?KCU"T."6\]_P=W&\*0A>>_W!CNVN][HP+I&V[-?_I^M3=NA:4)HDT_`@7 MUPFTQW5D87*=;7EU=`@[UPN2`_",^73]CUY=]Y@+@[D3[G'MNEX<8_CV/IEC M0F>X*G.Z]SR9O.XT9Y:>US'0O/0-M/O;UEYTEU87OZ7FO/2J^47Y1ZG29$!` M8>Z(_'7":P'ROX[-&+(2K!=\PLNK^X'=^HV&J*:/0A'DF6O&@HNRINP*70,' M`1

RKO"UI;!R_0R\7V`F]THKGX\Y[F M)G=_D+XEP`P!T%O_O49B)^"2=!*Z)TR3PDP7/F% MK".,%G>LKZ=O$>LF17MN0=SC_H+]8)K2"A`,;O!-3M$#6443G[D2!3!YK[RC M^3[L7(!SPY66R(]I11E;K`\NPLT()+_L=%W@7[$(V M'^6/H4ROWPU`#XG=5`QT$0"#OVF-Q&HC)$%!+`O4:?.%$3YF:7O0P?>`UP4" M^>#N^E![;:>O1\@7K9K>!S:5LDDN;*&D".#TL?E0.T M`)@`946DX/.=;%EK/`N\_!3P1`C#JH/OSGLB)7N-T+]?(T^0^7TL="_P M";"35=(-'ZU1^IC[(PP@+SD"DP$W0.]7!)`=Y56X^K8`,P`=?(-OSOAOMU%V MX4\`O\(C@-7@STC7W:73^G"-L.&C`51`%2D3D`.@X*V&44U@[2E6$>_U>\); M.AM\A'?YPK[/Z2>$K_&-^>0+`=2K8T90Q$H%$+U'^A)]B$8OY$82TG?MHPNP MXFWQ3(#ZY!7539I"?&-N^JZ94X#-^U2Q_3B3!O&E!]>#Q=$N?'S0.>B%SPOR MV:*WHV=@H\E6X%A!?1$B!52#84!'(@$&!_\I<7QK<$=X0B2?3AEU/`B:\V\R:_AXW\75NNOQ?-] MA>.#FKY7H$RTKVG4GCNF05DN#('Y(P0TTDB*-PDLWTFH?%&(J3;\AANHY?.Q MXYD`+X`C0!/@5,POT#Y2&WRZ=CYYGZ(1(NKE"G%Z7AYWSH>24Y57L/S M@\^"X'A$8ZYQ]/<:M,8'`]'HYP:+([(6*?FLW3=#^FB-L4WY\#+REYAK;,.W M@(/R0<&+(P6T;'G990-L_39\C_A(_"3^,OK&5,0"P?>1W@VP^H'T._KE(^)F M!(\`GSXS@HR@Z'A@I?SI&R]\,X/PP9X@#1!X)]IN^(@`Z8&<*/(R`2@'"%BR M`3_",FLD@"V>[XYC!47*9:&:IO,99]*$^6EJK,BWXBNG1/B,_-VT"H^8S#FO MKB6]&,?8`B&40BT/G1$.B7>SJ.#6)+VQ]&F59'X/&Y,`\W@0Y$C^+'^8][@Z M`12``,>P/*X`$%OK4[W7%;N`K/?Y8&QW"V\88-L"&%^$8,%&K+6OIHC?O4>C M;;,39D<>(XQO"Z!J-&4_YM&=&'DC_#+>"C_T>_!1+<6IEN0"H!5`K,@&!%M2 M`43SG$72?%(4:WF:G\=W^D3R1`"G'UJ^L8L+?,WW"/Z3?+[4^ZTO(:B`-Q8< M2'/SOD:I.)*R-[\=[*D"YXGQ8<"%_.E2]<!U"13\+K M2=<`:H#^`NYVJC@%<`$<`93P)CY7XU]U6>L"$,P/^?A\)SYSJC8\"9^2KY2& M^<#P<0'=[XJU";#UH_65+:-\:8`J/#-8SX@$W'1G*LGP27D.(D">?:PU9,-K MX$U\`_EY?,18>_KSHPK^+C,$*/@L*Y\O&`^L9,16"YU\U456;*?//?^;!D+* M;B\"%8&+@.1VP_<$8`VL^>:.:/AJUX01$!$8&***>[<`.8!58X*O!^BXA=!+ MZ+'$)U%:H@N@1-^B]]>6`5+TB4[I+XV^8)@B_L["%J^!Z$YQ/$Z^@0B7-Q(D M#IN._7DX7R#R&V@KQ$W[:[F0N-"@\MYAPICJ&VMN%ON[&8)UGPNQQK<$.,#W M%^;QGOEN.)TP6[!X3WC25^>648$?J1W>7)N=K/'MX;7T>TA$O=X0#1^^6$W* M!-P`O'=V9@<2#D@$L'<.`82O>#_6(/'4#0BPN`&T`-(0-H",8,2@4#+U@\'; M!`"&9'IZ^2_Q;OZJU*RBW2$"5`#RP'BWG$!N<"I`%>0":(6J`DEAK=`V@"M4 '"`8#3P#Y!R_Q ` end From owner-mpi-ptop@CS.UTK.EDU Tue Mar 16 04:55:14 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA14277; Tue, 16 Mar 93 04:55:14 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA15025; Tue, 16 Mar 93 04:54:46 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Tue, 16 Mar 1993 04:54:45 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA15006; Tue, 16 Mar 93 04:54:01 -0500 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA13913 (5.65c/IDA-1.4.4); Tue, 16 Mar 1993 10:52:14 +0100 Received: by f1neuman.gmd.de id AA15815; Tue, 16 Mar 1993 10:53:37 GMT Date: Tue, 16 Mar 1993 10:53:37 GMT From: Rolf.Hempel@gmd.de Message-Id: <9303161053.AA15815@f1neuman.gmd.de> To: mpi-collcomm@cs.utk.edu, mpi-ptop@cs.utk.edu Subject: Al's COLLCOMM proposal Cc: gmap10@f1neuman.gmd.de I would like to comment on Al's Collective Communications draft which he sent out a few days ago. First of all, I agree with Al in most points, especially that we should not attempt to include everything into MPI-1. Given the limited time available, it seems to me a good idea to leave the level-1 routines and all dynamic stuff for MPI-2. In section 2 Al says that "each group has a topology associated with it". As far as I know this is still an open issue. Do we agree that there is always a default topology (like a ring, to make the shift operation meaningful in all cases, or fully connected)? Otherwise a topology is an optional attribute which a group may or may not have. Another question then is how this assignment is done. At the last Dallas meeting we discussed two basic options: 1. A topology is defined after the creation of the group. The topology thus is an attribute which is assigned to the group, and which can be overwritten without creating a new group. 2. A topology definition always creates a new group (or even two of them, the second one being the collection of processes which are not used by the topology). The advantage of this choice is that the rank of a process within a group never changes. When a group with topology is created, the processes can be arranged in the optimal way from the very beginning. Personally, I prefer the second option. One additional advantage is the following: assume that the original group has 10 processes, and then a (3,3) grid topology is defined. Does a global operation on this group include all 10 processes? If the (3,3) grid formation creates a new subgroup of 9 processes, the answer is clear. The draft is not consistent in the relationship of groups and topologies. In the Introduction it says "Giving or forming groups with topological features is presented in section 4" which suggests option 2. above. On the other hand, under Group Functions it states that "Existing groups including the ALL group can switch their associated topology", which sounds like option 1. Do we all agree on choosing option 2? On page 2 the draft states that "The collective communication routines are implemented in terms of the topology associated with a given group. Although the function would be the same, a broadcast in a group with a ring topology could be implemented differently from a broadcast in a group with hypercube topology". I see a confusion of application and machine topologies here. The optimal implementation of a broadcast is guided by the machine topology, which could be a hypercube. Even if the logical group topology is a mesh, the global operation would follow the hypercube structure. However, this implementation detail is completely invisible to the user and should not be part of the standard. The only thing the user sees is the mesh topology and the result of the broadcast. The proposed MPI_SHIFT function could be made much more useful by adding another argument. Here's my proposal: Info = MPI_SHIFT(inbuf,outbuf,nitems,type,tag,group,direction,steps) The additional integer argument "direction" selects the coordinate direction in the group topology, and "steps" is the number of steps in that direction. In the case of cartesian structures the meaning is immediately clear. One could apply the function also in the case of a general graph. In this case "direction" would specify the neighbor number. "steps" could either be ignored, or we could define a transitive scheme of the kind "neighbor of neighbor of neighbor ...", with the indirection depth being specified by "steps". A hot topic for further discussions will be the "group names" proposed by Al. I see his point, but I don't see how the user-supplied group name solves the problem which arises if a new process wants to join a group. Even if the user tells MPI the global name of the group, global knowledge of all groups in the system is required to find the other group members to talk to. I agree to most points of Rik Littlefields comments. The only thing which does not convince me yet is the explicite caching mechanism. If the information caching is handled consistently between the group management and collective communication routines (in order to avoid usage of stale group information), I still hope that it could be done without showing up at the user interface. As a last point, I would like to forward the following note by Tom Henderson: > Rolf, > > Would it be a good idea to merge the mpi-collcomm and mpi-ptop > mailing lists? It seems like lots of stuff on that mailing list > now is closely related to process topology. I suppose the > mpi-collcomm stuff could just be forwarded to the mpi-ptop list. > > Tom I agree. What do others think? Rolf From owner-mpi-ptop@CS.UTK.EDU Tue Mar 16 11:04:50 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA20347; Tue, 16 Mar 93 11:04:50 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA04802; Tue, 16 Mar 93 11:01:02 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Tue, 16 Mar 1993 11:01:01 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from fslg8.fsl.noaa.gov by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA04794; Tue, 16 Mar 93 11:00:59 -0500 Received: by fslg8.fsl.noaa.gov (5.57/Ultrix3.0-C) id AA09360; Tue, 16 Mar 93 16:00:54 GMT Received: by macaw.fsl.noaa.gov (4.1/SMI-4.1) id AA16673; Tue, 16 Mar 93 08:59:39 MST Date: Tue, 16 Mar 93 08:59:39 MST From: hender@macaw.fsl.noaa.gov (Tom Henderson) Message-Id: <9303161559.AA16673@macaw.fsl.noaa.gov> To: mpi-ptop@cs.utk.edu Subject: Re: Al's COLLCOMM proposal Rolf writes, > Another question then is how this assignment is done. At the last > Dallas meeting we discussed two basic options: > 1. A topology is defined after the creation of the group. The topology > thus is an attribute which is assigned to the group, and which can > be overwritten without creating a new group. > 2. A topology definition always creates a new group (or even two of > them, the second one being the collection of processes which are not > used by the topology). The advantage of this choice is that the > rank of a process within a group never changes. When a group with > topology is created, the processes can be arranged in the optimal > way from the very beginning. > > Personally, I prefer the second option. One additional advantage is the > following: assume that the original group has 10 processes, and then > a (3,3) grid topology is defined. Does a global operation on this group > include all 10 processes? If the (3,3) grid formation creates a new > subgroup of 9 processes, the answer is clear. I also favor option 2. I think the draft should be modified as Rolf suggests (elsewhere in his message) to be consistent with option 2. > On page 2 the draft states that "The collective communication routines > are implemented in terms of the topology associated with a given group. > Although the function would be the same, a broadcast in a group with > a ring topology could be implemented differently from a broadcast in > a group with hypercube topology". I see a confusion of application > and machine topologies here. The optimal implementation of a broadcast > is guided by the machine topology, which could be a hypercube. Even > if the logical group topology is a mesh, the global operation would > follow the hypercube structure. However, this implementation detail > is completely invisible to the user and should not be part of the > standard. The only thing the user sees is the mesh topology and the > result of the broadcast. Agree. The concepts of machine (physical) topology and logical topology should be clearly separated. > The proposed MPI_SHIFT function could be made much more useful by > adding another argument. Here's my proposal: > > Info = MPI_SHIFT(inbuf,outbuf,nitems,type,tag,group,direction,steps) > > The additional integer argument "direction" selects the coordinate > direction in the group topology, and "steps" is the number of steps > in that direction. In the case of cartesian structures the meaning is > immediately clear. One could apply the function also in the case of a > general graph. In this case "direction" would specify the neighbor > number. "steps" could either be ignored, or we could define a > transitive scheme of the kind "neighbor of neighbor of neighbor ...", > with the indirection depth being specified by "steps". I would not have a problem restricting MPI_SHIFT() to Cartesian topologies (ring, grid, torus, etc.). If a user has defined a general graph, let the user define her own "shift". Tom From owner-mpi-ptop@CS.UTK.EDU Tue Mar 16 13:43:10 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA25615; Tue, 16 Mar 93 13:43:10 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA12835; Tue, 16 Mar 93 13:42:15 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Tue, 16 Mar 1993 13:42:13 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from pnlg.pnl.gov by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA12826; Tue, 16 Mar 93 13:42:10 -0500 Received: from carbon.pnl.gov (130.20.188.38) by pnlg.pnl.gov; Tue, 16 Mar 93 10:35 PST Received: from sodium.pnl.gov by carbon.pnl.gov (4.1/SMI-4.1) id AA01248; Tue, 16 Mar 93 10:33:37 PST Received: by sodium.pnl.gov (4.1/SMI-4.0) id AA07439; Tue, 16 Mar 93 10:33:34 PST Date: Tue, 16 Mar 93 10:33:34 PST From: rj_littlefield@pnlg.pnl.gov Subject: Re: Al's COLLCOMM proposal To: Rolf.Hempel@gmd.de, mpi-collcomm@cs.utk.edu, mpi-ptop@cs.utk.edu Cc: gmap10@f1neuman.gmd.de, rj_littlefield@pnlg.pnl.gov Message-Id: <9303161833.AA07439@sodium.pnl.gov> X-Envelope-To: mpi-ptop@cs.utk.edu, mpi-collcomm@cs.utk.edu Rolf Hempel writes: > I agree to most points of Rik > Littlefields comments. The only thing which does not convince me yet > is the explicite caching mechanism. If the information caching is > handled consistently between the group management and collective > communication routines (in order to avoid usage of stale group > information), I still hope that it could be done without showing up > at the user interface. Just a point of clarification. I do NOT propose that cacheing be visible at the interface between the application program and a collective communication routine that it calls. The example I provided was perhaps not explicit enough on this point. It said: efficient_global_op (grphandle, ...) struct group_descriptor_type *grphandle; I intended "..." to mean only the arguments that would be provided to any collective communication routine, e.g., data buffer, number of elements, and so on. Nothing about cacheing there. I think Rolf would agree that the standard collective communication routines need an internal facility like this to coordinate with the standard group management routines, if they are to achieve high efficiency. My proposal is essentially to standardize and export that facility so as to permit new collective communication routines to run as efficiently as the built-ins. In this vein, you may wish to think of standardized cacheing as a feature to increase MPI's extensibility. --Rik ---------------------------------------------------------------------- rj_littlefield@pnl.gov (alias 'd39135') Rik Littlefield Tel: 509-375-3927 Pacific Northwest Lab, MS K1-87 Fax: 509-375-6631 P.O.Box 999, Richland, WA 99352 From owner-mpi-ptop@CS.UTK.EDU Wed Mar 17 15:33:52 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA24684; Wed, 17 Mar 93 15:33:52 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA08846; Wed, 17 Mar 93 15:33:15 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Wed, 17 Mar 1993 15:33:14 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from watson.ibm.com by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA08837; Wed, 17 Mar 93 15:33:12 -0500 Message-Id: <9303172033.AA08837@CS.UTK.EDU> Received: from YKTVMV by watson.ibm.com (IBM VM SMTP V2R3) with BSMTP id 5657; Wed, 17 Mar 93 15:33:12 EST Date: Wed, 17 Mar 93 15:33:12 EST From: "Marc Snir" To: MPI-PTOP@CS.UTK.EDU Some comments on how to support topologies -- on top on current collective communications The current proposal for collective communications and assumes that a group does not carry topological information: It is just an ordered set of processes. Assume that each graph has a canonical numbering, i.e., that a graph consists of an ORDERED set of nodes + a set of (directed) edges. I.e., the nodes of a general graph are ordered according to the order in which they are listed in an adjacency list, the nodes of a mesh are ordered in row major order, etc. Then a topology is an MPI group + a set of directed edges + some additional structure, for regular graphs, like meshes. The basic question to address is: should the MPI communication library be aware of topological structures? There are two choices: (i) a topology is a basic MPI object, that can be used in various communication operations. (ii) MPI communication functions have no understanding of topologies; thus, a topology is represented by two objects: the first one is a group that MPI is cognizant of, and the second one is a separate data structure that captures the topological information. Let me illustrate these two approaches by examples (I apologize for not sticking to the official syntax) 1. Mapping a topology so as to ease communication along edges (i.e., creating a new group with processes numbered so that the embedding of graph edges on the communication hardware is efficient) (i) MPI_CREATE_GRID(oldtopology, newtopology, topology_parameters). Creates a new topology, with a new process numbering. (ii) GRID_MAP(oldgroup, topology_parameters, newrank). A call that computes the best permutation of the group processes for embedding the grid. MPI_CREATE_GROUP(oldgroup, newgroup,0,newrank) Create new group. The situation where not all processes belong to the new group, can be handled, say, by returning a negative "newrank" for the outsiders, and then executing MPI_CREATE(oldgroup, newgroup, (newrank <0), newrank). (By the way, I doubt such function will have much use) 2. Sticking a new topology on top of an existing one (i) I don't know the solution (ii) Just create the new topology object (e.g., the adjacency list) 3. Shift on a row (or a hyperplane) of a grid (i) SHIFT_GRID(topology, inbuf, outbuf, dimension, shift) (ii) MPI_CREATE_GROUP(oldgroup, newgroup, grid_index(grid_params, myid,dimension),0) A new group needs to be created (once). I assume an auxiliary function that, given a description of the grid, a rank and a dimension, returns the index of the node with that rank in that dimension. MPI_CSHIFT(newgroup, inbuf, outbuf, shift) (or MPI_EOSHIFT, for an open topology). And so on. Consider a k-dimensional grid. Assume one creates a separate group for each one-dimensional subgrid. I.e., each node belongs to k subgroups, one for each dimension. Further assume that the rank of each node in a subgroup is equal to its index in the corresponding grid dimension (this holds true if the relative global order of the nodes is preserved when the subgroups are created). Then (almost) all the topological structure of the grid is captured by these subgroups: One can find the index of a node in the i-th dimension by querying its rank in the i-th subgroup. One can shift in the i-th dimension by shifting in the i-th subgroup. One can create a new subgrid that contains all nodes in a hyperplane, by using the index in the i-th dimension to split the group; the subgroups that are contained in this hyperplane are not affected. And so on. (The only information that is not captured is whether the grid is cyclic or not in each dimension. I.e., whether a cyclic shift or a end-off shift should be used in each dimension.) Thus, once the subgroups are built, it would seem we don't need any further topological information. The data structure that represents the grid is needed only to build these subgroups. This seems to be one more reason not to integrate topological information into the MPI communication functions. Temporary conclusion: topology cognizant MPI communication functions will save some writing, but not much. For the sake of keeping the communication layer simple, I vote for keeping topologies out of the MPI communication layer. This does not mean that topologies disappear. It means that topology objects are only used in order to build groups, and that only one MPI function really needs to understand topologies, for the purpose of embedding them on the communication hardware. Further topology functions could be provided, if required, as syntactic sugar on top of existing group functions. From owner-mpi-ptop@CS.UTK.EDU Thu Mar 18 09:42:52 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA11664; Thu, 18 Mar 93 09:42:52 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA07125; Thu, 18 Mar 93 09:40:15 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Thu, 18 Mar 1993 09:40:10 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA07073; Thu, 18 Mar 93 09:40:03 -0500 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA05710 (5.65c/IDA-1.4.4 for ); Thu, 18 Mar 1993 15:38:30 +0100 Received: by f1neuman.gmd.de id AA16221; Thu, 18 Mar 1993 15:39:55 GMT Date: Thu, 18 Mar 1993 15:39:55 GMT From: Rolf.Hempel@gmd.de Message-Id: <9303181539.AA16221@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: Marc's proposal Cc: gmap10@f1neuman.gmd.de Thanks to Marc for explaining in some detail his view of process topologies in MPI. I would like to add a few comments: 1. Marc summarizes his proposal with: > This seems to be one more reason not to integrate topological > information into the MPI communication functions. I agree that the topology functions should not change at all the pt-pt communication functions. The current draft already fulfills this requirement. Also the definitions of collective communications are not affected by the topologies, with the exception of the shift function where the topological information can add some useful functionality. The only point where the topology functions interact with the rest of MPI is the management of process groups. I do not see this as a disadvantage, but using topology information there helps in bringing stucture into groups. It seems that Marc agrees to this role of topology functions when he concludes: > It means that topology objects are only used in order to build > groups, ... 2. The question then is whether it is worthwhile to integrate process topologies with group management. I argue that the use of topologies like rings, grids, cubes, ... is so widespread that creating groups with these structures should be as easy and elegant as possible. For the same purpose there should be a (small) set of inquiry functions to support the addressing of processes in topology-based programs. In this sense we would not only save the MPI user "some writing", but we would also make MPI programming more natural to those who think in terms of their process topology structure rather than in terms of a set of processes, ranked from 0 to n-1. 3. The representation of a grid by the k subgroups works well as long as one only looks at neighboring processes in the primary directions. The handling of diagonal neighbors or arbritrary processes in the grid is less straight-forward. In my opinion it is better to provide a look-up function which returns the process id (or the group rank of the grid group) for any tuple (i,j,k) of grid coordinates. 4. The "Sticking a new topology on top of an existing one" could be done by adding an argument to a definition function like MPI_CART which specifies whether the processes are allowed to be reordered or not (with the purpose of optimal mapping onto the hardware). In this way, the "primary" topology could be defined with mapping MPI_CART(old_gid, cart_gid, ..., MPI_REORDER), and then another topology could be defined like MPI_GRAPH(cart_gid, graph_gid, ..., MPI_NOREORDER). In this way the graph nodes 0, 1, ..., n-1 could be associated with ranks 0, 1, ..., n-1 in the cartesian group. Are there any comments from other PTOP members? Rolf Hempel From owner-mpi-ptop@CS.UTK.EDU Thu Mar 18 13:00:35 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA15916; Thu, 18 Mar 93 13:00:35 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA17376; Thu, 18 Mar 93 13:00:03 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Thu, 18 Mar 1993 13:00:02 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from fslg8.fsl.noaa.gov by CS.UTK.EDU with SMTP (5.61++/2.8s-UTK) id AA17359; Thu, 18 Mar 93 12:59:58 -0500 Received: by fslg8.fsl.noaa.gov (5.57/Ultrix3.0-C) id AA14818; Thu, 18 Mar 93 17:59:52 GMT Received: by macaw.fsl.noaa.gov (4.1/SMI-4.1) id AA20665; Thu, 18 Mar 93 10:58:36 MST Date: Thu, 18 Mar 93 10:58:36 MST From: hender@macaw.fsl.noaa.gov (Tom Henderson) Message-Id: <9303181758.AA20665@macaw.fsl.noaa.gov> To: Rolf.Hempel@gmd.de Subject: Re: Marc's proposal Cc: mpi-ptop@cs.utk.edu Rolf, > 2. The question then is whether it is worthwhile to integrate process > topologies with group management. I argue that the use of topologies > like rings, grids, cubes, ... is so widespread that creating groups > with these structures should be as easy and elegant as possible. > For the same purpose there should be a (small) set of inquiry > functions to support the addressing of processes in topology-based > programs. In this sense we would not only save the MPI user > "some writing", but we would also make MPI programming more natural > to those who think in terms of their process topology structure > rather than in terms of a set of processes, ranked from 0 to n-1. I think that this is analogous to the argument for supporting multi-dimensional arrays in C, Fortran, etc. We could do everything we want with one-dimensional storage, but few people think this is a good idea! > 3. The representation of a grid by the k subgroups works well as long > as one only looks at neighboring processes in the primary directions. > The handling of diagonal neighbors or arbritrary processes in the > grid is less straight-forward. In my opinion it is better to > provide a look-up function which returns the process id (or the > group rank of the grid group) for any tuple (i,j,k) of grid > coordinates. I agree. > 4. The "Sticking a new topology on top of an existing one" could be > done by adding an argument to a definition function like MPI_CART > which specifies whether the processes are allowed to be reordered > or not (with the purpose of optimal mapping onto the hardware). > In this way, the "primary" topology could be defined with mapping > MPI_CART(old_gid, cart_gid, ..., MPI_REORDER), > and then another topology could be defined like > MPI_GRAPH(cart_gid, graph_gid, ..., MPI_NOREORDER). > In this way the graph nodes 0, 1, ..., n-1 could be associated with > ranks 0, 1, ..., n-1 in the cartesian group. I'm still not comfortable with the idea that MPI should attempt to deal with relationships between multiple topologies. It seems to me that the only way to do a good job would be to specify all topologies used by a group at group creation. I would prefer to see a new group created with the same members each time a new topology is required. > Rolf Hempel Tom Henderson NOAA Forecast Systems Laboratory hender@fsl.noaa.gov From owner-mpi-ptop@CS.UTK.EDU Tue Mar 23 08:12:05 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA13729; Tue, 23 Mar 93 08:12:05 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61+IDA+UTK-930125/2.8s-UTK) id AA08005; Tue, 23 Mar 93 08:11:24 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Tue, 23 Mar 1993 08:11:21 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61+IDA+UTK-930125/2.8s-UTK) id AA07997; Tue, 23 Mar 93 08:11:16 -0500 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA24548 (5.65c/IDA-1.4.4 for ); Tue, 23 Mar 1993 14:09:35 +0100 Received: by f1neuman.gmd.de id AA14713; Tue, 23 Mar 1993 14:11:03 GMT Date: Tue, 23 Mar 1993 14:11:03 GMT From: Rolf.Hempel@gmd.de Message-Id: <9303231411.AA14713@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: Next meeting Cc: gmap10@f1neuman.gmd.de Dear PTOP working group member, we have to prepare our discussions at the next Dallas meeting. Therefore, I just present you a list of points which I think we should address. All comments and additions are welcome. 1. The most important question seems to be how we want to support process topologies. The basic choice is between the positions which were presented by Marc Snir's and myself: a): MPI knows only about process groups which are linearly ordered collections of processes without any further structure. Process topologies are handled by libraries which are not part of MPI. b): Topological information is an integral part of process groups. There should be process definition functions for the most common topologies (cartesian and graph structures) which create process groups. Topologies can be used to define collective communication functions like shift. In the following I assume that we follow path b). (Otherwise we would not have much to do, anyway.) 2. It seems reasonable to me that we change the current draft such that a topology definition creates a new group. This removes the problem that the rank within the group may change. My proposal for the cartesian definition function is: active = MPI_CART (gid, parent_gid, ndim, dims, period). with: parent_gid: group id of existing group (input) gid: group id of newly created group (output) active: =.true., if calling process is contained in grid =.false. otherwise Depending on the return value, gid is the group id of the grid group or the remainder group. Thus, in case there are more processes in the parent group than are needed by the topology, two groups are created. If there is a perfect match, only one group is created (in this case there is no process which returns with active=.false. ===> there is no remainder gid.). The corresponding graph definition function then looks like: active = MPI_GRAPH (gid, parent_gid, nodes, index, nglist) The number of processes in the topology is no longer returned from these functions. This was rather redundant in the first place, and this information can easily be obtained from the inquiry function for the number of processes in a group. 3. The above definition functions could be extended by two arguments: 1. How should the mapping be done? One out of MPI_REORDER | MPI_NOREORDER | MPI_RANDOM MPI_REORDER: MPI is allowed to change the ranking of processes as compared to the parent group. MPI_NOREORDER: the rank of a process is the same in parent and new group. MPI_RANDOM: the rank in the new group is set randomly. (Is the third option necessary?) This argument allows an ordered overlay of two topologies, as in the following example: active = MPI_CART (cart_gid, parent_gid, ndim, dims, period, MPI_REORDER) active = MPI_GRAPH (graph_gid, cart_gid, nodes, index, nglist, MPI_NOREORDER) This way the graph nodes 0,1, ..., n-1 are associated with the processes with rank 0,1, ..., n-1 in the cartesian group. 2. For cartesian topologies: definition of ranking in the group: MPI_ROWM | MPI_COLM for row-major / column-major ordering. 4. Does every group have a default topology, and what could it be? My proposal is that we choose a ring topology as default, i.e., the rank of a process 0, ..., n-1 defines its position in the ring. No attempt is made to efficiently map the processes to the hardware. As an alternative, we could argue that there is no reason for closing the linear structure, and take a linear structure instead. I think that a default topology makes the interface more consistent. 5. Can processes in different topologies talk to each other? This question seems closely related to whether processes in different groups can communicate. I think that we can leave this decision to the more general debate on groups. The topology inquiry functions can be applied for groups of which the calling process is not a member, if this is decided to be necessary for other group inquiry functions. 6. Similarly, I think that we should not touch the question of what a process id is. The topology functions can adopt whatever scheme is decided in pt-to-pt. Any comments/additions? Rolf Hempel From owner-mpi-ptop@CS.UTK.EDU Tue Mar 23 11:14:56 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA20707; Tue, 23 Mar 93 11:14:56 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61+IDA+UTK-930125/2.8s-UTK) id AA16737; Tue, 23 Mar 93 11:13:50 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Tue, 23 Mar 1993 11:13:49 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from fslg8.fsl.noaa.gov by CS.UTK.EDU with SMTP (5.61+IDA+UTK-930125/2.8s-UTK) id AA16729; Tue, 23 Mar 93 11:13:47 -0500 Received: by fslg8.fsl.noaa.gov (5.57/Ultrix3.0-C) id AA23711; Tue, 23 Mar 93 16:13:41 GMT Received: by macaw.fsl.noaa.gov (4.1/SMI-4.1) id AA28595; Tue, 23 Mar 93 09:12:24 MST Date: Tue, 23 Mar 93 09:12:24 MST From: hender@macaw.fsl.noaa.gov (Tom Henderson) Message-Id: <9303231612.AA28595@macaw.fsl.noaa.gov> To: mpi-ptop@cs.utk.edu Subject: Re: Next meeting Rolf writes, > 2. It seems reasonable to me that we change the current draft such that > a topology definition creates a new group. This removes the problem > that the rank within the group may change. My proposal for the > cartesian definition function is: > > active = MPI_CART (gid, parent_gid, ndim, dims, period). > > with: > parent_gid: group id of existing group (input) > gid: group id of newly created group (output) > active: =.true., if calling process is contained in grid > =.false. otherwise > > Depending on the return value, gid is the group id of the grid > group or the remainder group. Thus, in case there are more processes > in the parent group than are needed by the topology, two groups > are created. If there is a perfect match, only one group is > created (in this case there is no process which returns > with active=.false. ===> there is no remainder gid.). > > The corresponding graph definition function then looks like: > > active = MPI_GRAPH (gid, parent_gid, nodes, index, nglist) > > The number of processes in the topology is no longer returned from > these functions. This was rather redundant in the first place, and > this information can easily be obtained from the inquiry function > for the number of processes in a group. I think this is a good idea. > 3. The above definition functions could be extended by two arguments: > > 1. How should the mapping be done? One out of > > MPI_REORDER | MPI_NOREORDER | MPI_RANDOM > > MPI_REORDER: MPI is allowed to change the ranking of processes > as compared to the parent group. > MPI_NOREORDER: the rank of a process is the same in parent and > new group. > MPI_RANDOM: the rank in the new group is set randomly. > (Is the third option necessary?) > > This argument allows an ordered overlay of two topologies, as > in the following example: > > active = MPI_CART (cart_gid, parent_gid, ndim, dims, period, > MPI_REORDER) > active = MPI_GRAPH (graph_gid, cart_gid, nodes, index, nglist, > MPI_NOREORDER) > > This way the graph nodes 0,1, ..., n-1 are associated with the > processes with rank 0,1, ..., n-1 in the cartesian group. I currently favor permitting MPI_REORDER only. It seems to me that the MPI_NOREORDER case could be layered on top of the rest of MPI if we decide we really want it. The current MPI point-to-point draft allows no communication between processes in group graph_gid and group cart_gid without reference to group cart_gid (or even parent_gid). In other words, I have to know my rank in group cart_gid to communicate anyway so why do I care if the rank in group graph_gid is the same? If I've hard-coded "rank" into my program, I can just look up my rank in group cart_gid. I think that including both MPI_REORDER and MPI_NOREORDER may cause some confusion. The purpose of MPI_REORDER is to provide better communication performance on some machines. The purpose of MPI_NOREORDER seems to be to provide a look-up table for convenience (performance would likely be worse on many machines). I might be persuaded otherwise if I saw a more detailed example of some pseudo-code that uses the topology overlay feature with the current MPI point-to-point draft (no inter-group communication). Could someone do that? > 2. For cartesian topologies: definition of ranking in the group: > > MPI_ROWM | MPI_COLM > > for row-major / column-major ordering. This may be the only sane way to handle this. I can't come up with anything better at the moment. :-) > 4. Does every group have a default topology, and what could it be? > My proposal is that we choose a ring topology as default, i.e., > the rank of a process 0, ..., n-1 defines its position in the ring. > No attempt is made to efficiently map the processes to the hardware. > As an alternative, we could argue that there is no reason for > closing the linear structure, and take a linear structure instead. > I think that a default topology makes the interface more consistent. I think a default topology is a good idea for all groups (including the default "ALL" group). Ring seems slightly better than linear since the collective communication proposal currently contains circular shift operations. > 5. Can processes in different topologies talk to each other? This > question seems closely related to whether processes in different > groups can communicate. I think that we can leave this decision > to the more general debate on groups. The topology inquiry functions > can be applied for groups of which the calling process is not a > member, if this is decided to be necessary for other group inquiry > functions. Agree. The context subcommittee and the point-to-point subcommittee can hash this out. > 6. Similarly, I think that we should not touch the question of what > a process id is. The topology functions can adopt whatever scheme > is decided in pt-to-pt. Agree. > Rolf Hempel > Tom Henderson From owner-mpi-ptop@CS.UTK.EDU Wed Mar 24 04:39:54 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA09656; Wed, 24 Mar 93 04:39:54 -0500 Received: from localhost by CS.UTK.EDU with SMTP (5.61+IDA+UTK-930125/2.8s-UTK) id AA01122; Wed, 24 Mar 93 04:39:17 -0500 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Wed, 24 Mar 1993 04:39:12 EST Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61+IDA+UTK-930125/2.8s-UTK) id AA01114; Wed, 24 Mar 93 04:39:07 -0500 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA12055 (5.65c/IDA-1.4.4 for ); Wed, 24 Mar 1993 10:37:31 +0100 Received: by f1neuman.gmd.de id AA15364; Wed, 24 Mar 1993 10:38:58 GMT Date: Wed, 24 Mar 1993 10:38:58 GMT From: Rolf.Hempel@gmd.de Message-Id: <9303241038.AA15364@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: Next Meeting Cc: gmap10@f1neuman.gmd.de Tom Henderson writes: > I currently favor permitting MPI_REORDER only. It seems to me that > the MPI_NOREORDER case could be layered on top of the rest of MPI if > we decide we really want it. The current MPI point-to-point draft > allows no communication between processes in group graph_gid and > group cart_gid without reference to group cart_gid (or even > parent_gid). In other words, I have to know my rank in group > cart_gid to communicate anyway so why do I care if the rank in group > graph_gid is the same? If I've hard-coded "rank" into my program, I > can just look up my rank in group cart_gid. > > I think that including both MPI_REORDER and MPI_NOREORDER may cause > some confusion. The purpose of MPI_REORDER is to provide better > communication performance on some machines. The purpose of > MPI_NOREORDER seems to be to provide a look-up table for convenience > (performance would likely be worse on many machines). I'm not completely convinced either that we need to have MPI_NOREORDER, and I don't have seen any "real" application yet where this feature is necessary. The reasoning which led to this feature was as follows: Assume you have ordered the group in a Cartesian grid, such that each process has been assigned its role in the topology. Now assume that you have an application which uses the grid in one phase, but it needs to communicate along a tree in another phase. For some reasons it could be important that the user can specify the position in the tree for each process in the grid, that, for example, process (1,1) becomes the root of the tree. A reason for this requirement could be that process (1,1) keeps some local data which the root of the tree needs, also. So, if the mapping optimization is done for both topologies (MPI_REORDER), then - most likely - two different processes will be selected, which means additional communication. Layering the MPI_NOREORDER functions on top of MPI works, but the disadvantage is that you have to duplicate the whole set of topology functions, including the inquiry routines. If one selects MPI_REORDER/MPI_NOREORDER in the topology definition, then all other functions can work the same for both options. On the issue of default topologies Tom writes: > I think a default topology is a good idea for all groups (including > the default "ALL" group). Ring seems slightly better than linear > since the collective communication proposal currently contains > circular shift operations. If we base the shift operation on topological information, then we do not have to support both the circular shift and the end-off shift, as in the current COLLCOMM proposal. I posted a corresponding definition on the COLLCOMM mailing list: Info = MPI_SHIFT(inbuf,outbuf,nitems,type,tag,group,direction,steps) The additional "direction" argument specifies the coordinate direction of the shift. If the coordinate direction is "periodic", then the shift is circular, otherwise it is end-off. I cannot immagine that one wants to have a circular shift in a non-periodic topology, or vice versa. The only remaining problem is the ALL group, because its topology cannot be specified by the user. Thus, there is only one shift version in that group. If the user wants the other one, he has to duplicate the ALL group and give it the other periodicity type. Does this sound reasonable? Rolf Hempel From owner-mpi-ptop@CS.UTK.EDU Mon Apr 19 10:33:27 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA02674; Mon, 19 Apr 93 10:33:27 -0400 Received: from localhost by CS.UTK.EDU with SMTP (5.61+IDA+UTK-930125/2.8s-UTK) id AA13408; Mon, 19 Apr 93 10:32:51 -0400 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Mon, 19 Apr 1993 10:32:50 EDT Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61+IDA+UTK-930125/2.8s-UTK) id AA13399; Mon, 19 Apr 93 10:32:47 -0400 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA04253 (5.65c/IDA-1.4.4 for ); Mon, 19 Apr 1993 16:32:38 +0200 Received: by f1neuman.gmd.de id AA12728; Mon, 19 Apr 1993 16:32:36 GMT Date: Mon, 19 Apr 1993 16:32:36 GMT From: Rolf.Hempel@gmd.de Message-Id: <9304191632.AA12728@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: new ptop definitions Cc: gmap10@f1neuman.gmd.de Dear PTOP subcommittee member, At the last Dallas meeting it was decided to base the topology functions in MPI on the lower level group definition functions. In the following I want to present to you my first thoughts about how this could be done. If you have any suggestions/corrections/brilliant ideas, please help. Proposal for process topologies in MPI: --------------------------------------- 1. The basic group definition function will look something like MPI_CREATE_GROUP (gid_old, gid_new, key, newrank) This function was proposed by Marc Snir, and it allows both a partitioning by key (using the ranking in the original group) and a reordering using the user supplied newrank value. The alternative which will be discussed is to provide two functions, one for partitioning and one for reordering. I think we could do with both options. We now have to base the topology definition on this group creation mechanism. 2. The first thing we need is for each topology a function which computes an appropriate reordering of a given group for a good mapping. As Marc suggested, this would obviously look like MPI_MAP_CART (gid_old, ndim, dims, period, newrank), and MPI_MAP_GRAPH (gid_old, nodes, index, nglist, newrank). The definition of the function arguments ndim, dims, period, nodes, index, nglist is the same as in the original proposal. These functions compute the newrank value to be used by the above group creation function (key can be set to any fixed value). 3. In the spirit of the layering of MPI functions, it could be a good idea to define high-level functions for topological group creation which are defined in terms of the above low-level functions. In this way we could define two functions like MPI_CART (gid_old, gid_cart, ndims, dims, period) and which is MPI_MAP_CART (gid_old, ndim, dims, period, newrank) MPI_CREATE_GROUP (gid_old, gid_cart, 0, newrank) and MPI_GRAPH (gid_old, gid_graph, nodes, index, nglist). which is MPI_MAP_GRAPH (gid_old, nodes, index, nglist, newrank) MPI_CREATE_GROUP (gid_old, gid_graph, 0, newrank) This way we would free the applications programmer from handling the rank reordering explicitely. Also, this definition allows an optimized implementation of MPI_CART without use of the lower level functions (if the implementor chooses to do so). 4. A general question is whether there is a standard way of caching internal information for process groups in MPI. This has been proposed in the Context subcommittee, and it was characterized as a possible addition to all the different context proposals. If this mechanism is available, the functions MPI_CART and MPI_GRAPH can store information about the topology with the newly created group. The following data would have to be kept: 1. topology type (cartesian/graph) 2. for cartesian structure: ndim, dims, period, pos (own position in grid) 3. for graph structure: nodes, pos This information can be used by later calls to inquiry functions MPI_INQMAP(gid) MPI_INQCART(maxdim, size, period, pos) MPI_INQGR(gid, nodes) The MPI_INQPID function should perhaps be redefined, so that it just returns the rank of a single process rather than a vector with several ids. The reason for this proposed change is that with the new strategy MPI_INQPID is a local operation without any communication. Thus, a seperate call for each look-up does not cost too much. As you see, the availability of a information caching facility makes it possible to retain the functionality of our previous proposal without the need to interfere with the group creation functions. The only (non-portable) low-level functions would be the two mapping functions. For an easy first implementation, they could choose the mapping permutation to be the identity. 5. If the information cacheing is not available, we have to find other ways for defining the inquiry functions. Obviously, the MPI_INQMAP function would go away. It would be the responsibility of the application programmer to remember the topology assigned to a given group. The same is true for MPI_INQCART and MPI_INQGR. MPI_INQPID would have to be changed. As input one would have to add the dimensions of the overall grid. The rank calculation of a given process then is almost trivial. 6. The MPI_PARTC(gid_old, gid_new, lostdim) function can be based on top of MPI_CREATE_GROUP (gid_old, gid_new, key, 0). Argument key is simply set to the index of the calling process in the coordinate direction "lostdim". The result is a collection of hyperplane groups which are orthogonal to that direction. I need your input for creating a proposal for the next Dallas meeting. Every comment is welcome. Best regards, Rolf From owner-mpi-ptop@CS.UTK.EDU Wed Apr 21 05:44:13 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA01566; Wed, 21 Apr 93 05:44:13 -0400 Received: from localhost by CS.UTK.EDU with SMTP (5.61+IDA+UTK-930125/2.8s-UTK) id AA01846; Wed, 21 Apr 93 05:43:37 -0400 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Wed, 21 Apr 1993 05:43:36 EDT Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61+IDA+UTK-930125/2.8s-UTK) id AA01838; Wed, 21 Apr 93 05:43:28 -0400 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA20888 (5.65c/IDA-1.4.4 for ); Wed, 21 Apr 1993 11:43:18 +0200 Received: by f1neuman.gmd.de id AA16453; Wed, 21 Apr 1993 11:43:18 GMT Date: Wed, 21 Apr 1993 11:43:18 GMT From: Rolf.Hempel@gmd.de Message-Id: <9304211143.AA16453@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: Information cacheing Cc: gmap10@f1neuman.gmd.de Recently I got a (private) email from Tom Henderson in which he comments on my latest proposal on process topologies. I think that it is interesting to the whole committee, and I hope it's okay with Tom if I forward it to you all: >> 4. A general question is whether there is a standard way of caching >> internal information for process groups in MPI. This has been >> proposed in the Context subcommittee, and it was characterized as >> a possible addition to all the different context proposals. If >> this mechanism is available, the functions MPI_CART and MPI_GRAPH >> can store information about the topology with the newly created >> group... >> >> Best regards, >> Rolf >> > >Rolf, > >It seems to me that we could still do this even if a caching facility >is not available to library users. All we need is some kind of >internal caching. It could possibly be implemented differently on >each system... > >Does this make sense? > >Tom The question of whether the PTOP functions can pass information between each other without the detour of the user argument lists is crucial for the design of the user interface. If the information cacheing is standardized in MPI, then the answer obviously is yes, and we can use that facility in PTOP. I agree with Tom that this is no necessary condition for an implementation of information cacheing for PTOP, but it would certainly help. Do you think we should rely on having some way or other of storing data internally? The note from Tom was the only reaction I got so far on my proposal. I really need your input, and I need it soon in order to get something done before the next meeting. I want to make sure that the proposal we produce on process topologies represents a consensus in the subcommittee as well as possible. Rolf From owner-mpi-ptop@CS.UTK.EDU Wed Apr 21 08:50:45 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA24140; Wed, 21 Apr 93 08:50:45 -0400 Received: from localhost by CS.UTK.EDU with SMTP (5.61+IDA+UTK-930125/2.8s-UTK) id AA12654; Wed, 21 Apr 93 08:49:33 -0400 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Wed, 21 Apr 1993 08:49:32 EDT Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from super.super.org by CS.UTK.EDU with SMTP (5.61+IDA+UTK-930125/2.8s-UTK) id AA12632; Wed, 21 Apr 93 08:49:29 -0400 Received: from b125.super.org by super.super.org (4.1/SMI-4.1) id AA22879; Wed, 21 Apr 93 08:49:23 EDT Received: by b125.super.org (4.1/SMI-4.1) id AA13747; Wed, 21 Apr 93 08:49:22 EDT Date: Wed, 21 Apr 93 08:49:22 EDT From: lederman@b125.super.org (Steve Huss-Lederman) Message-Id: <9304211249.AA13747@b125.super.org> To: mpi-ptop@cs.utk.edu Cc: gmap10@f1neuman.gmd.de In-Reply-To: Rolf.Hempel@gmd.de's message of Wed, 21 Apr 1993 11:43:18 GMT <9304211143.AA16453@f1neuman.gmd.de> Subject: Information cacheing I should have responded sooner to Rolf's note. I think it is fine to use this to write the next draft of the ptop section. The question of caching seems like a convenience if I understand it. If MPI does not support it, then a structure would have to be passed back to the user and used as input to other functions. Is this correct? I say we say there is a mechanism for getting such information and leave it vague until context is settled. Then the agrument lists, etc for the routines can be settled. Steve From owner-mpi-ptop@CS.UTK.EDU Wed Apr 21 11:01:15 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA29374; Wed, 21 Apr 93 11:01:15 -0400 Received: from localhost by CS.UTK.EDU with SMTP (5.61+IDA+UTK-930125/2.8s-UTK) id AA25596; Wed, 21 Apr 93 10:59:35 -0400 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Wed, 21 Apr 1993 10:59:35 EDT Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from gw1.fsl.noaa.gov by CS.UTK.EDU with SMTP (5.61+IDA+UTK-930125/2.8s-UTK) id AA25587; Wed, 21 Apr 93 10:59:33 -0400 Received: by gw1.fsl.noaa.gov (5.57/Ultrix3.0-C) id AA15249; Wed, 21 Apr 93 14:59:30 GMT Received: by macaw.fsl.noaa.gov (4.1/SMI-4.1) id AA24694; Wed, 21 Apr 93 08:58:05 MDT Date: Wed, 21 Apr 93 08:58:05 MDT From: hender@macaw.fsl.noaa.gov (Tom Henderson) Message-Id: <9304211458.AA24694@macaw.fsl.noaa.gov> To: mpi-ptop@cs.utk.edu Subject: Re: Information cacheing > From owner-mpi-ptop@cs.utk.edu Wed Apr 21 06:50:11 1993 > X-Resent-To: mpi-ptop@CS.UTK.EDU ; Wed, 21 Apr 1993 08:49:32 EDT Steve Huss-Lederman writes: > I should have responded sooner to Rolf's note. I think it is fine to > use this to write the next draft of the ptop section. The question of > caching seems like a convenience if I understand it. If MPI does not > support it, then a structure would have to be passed back to the user > and used as input to other functions. Is this correct? I say we > say there is a mechanism for getting such information and leave it > vague until context is settled. Then the agrument lists, etc for the > routines can be settled. > > Steve I agree. Tom Henderson From owner-mpi-ptop@CS.UTK.EDU Fri Apr 30 05:35:29 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA17837; Fri, 30 Apr 93 05:35:29 -0400 Received: from localhost by CS.UTK.EDU with SMTP (5.61+IDA+UTK-930125/2.8s-UTK) id AA27107; Fri, 30 Apr 93 05:34:40 -0400 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Fri, 30 Apr 1993 05:34:39 EDT Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61+IDA+UTK-930125/2.8s-UTK) id AA27090; Fri, 30 Apr 93 05:34:24 -0400 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA14184 (5.65c/IDA-1.4.4 for ); Fri, 30 Apr 1993 11:33:38 +0200 Received: by f1neuman.gmd.de id AA15005; Fri, 30 Apr 1993 11:33:51 GMT Date: Fri, 30 Apr 1993 11:33:51 GMT From: Rolf.Hempel@gmd.de Message-Id: <9304301133.AA15005@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: New draft Cc: gmap10@f1neuman.gmd.de Dear Collegue, time is going by quickly, and the next MPI meeting is approaching. I have put together a new draft of the topology section, which I send you for comment with this note as LaTex source. If you have any problem with processing it, please let me know. I will then send you a PostScript file. Please read the document carefully, and send me all your comments. I tried to follow the general decisions at the last MPI meeting, at the same time retaining most features of the previous version. I didn't try to write the text in such a polished style as before, with man pages and everything. Instead I used the same macros as are used in the PT2PT and COLLCOMM sections. This makes the text hardly readable, but this is not of highest priority at the moment. I really need your input, so that I can produce the new official version by next week. Rolf Hempel -------------------------- cut here ------------------------------------ % MPI Authors: % This is MY version of YOUR chapter. It has a wrapper so that you % should be able to simply LaTeX this. % % Please work from this text so that we are in synch. % % --Steve Otto \documentstyle[twoside,11pt]{report} \pagestyle{headings} %\markright{ {\em Draft Document of the MPI Standard,\/ \today} } \marginparwidth 0pt \oddsidemargin=.25in \evensidemargin .25in \marginparsep 0pt \topmargin=-.5in \textwidth=6.0in \textheight=9.0in \parindent=2em % ---------------------------------------------------------------------- % mpi-macs.tex --- man page macros, % discuss, missing, mpifunc macros % % ---------------------------------------------------------------------- % a couple of commands from Marc Snir, modified S. Otto \newlength{\discussSpace} \setlength{\discussSpace}{.7cm} \newcommand{\discuss}[1]{\vspace{\discussSpace} {\small {\bf Discussion:} #1} \vspace{\discussSpace} } \newcommand{\missing}[1]{\vspace{\discussSpace} {\small {\bf Missing:} #1} \vspace{\discussSpace} } \newlength{\codeSpace} \setlength{\codeSpace}{.3cm} \newcommand{\mpifunc}[1]{\vspace{\codeSpace} {\bf #1} \vspace{\codeSpace} } % ----------------------------------------------------------------------- % A few commands to help in writing MPI man pages % \def\twoc#1#2{ \begin{list} {\hbox to95pt{#1\hfil}} {\setlength{\leftmargin}{120pt} \setlength{\labelwidth}{95pt} \setlength{\labelsep}{0pt} \setlength{\partopsep}{0pt} \setlength{\parskip}{0pt} \setlength{\topsep}{0pt} } \item {#2} \end{list} } \outer\long\def\onec#1{ \begin{list} {} {\setlength{\leftmargin}{25pt} \setlength{\labelwidth}{0pt} \setlength{\labelsep}{0pt} \setlength{\partopsep}{0pt} \setlength{\parskip}{0pt} \setlength{\topsep}{0pt} } \item {#1} \end{list} } \def\manhead#1{\noindent{\bf{#1}}} \hyphenation{RE-DIS-TRIB-UT-ABLE sub-script mul-ti-ple} \begin{document} \setcounter{page}{1} \pagenumbering{roman} \title{ {\em D R A F T} \\ Document for a Standard Message-Passing Interface} \author{Scott Berryman, {\em Yale Univ} \\ James Cownie, {\em Meiko Ltd} \\ Jack Dongarra, {\em Univ. of Tennessee and ORNL} \\ Al Geist, {\em ORNL} \\ Bill Gropp, {\em ANL} \\ Rolf Hempel, {\em GMD} \\ Bob Knighten, {\em Intel} \\ Rusty Lusk, {\em ANL} \\ Steve Otto, {\em Oregon Graduate Inst} \\ Tony Skjellum, {\em Missisippi State Univ} \\ Marc Snir, {\em IBM T. J. Watson} \\ David Walker, {\em ORNL} \\ Steve Zenith, {\em Kuck \& Associates} } \date{March 25, 1993 \\ This work was supported by ARPA and NSF under contract number \#\#\#, by the National Science Foundation Science and Technology Center Cooperative Agreement No. CCR-8809615. } \maketitle \hfuzz=5pt %\tableofcontents %\begin{abstract} %We don't have an abstract yet. %\end{abstract} \setcounter{page}{1} \pagenumbering{arabic} \chapter{Process Topologies} \label{sec:topol} %----------------------------------------------------------------------- % % file topol.tex % \section{Introduction} A process group in MPI is a collection of n processes. Each process in the group is assigned a rank between 0 and n-1. In the majority of parallel applications this linear ordering does not reflect the communication structure of the processes which is usually determined by the underlying problem geometry and the numerical algorithm used. Often the processes are arranged in topological patterns as simple as a two-- or three--dimensional grid. For the user who programs an application on such a process structure the most natural way of referring to individual processes is to specify their coordinates in the process grid. Usually a process does not have to know the PID numbers of {\em all} processes in the grid, because communication only takes place with processes in a certain neighborhood, and perhaps with a few distant processes. Non--scalable features like the explicit listing of all participating processes in the application could, therefore, be avoided, if there were a means to specify the information which is really required. Additionally, the user--defined process structure can be exploited by the system in the assignment of processes to hardware processors, if this helps to improve the communication performance on a given machine. How this mapping is done in detail is outside the scope of this standardization proposal, since it is too much dependent on the particular hardware. The description of a ``virtual process topology'', however, only depends on the application, and is, therefore, machine--independent. The functions which are proposed in this paper only deal with this side of the mapping problem. There are well--known techniques for mapping grid/torus structures to hardware to\-po\-logies such as hypercubes or grids. But even for more complicated graph structures good heuristics often yield near--optimal results \cite{suprenum}. On the other hand, if there is no way for the user to specify the logical process arrangement as a ``virtual topology'', this information is most difficult to recover later. A random mapping is most likely the consequence. On some machines this will lead to unnecessary contention in the interconnect network. Some details about predicted and measured performance improvements which result from good process--to--processor mapping on modern wormhole--routing architectures can be found in \cite{wormhole1,wormhole2}. % \section{Virtual Topologies} % In the most general case the connection structure of a set of processes can be represented by a graph. The nodes stand for the processes, and the edges connect processes which communicate with each other. MPI provides message--passing between any pair of processes in a group. There is no requirement for opening a channel explicitly. Therefore, a ``missing link'' in the user--defined process graph does not prevent the corresponding processes from exchanging messages. It only means that this connection is neglected in the virtual topology. As a consequence, an automatic mapping tool will not try to assign the processes to neighboring processors. The graph definition function MPI\_GRAPH presented in this paper does not include any weighting of the communication volume along the different edges. Processes are either neighbors or not. Experience with similar techniques in PARMACS \cite{parmacs1,parmacs2} show that this information is usually sufficient for a good mapping. Additionally, a more precise specification would be more difficult for the user to set up, and it would make the interface functions more complicated. In principle a function for the definition of such a graph structure would be sufficient for all applications. However, in many applications the graph structure is very regular, and the detailed set--up of the graph would be inconvenient for the user and resource--consuming during run time. A large fraction of all parallel applications use process topologies like rings, or two-- or higher--dimensional grids or tori. These structures are completely defined by the number of dimensions and the numbers of processes in each coordinate direction. Also, the mapping of grids and tori is generally an easier problem then that of general graphs. Thus, it is desirable to do the mapping as grids and tori rather than obscuring the information in a graph. Therefore, in addition to the general graph cartesian structures of arbitrary dimension are supported. The topology definition function MPI\_CART can be used to describe both grids and tori by specifying per coordinate direction whether the process structure is periodic or not. Similar functions are contained in EXPRESS \cite{express} and PARMACS. Note that an n--dimensional hypercube is an n--dimensional torus with 2 processes per coordinate direction. Thus, special support for hypercube structures is not necessary. Based on the positions in the virtual topology a process can look up the PID numbers of other processes by calling MPI\_INQPID. The processes are specified by their logical node numbers (graph) or rectangular process coordinates (grid/torus). Thus, it is possible to inquire for the PID numbers of only those (few) processes with which messages have to be exchanged. \discuss{In the PT2PT discussion it is not yet decided how processes are addressed. This could, for example, be done by unique identifiers, or by their ranks in a group. The topology handling will be adapted to whatever decision will be taken in PT2PT.} % \section{Embedding in MPI} % The support for virtual topologies has to be fully integrated in the framework of MPI. In particular, the virtual topology functions do not interfere with the functions for process group creation and with those for collective communication. The current design of the process topology functions is based on the following assumptions: \begin{itemize} \item MPI is free to choose any process in a group for any location in the topology. There are no special processes which have to be aligned with topology substructures. This assumption fits best to a ``local SPMD model'' in which all processes in the same group execute the same program. \item One important feature of MPI are the process groups. The treatment of virtual topologies has to take this concept into account. The functions presented in this paper are not only compatible with the process groups, they even take advantage of them for the definition of mapping overlays and dynamic remapping during execution time. Virtual process topologies help in the subdivision of process groups which in many applications has to be done along process coordinate lines. Furthermore, the definition of collective gather--type communication calls benefits from the natural ordering of processes in a logical topology. \end{itemize} The virtual topology handling follows the general MPI rule of functional layering. At the bottom it is based on the general process group creation function MPI\_CREATE\_GROUP plus a few topology specific mapping functions. In principle this is sufficient for defining topologies and use them in applications. A number of higher--level functions provide an interface which is similar to what users of packages like Express and PARMACS are familiar with. This level can be written in a portable way in terms of the low--level functions. Implementations of the high--level functions which bypass the low level may be advantageous on some systems in terms of memory and cpu time usage, but this is not a requirement. % \section{Overview of the proposed MPI functions} % This chapter assumes the availability of the following MPI features which are defined in other chapters: \begin{enumerate} \item An existing process group (e.g., the ALL group) can be partitioned and/or renumbered using \mpifunc{MPI\_CREATE\_GROUP(gid\_old, gid\_new, key, newrank).} This function is called by all processes in group {\em gid\_old} loosely synchronously. It creates as many different subgroups as there are different values of {\em key}. The ranking in the newly created groups is determined by the {\em newrank} values, with the process ranks in the parent group for tie breaking. \item There is an internal information cacheing facility, which can be used by MPI routines to store and retrieve information related to a process group. This facility was proposed in the Context Subcommittee. High--level topology functions use this possibility to store structure information for a process group, which is then later used by topology inquiry functions. Also, special shift functions which are based on the topological structure can take advantage of this information. \end{enumerate} \discuss{If the information cacheing will not be part of MPI, the topology information can be stored in opaque objects, handles to which then have to be passed to the topology functions. Of course, this would be less elegant than an internal handling.} % \subsection{Low--level topology functions} % On the lowest level groups are created by the general function MPI\_CREATE\_GROUP. Using the input argument {\em newrank} the caller can change the process ranks in the new group as compared to the parent group. This feature can be used to efficiently map the processes in the group to the positions in the topology. For each class of topologies, therefore, there is a function which computes the rank of the calling process in the new group, based on the size and structure of the topology. Note that it is a valid implementation if the new ranks are identical to the old ranks. So, if for a given machine the placement of processes makes no difference, the implementation of the rank computation functions is trivial. For cartesian structures the function is \mpifunc{MPI\_MAP\_CART(gid\_old, ndim, dims, period, newrank)} with \begin{description} \item[IN gid\_old] group id of parent group, \item[IN ndim, dims, period] definition of cartesian structure, \item[OUT newrank] rank of calling process in reordered group. \end{description} The corresponding function for general graph structures is \mpifunc{MPI\_MAP\_GRAPH(god\_old, nodes, index, nglist, newrank)} with \begin{description} \item[IN gid\_old] group id of parent group, \item[IN nodes, index, nglist] definition of graph structure, \item[OUT newrank] rank of calling process in reordered group. \end{description} If the topology contains fewer processes than there are processes in the parent group, the newrank values for the remaining processes are set to MPI\_NORANK. % \subsection{High--level topology functions} % In principle, all what is needed to create an MPI process group with topological ordering are the two mapping functions and the general group creation routine. However, this is not what the applications programmer expects to see on his level of abstraction. Therefore, another layer of functions is defined on top of the low--level functions. This higher level can be portably implemented in terms of the level below. Of course MPI implementors are free to bypass the lower level for the sake of better performance or resource usage. For cartesian and graph structures the functions MPI\_CART and MPI\_GRAPH create a process group with topological structure. They combine the call to the mapping function with the call to the group creation routine. The user can decide whether the ranking in the new group may be reordered (for a more efficient mapping) or not. In the latter case the rank of the calling process is copied from the parent group. MPI\_CART is defined as \mpifunc{MPI\_CART(gid\_old, gid\_cart, ndims, dims, period, reorder)} with \begin{description} \item[IN gid\_old] group id of parent group, \item[OUT gid\_cart] group id of cartesian group (if calling process is part of it, else id of rest group), \item[IN ndim, dims, period] definition of cartesian structure, \item[IN reorder] ranking may be reordered or not. \end{description} The return value is the rank of the calling process in the cartesian group (or MPI\_NORANK). Apart from some local computations, MPI\_CART is based on the two function calls \mpifunc{MPI\_MAP\_CART (gid\_old, ndim, dims, period, newrank)} and \mpifunc{MPI\_CREATE\_GROUP (gid\_old, gid\_cart, 0, newrank)}. % \vspace*{3.0ex} For graph structures the corresponding routine is \mpifunc{MPI\_GRAPH(gid\_old, gid\_graph, nodes, index, nglist, reorder) } with \begin{description} \item[IN gid\_old] group id of parent group, \item[OUT gid\_graph] group id of graph group (if calling process is part of it, else id of rest group), \item[IN nodes, index, nglist] definition of graph structure, \item[IN reorder] ranking may be reordered or not. \end{description} % \vspace*{3.0ex} In the context subcommittee a mechanism has been proposed for cacheing information which is associated with a process group. If that feature will be part of MPI, the high--level topology functions can store the following information: \begin{enumerate} \item Type of topology (cartesian/graph) \item For a cartesian topology: \begin{itemize} \item ndim (number of dimensions), \item dims(*) (numbers of processes per coordinate direction), \item period(*) (periodicity information), \item own\_position(*) (own position in grid, could also be computed from rank and dims(*)) \end{itemize} \end{enumerate} For a graph structure no information has to be stored, because the graph node number is the rank in the graph group, and the size of the graph is the number of processes in that group. In case the cacheing mechanism does not become part of MPI, the above data have to be stored in an opaque object, a handle to which would get included in the argument lists of all related topology functions. Since this alternative would be much less elegant, the topology section would very much benefit from the inclusion of the cacheing facility. \subsection{Topology inquiry functions} If a topology has been defined with one of the above high--level functions, the topology information can be looked up using inquiry functions. They all are local calls, and do not require any communication. In the following the assumption is made that the information cacheing is available. Otherwise the handle to the opaque information object has to be added to the argument lists (or replace the group identifier argument). The function MPI\_INQMAP returns the type of mapping which is assigned to a group. \mpifunc{map = MPI\_INQMAP(gid).} The return value is one of the following: \begin{itemize} \item MPI\_CARTTOP: cartesian topology, \item MPI\_GRAPHTOP: graph topology, \item MPI\_NOTOP: no topology. \end{itemize} For a process group with cartesian ordering which has been created by MPI\_CART, MPI\_INQCART looks up the following info: \mpifunc{MPI\_INQCART(gid, maxdim, size, period, pos)} with \begin{description} \item[IN gid] group id of group with cartesian structure, \item[IN maxdim] length of vectors size(*), period(*), and pos(*) in the calling program, \item[OUT size] number of processes for each coordinate direction, \item[OUT period] periodicity yes/no for each coordinate direction, \item[OUT pos] position of calling process in cartesian structure. \end{description} For a process group with cartesian structure, the function MPI\_INQPID translates the logical process coordinates to process ids as they are used by the point--to--point routines: \mpifunc{pid = MPI\_INQPID(gid, location)} with \begin{description} \item[IN gid] group id of group with cartesian structure, \item[IN location] coordinate vector of process. \end{description} \discuss{Depending on the decisions taken in the PT2PT subcommittee, MPI\_INQPID might be changed to return the rank in the group rather than a pid.} \subsection{Partitioning of cartesian structures} If a group has been created with MPI\_CART, the function MPI\_PARTC can be used to partition the group into subgroups which form hyperplanes of the original process structure. As an example, assume that MPI\_CART has defined a two--dimensional grid structure on a process group. With a call to MPI\_PARTC this group can then be split up into disjoint subgroups which represent rows or columns of the original group. The function syntax is: \mpifunc{MPI\_PARTC(gid, lostdim)} with \begin{description} \item[IN gid] group id of group with cartesian structure, \item[IN lostdim] direction perpendicular to resulting hyperplanes. \end{description} The function returns the group identifier of the hyperplane group in which the calling process is contained. MPI\_PARTC uses the function MPI\_CREATE\_GROUP to partition the parent group, using the process coordinate in direction {\tt lostdim} as the key value. After a group has been partitioned by MPI\_PARTC, the inquiry functions MPI\_INQCART and MPI\_INQPID consistently refer to the grid sections if applied to a subgroup. If the process topology is a cartesian structure, a collective shift operation most likely is used along a coordinate direction. The function MPI\_SHIFTGR partitions a cartesian group into one--dimensional subgroups, which are all oriented in one coordinate direction. Every process is contained in exactly one subgroup: \mpifunc{MPI\_SHIFTGR(gid, direction)} with \begin{description} \item[IN gid] group id of group with cartesian structure, \item[IN direction] coordinate direction of one--dimensional subgroups. \end{description} MPI\_SHIFTGR uses the function MPI\_CREATE\_GROUP to partition the parent group. The return value is the group identifier of the one--dimensional subgroup to which the calling process belongs. \discuss{Instead of creating one--dimensional subgroups and then use the standard collective communication (circular or end--off) shift functions, one could also define a topology based shift function. It would look very much like the standard shift, with an additional argument defining the coordinate direction for the shift. Depending on the periodicity information of the cartesian structure, the shift function could choose the circular or end--off version automatically.} In a portable program the number of available processors usually is not known at compile time. In the case of a cartesian topology the function MPI\_MAKDIM helps in setting the numbers of processes per coodinate direction, depending on the number of processes in the group and optional constraints which can be specified by the user: \mpifunc{MPI\_MAKDIM(gid, ndim, dims)} with \begin{description} \item[IN gid] group id of group which is to be structured, \item[IN ndim] number of cartesian dimensions, \item[IN/OUT dims] number of processes in each coordinate direction. \end{description} The processes of the group with ID number {\tt gid} are divided as evenly as possible among {\tt ndim} dimensions, subject to the constraint that all processes in the group are used. The caller may further constrain the operation of this routine by specifying elements of array {\tt dims}. If {\tt dims(i)} is set to a positive number, the routine will fix the number of processes in coordinate direction {\tt i} to {\tt dims(i)}. Processes will then be distributed among directions where {\tt dims(i)} is set to 0, subject to the above constraints. Note that an error will occur if {\tt dims(i)} is positive and does not evenly divide the number of processes in the group. Negative input values of {\tt dims(i)} will also cause an error. All elements of {\tt dims(*)} must be set to 0 for completely automatic process distribution. When the function returns, {\tt dims(i)} will be set to the number of processes in coordinate direction {\tt i}. For {\tt dims(i)} not set by the user, {\tt dims(i)} will be ordered (MAXIMUM, ..., MINIMUM). Array {\tt dims(*)} is suitable for use as input to routine MPI\_CART. % \section{A simple example} % The example in figure \ref{poisson} shows how the grid definition and inquiry functions can be used in an application program. A partial differential equation, for example the Poisson equation, is to be solved on a rectangular domain. First the processes organize themselves in a two--dimensional structure. Each process then inquires for the PID numbers of its neighbors in the four directions. The numerical problem is solved by an iterative method, the details of which are hidden in the subroutine {\tt relax}. In each relaxation step each process computes new values for the solution grid function at all points owned by the process. Then the values at interprocess boundaries have to be exchanged with neighboring processes. For example, the exchange subroutine might contain a call like MPI\_SENDB(...,neigh\_pid(1),...) to send updated values to the left--hand neighbor (i-1,j). Note that there is only local communication in this exchange operation. \begin{figure} \begin{tabbing} =====\===\======\kill % \>integer ndim, mx\_neigh \\ \>parameter (ndim=2, mx\_neigh=4) \\ \>integer gid, gid\_cart, dims(ndim), neigh\_def(2), neigh\_pid(mx\_neigh) \\ \>integer own\_position(ndim), period(ndim), i, j, num\_neigh, ret, rank \\ \>real$\ast$8 u(0:101,0:101), f(0:101,0:101) \\ \>data dims / ndim $\ast$ 0 / \\ C \hspace{5mm} Create a grid structure in ALL group and inquire about own position\\ \>gid = ALL \\ C \hspace{5mm} Set process grid size and periodicity \\ \>ret = MPI\_MAKDIM(gid, ndim, dims) \\ \>if (ret .lt. 0) call error\_handler(ret) \\ \>period(1) = MPI\_PER \\ \>period(2) = MPI\_PER \\ \>rank = MPI\_CART (gid, gid\_cart, ndim, dims, period) \\ C \hspace{5mm} If not a grid process: stop \\ \>if (rank .eq. MPI\_NORANK) stop \\ \>call MPI\_INQCART (gid\_cart, ndim, dims, period, own\_position) \\ C \hspace{5mm} Look up the PID numbers for the neighbors. Own process coordinates are (i,j). \\ C \hspace{5mm} Neighbors are (i--1,j), (i+1,j), (i,j--1), (i,j+1) \\ \>i = own\_position(1) \\ \>j = own\_position(2) \\ \>neigh\_def(1) = i--1 \\ \>neigh\_def(2) = j \\ \>neigh\_pid(1) = MPI\_INQPID (gid, neigh\_def) \\ \>neigh\_def(1) = i+1 \\ \>neigh\_def(2) = j \\ \>neigh\_pid(2) = MPI\_INQPID (gid, neigh\_def) \\ \>neigh\_def(1) = i \\ \>neigh\_def(2) = j--1 \\ \>neigh\_pid(3) = MPI\_INQPID (gid, neigh\_def) \\ \>neigh\_def(1) = i \\ \>neigh\_def(2) = j+1 \\ \>neigh\_pid(4) = MPI\_INQPID (gid, neigh\_def) \\ \>num\_neigh = mx\_neigh \\ C \hspace{5mm} Initialize the grid functions and start the iteration \\ \>call init (u, f) \\ \>do 10 it=1,100 \\ \>\>call relax (u, f) \\ C \hspace{5mm} Exchange data with neighbor processes \\ \>\>call exchange (u, neigh\_pid) \\ 10 \>continue \\ \>call output (u) \\ \>end \\ \end{tabbing} \caption{Set--up of process structure for two--dimensional parallel Poisson solver.} \label{poisson} \end{figure} \end{document} From owner-mpi-ptop@CS.UTK.EDU Tue May 4 11:09:49 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA18199; Tue, 4 May 93 11:09:49 -0400 Received: from localhost by CS.UTK.EDU with SMTP (5.61+IDA+UTK-930125/2.8s-UTK) id AA11791; Tue, 4 May 93 11:08:28 -0400 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Tue, 4 May 1993 11:08:27 EDT Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from super.super.org by CS.UTK.EDU with SMTP (5.61+IDA+UTK-930125/2.8s-UTK) id AA11783; Tue, 4 May 93 11:08:25 -0400 Received: from b124.super.org by super.super.org (4.1/SMI-4.1) id AA10322; Tue, 4 May 93 11:08:22 EDT Received: by b124.super.org (4.1/SMI-4.1) id AA02846; Tue, 4 May 93 11:08:19 EDT Date: Tue, 4 May 93 11:08:19 EDT From: lederman@super.org (Steve Huss-Lederman) Message-Id: <9305041508.AA02846@b124.super.org> To: mpi-ptop@cs.utk.edu Subject: comments on 4/25 draft I generally like Rolf's chapter. Thanks for doing it :-). Below are a few points: On p. 2 is a discussion of why not to have weights in MPI_GRAPH. Why not have a second call that allows weights. If the user knows them then this information can be useful. To use Rolf's own argument against himself: It is most difficult to recover later :-) I agree it takes more work, so have the two level approach and make a separate call for the more complex case. An implementation can always ignore the weights and do a regular MPI_GRAPH. (which can also be ignored ;-) On p. 3 near the top the first bullet says: "fits best to a local SPMD model in which...." Why is this the case because we don't have alignment with substructures? I am probably being slow because I don't see the argument. In the MPI_CREATE_GROUP (bottom p. 3) says "The ranking in the newly created groups is determined by the newrank values, with the process ranks in the parent group for tie breaking." I'm missing something. Is newrank both input and output? How does the newrank determine the rank? I assume a later draft will have more details on the arguments concerning the topologies in MPI_MAP_CART/GRAPH (p. 4). The issue of how we number the location (coordinates) in MPI_INQPID seems lost (p. 6 near bottom). Did we vote to have both and place that information in the descriptor. (Both being 0 and 1 up numbering). Is this correct? Either way the draft needs to state which one. If it is both then the create and inquire functions need some changes. And now for a small point to wonder about when you have nothing else to do: In the example on p. 9 do you need the line: if(rank .eq. MPI_NORANK) stop? If MPI_MAKDIM works correctly then it places all the nodes in one of the dimensions. thus all processes get a group. Correct? Just added this one to let you know I actually read the whole thing :-) Steve From owner-mpi-ptop@CS.UTK.EDU Wed May 5 05:27:50 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA18080; Wed, 5 May 93 05:27:50 -0400 Received: from localhost by CS.UTK.EDU with SMTP (5.61+IDA+UTK-930125/2.8s-UTK) id AA01022; Wed, 5 May 93 05:26:38 -0400 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Wed, 5 May 1993 05:26:36 EDT Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61+IDA+UTK-930125/2.8s-UTK) id AA00833; Wed, 5 May 93 05:26:20 -0400 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA21501 (5.65c/IDA-1.4.4 for ); Wed, 5 May 1993 11:25:26 +0200 Received: by f1neuman.gmd.de id AA15664; Wed, 5 May 1993 11:25:42 GMT Date: Wed, 5 May 1993 11:25:42 GMT From: Rolf.Hempel@gmd.de Message-Id: <9305051125.AA15664@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: Revised draft Cc: gmap10@f1neuman.gmd.de Hi there, many thanks to Steve and Tom for reading my draft so carefully, and for their many suggestions for improvements. I tried to implement all their comments into a revised version which I send you with this note. In an additional message you will receive the corresponding PostScript. Please send me your comments as soon as possible, because I will have to give the text to Steve Otto for inclusion into the MPI document soon. Apart from the notes of Steve and Tom I did not get any comments. This can either mean that the text as it stands is okay with everybody else, or that nobody else cares any more. Pessimistic as I usually am, I fear the second alternative comes nearer to the truth. Any comment? Steve raised two points which I would like to shortly answer here. He wrote: > On p. 3 near the top the first bullet says: "fits best to a local SPMD > model in which...." Why is this the case because we don't have > alignment with substructures? I am probably being slow because I > don't see the argument. What I meant is the following: the system is free to choose the topological location for a process (which is already running) only if they all run the same program. If, for example, a grid boundary were handled by a different code, this assumption would no longer hold. This is what I called the local SPMD program (local, because it doesn't matter if other process groups run different programs). Steve then continues: > In the MPI_CREATE_GROUP (bottom p. 3) says "The ranking in the newly > created groups is determined by the newrank values, with the process > ranks in the parent group for tie breaking." I'm missing something. > Is newrank both input and output? How does the newrank determine the > rank? I referred to the group creation funtion which was proposed by Marc Snir (under the name MPI_NEW_CONTEXT). I renamed his input argument "index" to "newrank" to avoid confusion with the index vector of the graph mapping function. (Sorry that this way I obviously created more confusion than I avoided). Thus, newrank is an input argument, and it is used for determining the rank in the newly created group. If two or more processes call the function with the same newrank value, their ranks in the parent group are used for tie breaking. Best regards, Rolf ---------------------- LaTex source follows ---------------------------- % MPI Authors: % This is MY version of YOUR chapter. It has a wrapper so that you % should be able to simply LaTeX this. % % Please work from this text so that we are in synch. % % --Steve Otto \documentstyle[twoside,11pt]{report} \pagestyle{headings} %\markright{ {\em Draft Document of the MPI Standard,\/ \today} } \marginparwidth 0pt \oddsidemargin=.25in \evensidemargin .25in \marginparsep 0pt \topmargin=-.5in \textwidth=6.0in \textheight=9.0in \parindent=2em % ---------------------------------------------------------------------- % mpi-macs.tex --- man page macros, % discuss, missing, mpifunc macros % % ---------------------------------------------------------------------- % a couple of commands from Marc Snir, modified S. Otto \newlength{\discussSpace} \setlength{\discussSpace}{.7cm} \newcommand{\discuss}[1]{\vspace{\discussSpace} {\small {\bf Discussion:} #1} \vspace{\discussSpace} } \newcommand{\missing}[1]{\vspace{\discussSpace} {\small {\bf Missing:} #1} \vspace{\discussSpace} } \newlength{\codeSpace} \setlength{\codeSpace}{.3cm} \newcommand{\mpifunc}[1]{\vspace{\codeSpace} {\bf #1} \vspace{\codeSpace} } % ----------------------------------------------------------------------- % A few commands to help in writing MPI man pages % \def\twoc#1#2{ \begin{list} {\hbox to95pt{#1\hfil}} {\setlength{\leftmargin}{120pt} \setlength{\labelwidth}{95pt} \setlength{\labelsep}{0pt} \setlength{\partopsep}{0pt} \setlength{\parskip}{0pt} \setlength{\topsep}{0pt} } \item {#2} \end{list} } \outer\long\def\onec#1{ \begin{list} {} {\setlength{\leftmargin}{25pt} \setlength{\labelwidth}{0pt} \setlength{\labelsep}{0pt} \setlength{\partopsep}{0pt} \setlength{\parskip}{0pt} \setlength{\topsep}{0pt} } \item {#1} \end{list} } \def\manhead#1{\noindent{\bf{#1}}} \hyphenation{RE-DIS-TRIB-UT-ABLE sub-script mul-ti-ple} \begin{document} \setcounter{page}{1} \pagenumbering{roman} \title{ {\em D R A F T} \\ Document for a Standard Message-Passing Interface} \author{Scott Berryman, {\em Yale Univ} \\ James Cownie, {\em Meiko Ltd} \\ Jack Dongarra, {\em Univ. of Tennessee and ORNL} \\ Al Geist, {\em ORNL} \\ Bill Gropp, {\em ANL} \\ Rolf Hempel, {\em GMD} \\ Bob Knighten, {\em Intel} \\ Rusty Lusk, {\em ANL} \\ Steve Otto, {\em Oregon Graduate Inst} \\ Tony Skjellum, {\em Missisippi State Univ} \\ Marc Snir, {\em IBM T. J. Watson} \\ David Walker, {\em ORNL} \\ Steve Zenith, {\em Kuck \& Associates} } \date{March 25, 1993 \\ This work was supported by ARPA and NSF under contract number \#\#\#, by the National Science Foundation Science and Technology Center Cooperative Agreement No. CCR-8809615. } \maketitle \hfuzz=5pt %\tableofcontents %\begin{abstract} %We don't have an abstract yet. %\end{abstract} \setcounter{page}{1} \pagenumbering{arabic} \chapter{Process Topologies} \label{sec:topol} %----------------------------------------------------------------------- % % file topol.tex % \section{Introduction} A process group in MPI is a collection of n processes. Each process in the group is assigned a rank between 0 and n-1. In the majority of parallel applications this linear ordering does not reflect the communication structure of the processes which is usually determined by the underlying problem geometry and the numerical algorithm used. Often the processes are arranged in topological patterns as simple as a two-- or three--dimensional grid. For the user who programs an application on such a process structure the most natural way of referring to individual processes is to specify their coordinates in the process grid. Usually a process does not have to know the PID numbers of {\em all} processes in the grid, because communication only takes place with processes in a certain neighborhood, and perhaps with a few distant processes. Non--scalable features like the explicit listing of all participating processes in the application could, therefore, be avoided, if there were a means to specify the information which is really required. Additionally, the user--defined process structure can be exploited by the system in the assignment of processes to hardware processors, if this helps to improve the communication performance on a given machine. How this mapping is done in detail is outside the scope of this standardization proposal, since it is too much dependent on the particular hardware. The description of a ``virtual process topology'', however, only depends on the application, and is, therefore, machine--independent. The functions which are proposed in this paper only deal with this side of the mapping problem. There are well--known techniques for mapping grid/torus structures to hardware to\-po\-logies such as hypercubes or grids. But even for more complicated graph structures good heuristics often yield near--optimal results \cite{suprenum}. On the other hand, if there is no way for the user to specify the logical process arrangement as a ``virtual topology'', this information is most difficult to recover later. A random mapping is most likely the consequence. On some machines this will lead to unnecessary contention in the interconnect network. Some details about predicted and measured performance improvements which result from good process--to--processor mapping on modern wormhole--routing architectures can be found in \cite{wormhole1,wormhole2}. % \section{Virtual Topologies} % In the most general case the connection structure of a set of processes can be represented by a graph. The nodes stand for the processes, and the edges connect processes which communicate with each other. MPI provides message--passing between any pair of processes in a group. There is no requirement for opening a channel explicitly. Therefore, a ``missing link'' in the user--defined process graph does not prevent the corresponding processes from exchanging messages. It only means that this connection is neglected in the virtual topology. As a consequence, an automatic mapping tool will not try to assign the processes to neighboring processors. The graph definition function MPI\_GRAPH presented in this paper does not include any weighting of the communication volume along the different edges. Processes are either neighbors or not. Experience with similar techniques in PARMACS \cite{parmacs1,parmacs2} show that this information is usually sufficient for a good mapping. Additionally, a more precise specification would be more difficult for the user to set up, and it would make the interface functions more complicated. \discuss{One could also have two graph mapping functions, one of which uses communication weighting, whereas the other does not. The mapping function with weighting, \linebreak MPI\_MAP\_WGRAPH, would then have an additional input argument {\tt weights} with the same structure as the neighbor list {\tt nglist}. On the higher abstraction level the function MPI\_GRAPH would be complemented by MPI\_WGRAPH in the same way.} In principle a function for the definition of such a graph structure would be sufficient for all applications. However, in many applications the graph structure is very regular, and the detailed set--up of the graph would be inconvenient for the user and resource--consuming during run time. A large fraction of all parallel applications use process topologies like rings, or two-- or higher--dimensional grids or tori. These structures are completely defined by the number of dimensions and the numbers of processes in each coordinate direction. Also, the mapping of grids and tori is generally an easier problem then that of general graphs. Thus, it is desirable to do the mapping as grids and tori rather than obscuring the information in a graph. Therefore, in addition to the general graph cartesian structures of arbitrary dimension are supported. The topology definition function MPI\_CART can be used to describe both grids and tori by specifying per coordinate direction whether the process structure is periodic or not. Similar functions are contained in EXPRESS \cite{express} and PARMACS. Note that an n--dimensional hypercube is an n--dimensional torus with 2 processes per coordinate direction. Thus, special support for hypercube structures is not necessary. Based on the positions in the virtual topology a process can look up the PID numbers of other processes by calling MPI\_INQPID. The processes are specified by their logical node numbers (graph) or rectangular process coordinates (grid/torus). Thus, it is possible to inquire for the PID numbers of only those (few) processes with which messages have to be exchanged. \discuss{In the PT2PT discussion it is not yet decided how processes are addressed. This could, for example, be done by unique identifiers, or by their ranks in a group. The topology handling will be adapted to whatever decision will be taken in PT2PT.} % \section{Embedding in MPI} % The support for virtual topologies has to be fully integrated in the framework of MPI. In particular, the virtual topology functions do not interfere with the functions for process group creation and with those for collective communication. The current design of the process topology functions is based on the following assumptions: \begin{itemize} \item MPI is free to choose any process in a group for any location in the topology. There are no special processes which have to be aligned with topology substructures. This assumption fits best to a ``local SPMD model'' in which all processes in the same group execute the same program. \item One important feature of MPI are the process groups. The treatment of virtual topologies has to take this concept into account. The functions presented in this paper are not only compatible with the process groups, they even take advantage of them for the definition of mapping overlays and dynamic remapping during execution time. Virtual process topologies help in the subdivision of process groups which in many applications has to be done along process coordinate lines. Furthermore, the definition of collective gather--type communication calls benefits from the natural ordering of processes in a logical topology. \end{itemize} The virtual topology handling follows the general MPI rule of functional layering. At the bottom it is based on the general process group creation function MPI\_CREATE\_GROUP plus a few topology specific mapping functions. In principle this is sufficient for defining topologies and use them in applications. A number of higher--level functions provide an interface which is similar to what users of packages like Express and PARMACS are familiar with. This level can be written in a portable way in terms of the low--level functions. Implementations of the high--level functions which bypass the low level may be advantageous on some systems in terms of memory and cpu time usage, but this is not a requirement. % \section{Overview of the proposed MPI functions} % This chapter assumes the availability of the following MPI features which are defined in other chapters: \begin{enumerate} \item An existing process group (e.g., the ALL group) can be partitioned and/or renumbered using \mpifunc{MPI\_CREATE\_GROUP(gid\_old, gid\_new, key, newrank).} This function is called by all processes in group {\em gid\_old} loosely synchronously. It creates as many different subgroups as there are different values of {\em key}. The ranking in the newly created groups is determined by the {\em newrank} values, with the process ranks in the parent group for tie breaking. \item There is an internal information cacheing facility, which can be used by MPI routines to store and retrieve information related to a process group. This facility was proposed in the Context Subcommittee. High--level topology functions use this possibility to store structure information for a process group, which is then later used by topology inquiry functions. Also, special shift functions which are based on the topological structure can take advantage of this information. \end{enumerate} \discuss{If the information cacheing will not be part of MPI, passing the topology information for a group from one topology function to the other can be handled in several other ways which will be discussed later.} There has been a general decision in MPI to start all numbering with 0 instead of 1. Thus, coordinates in a cartesian structure are 0,1, ..., dim-1. In the same way, nodes in a general graph are numbered zero up. % \subsection{Low--level topology functions} % On the lowest level groups are created by the general function MPI\_CREATE\_GROUP. Using the input argument {\em newrank} the caller can change the process ranks in the new group as compared to the parent group. This feature can be used to efficiently map the processes in the group to the positions in the topology. For each class of topologies, therefore, there is a function which computes the rank of the calling process in the new group, based on the size and structure of the topology. Note that it is a valid implementation if the new ranks are identical to the old ranks. So, if for a given machine the placement of processes makes no difference, the implementation of the rank computation functions is trivial. For cartesian structures the function is \mpifunc{newrank = MPI\_MAP\_CART(gid\_old, ndim, dims, period)} with \begin{description} \item[IN gid\_old] group id of parent group, \item[IN ndim] number of dimensions of cartesian structure, \item[IN dims(*)] number of processes in each coordinate direction, \item[IN period(*)] periodicity specification in each coordinate direction. \end{description} The return value is the reordered rank of the calling process. \\[4.0ex] The corresponding function for general graph structures is \mpifunc{newrank = MPI\_MAP\_GRAPH(gid\_old, nodes, index, nglist)} with \begin{description} \item[IN gid\_old] group id of parent group, \item[IN nodes] number of graph nodes, \item[IN index(*), nglist(*)] definition of graph structure: For each graph node its set of neighbors is entered, i.~e., the set of processes with which the node process is going to communicate most frequently. Since there is no weighting of node connections, it might be sensible to specify only those neighbors with which there is a considerable amount of communication. For each node $i\;\:(0 \leq i \leq nodes-1)$ the list of the neighbor nodes is stored at $nglist(j),\;\: index(i) \leq j \leq index(i+1)-1$. For the last process $index(i+1)$ points to the first free position in vector nglist. Note that the matrix is symmetric, i.~e., if node {\em i} is a neighbor of node {\em j}, than node {\em j} is also a neighbor of node {\em i}. This property can be used for an internal checking of the input arguments. \end{description} Again, the function returns the reordered rank of calling process. \\[4.0ex] If the topology contains fewer processes than there are processes in the parent group, the newrank values for the remaining processes are set to MPI\_NORANK. % \subsection{High--level topology functions} % In principle, all that is needed to create an MPI process group with topological ordering are the two mapping functions and the general group creation routine. However, this is not what the applications programmer expects to see on his level of abstraction. Therefore, another layer of functions is defined on top of the low--level functions. This higher level can be portably implemented in terms of the level below. Of course MPI implementors are free to bypass the lower level for the sake of better performance or resource usage. For cartesian and graph structures the functions MPI\_CART and MPI\_GRAPH create a process group with topological structure. They combine the call to the mapping function with the call to the group creation routine. The user can decide whether the ranking in the new group may be reordered (for a more efficient mapping) or not. In the latter case the rank of the calling process is copied from the parent group. The option of not reordering the process ranks can be useful in the case of an overlay of different topologies. The user can then fully control which location in the first topology coincides with which location in the second one. The price which one obviously has to pay in this case is that for only one topology the process arrangement can be optimized. MPI\_CART is defined as \mpifunc{MPI\_CART(gid\_old, gid\_cart, ndims, dims, period, reorder)} with \begin{description} \item[IN gid\_old] group id of parent group, \item[OUT gid\_cart] group id of cartesian group (if calling process is part of it, else id of rest group), \item[IN ndim, dims(*), period(*)] definition of cartesian structure, (for details see the definition of MPI\_MAP\_CART) \item[IN reorder] ranking may be reordered or not. \end{description} The return value is the rank of the calling process in the cartesian group (or MPI\_NORANK). Apart from some local computations, MPI\_CART is based on the two function calls \mpifunc{MPI\_MAP\_CART (gid\_old, ndim, dims, period, newrank)} and \mpifunc{MPI\_CREATE\_GROUP (gid\_old, gid\_cart, 0, newrank)}. % \vspace*{3.0ex} For graph structures the corresponding routine is \mpifunc{MPI\_GRAPH(gid\_old, gid\_graph, nodes, index, nglist, reorder) } with \begin{description} \item[IN gid\_old] group id of parent group, \item[OUT gid\_graph] group id of graph group (if calling process is part of it, else id of rest group), \item[IN nodes, index(*), nglist(*)] definition of graph structure, (for details see the definition of MPI\_MAP\_GRAPH) \item[IN reorder] ranking may be reordered or not. \end{description} % \vspace*{3.0ex} In the context subcommittee a mechanism has been proposed for cacheing information which is associated with a process group. If that feature will be part of MPI, it can be used by the high--level topology functions to store the following information: \begin{enumerate} \item Type of topology (cartesian/graph) \item For a cartesian topology: \begin{itemize} \item ndim (number of dimensions), \item dims(*) (numbers of processes per coordinate direction), \item period(*) (periodicity information), \item own\_position(*) (own position in grid, could also be computed from rank and dims(*)) \end{itemize} \end{enumerate} For a graph structure no information has to be stored, because the graph node number is the rank in the graph group, and the size of the graph is the number of processes in that group. In case the cacheing mechanism does not become part of MPI, there are two other possibilities: since the storage and retrieval of the above information is only done by the topology functions, it can internally be stored with other group information in an implementation--dependent way, which each vendor can do as is most convenient. As an alternative, the information could be stored in an opaque object, a handle to which would have to be included in the argument lists of all related topology functions. However, this alternative would be much less elegant. \subsection{Topology inquiry functions} If a topology has been defined with one of the above high--level functions, the topology information can be looked up using inquiry functions. They all are local calls, and do not require any communication. In the following the assumption is made that the information cacheing is available. Otherwise the handle to the opaque information object has to be added to the argument lists (or replace the group identifier argument). The function MPI\_INQMAP returns the type of mapping which is assigned to a group. \mpifunc{map = MPI\_INQMAP(gid).} The return value is one of the following: \begin{itemize} \item MPI\_CARTTOP: cartesian topology, \item MPI\_GRAPHTOP: graph topology, \item MPI\_NOTOP: no topology. \end{itemize} For a process group with cartesian ordering which has been created by MPI\_CART, MPI\_INQCART looks up the following info: \mpifunc{dim = MPI\_INQCART(gid, maxdim, size, period, pos)} with \begin{description} \item[IN gid] group id of group with cartesian structure, \item[IN maxdim] length of vectors size(*), period(*), and pos(*) in the calling program, \item[OUT size(*)] number of processes for each coordinate direction, \item[OUT period(*)] periodicity yes/no for each coordinate direction, \item[OUT pos(*)] position of calling process in cartesian structure. \end{description} The return value is the dimension of the cartesian structure. For a process group with cartesian structure, the function MPI\_INQPID translates the logical process coordinates to process ids as they are used by the point--to--point routines: \mpifunc{pid = MPI\_INQPID(gid, location)} with \begin{description} \item[IN gid] group id of group with cartesian structure, \item[IN location] coordinate vector of process. \end{description} In a sense, the inverse operation to MPI\_INQPID is provided by the function: \mpifunc{MPI\_INQLOC(gid, rank, location, maxdim)} with \begin{description} \item[IN gid] group id of group with cartesian structure, \item[IN rank] rank of a process within group {\tt gid}, \item[OUT location(*)] coordinate vector of process, \item[IN maxdim] length of vector {\tt location} in the calling program (only for error checking). \end{description} MPI\_INQLOC translates a rank in the group into a topology location. \discuss{Depending on the decisions taken in the PT2PT subcommittee, MPI\_INQPID might be changed to return the rank in the group rather than a pid.} \subsection{Partitioning of cartesian structures} If a group has been created with MPI\_CART, the function MPI\_PARTC can be used to partition the group into subgroups which form hyperplanes of the original process structure. As an example, assume that MPI\_CART has defined a two--dimensional grid structure on a process group. With a call to MPI\_PARTC this group can then be split up into disjoint subgroups which represent rows or columns of the original group. The function syntax is: \mpifunc{MPI\_PARTC(gid, lostdim)} with \begin{description} \item[IN gid] group id of group with cartesian structure, \item[IN lostdim] direction perpendicular to resulting hyperplanes. \end{description} The function returns the group identifier of the hyperplane group in which the calling process is contained. MPI\_PARTC uses the function MPI\_CREATE\_GROUP to partition the parent group, using the process coordinate in direction {\tt lostdim} as the key value. After a group has been partitioned by MPI\_PARTC, the inquiry functions MPI\_INQCART and MPI\_INQPID consistently refer to the grid sections if applied to a subgroup. If the process topology is a cartesian structure, a collective shift operation most likely is used along a coordinate direction. The function MPI\_SHIFTGR partitions a cartesian group into one--dimensional subgroups, which are all oriented in one coordinate direction. Every process is contained in exactly one subgroup: \mpifunc{MPI\_SHIFTGR(gid, direction)} with \begin{description} \item[IN gid] group id of group with cartesian structure, \item[IN direction] coordinate direction of one--dimensional subgroups. \end{description} MPI\_SHIFTGR uses the function MPI\_CREATE\_GROUP to partition the parent group. The return value is the group identifier of the one--dimensional subgroup to which the calling process belongs. \discuss{Instead of creating one--dimensional subgroups and then use the standard collective communication (circular or end--off) shift functions, one could also define a topology based shift function. It would look very much like the standard shift, with an additional argument defining the coordinate direction for the shift. Depending on the periodicity information of the cartesian structure, the shift function could choose the circular or end--off version automatically.} In a portable program the number of available processors usually is not known at compile time. In the case of a cartesian topology the function MPI\_MAKDIM helps in setting the numbers of processes per coodinate direction, depending on the number of processes in the group and optional constraints which can be specified by the user: \mpifunc{MPI\_MAKDIM(gid, ndim, dims)} with \begin{description} \item[IN gid] group id of group which is to be structured, \item[IN ndim] number of cartesian dimensions, \item[IN/OUT dims] number of processes in each coordinate direction. \end{description} The processes of the group with ID number {\tt gid} are divided as evenly as possible among {\tt ndim} dimensions, subject to the constraint that all processes in the group are used. The caller may further constrain the operation of this routine by specifying elements of array {\tt dims}. If {\tt dims(i)} is set to a positive number, the routine will fix the number of processes in coordinate direction {\tt i} to {\tt dims(i)}. Processes will then be distributed among directions where {\tt dims(i)} is set to 0, subject to the above constraints. Note that an error will occur if {\tt dims(i)} is positive and does not evenly divide the number of processes in the group. Negative input values of {\tt dims(i)} will also cause an error. All elements of {\tt dims(*)} must be set to 0 for completely automatic process distribution. When the function returns, {\tt dims(i)} will be set to the number of processes in coordinate direction {\tt i}. For {\tt dims(i)} not set by the user, {\tt dims(i)} will be ordered (MAXIMUM, ..., MINIMUM). Array {\tt dims(*)} is suitable for use as input to routine MPI\_CART. % \section{A simple example} % The example in figure \ref{poisson} shows how the grid definition and inquiry functions can be used in an application program. A partial differential equation, for example the Poisson equation, is to be solved on a rectangular domain. First the processes organize themselves in a two--dimensional structure. Each process then inquires for the PID numbers of its neighbors in the four directions. The numerical problem is solved by an iterative method, the details of which are hidden in the subroutine {\tt relax}. In each relaxation step each process computes new values for the solution grid function at all points owned by the process. Then the values at interprocess boundaries have to be exchanged with neighboring processes. For example, the exchange subroutine might contain a call like MPI\_SENDC(...,neigh\_pid(1),...) to send updated values to the left--hand neighbor (i-1,j). Note that there is only local communication in this exchange operation. \begin{figure} \begin{tabbing} =====\===\======\kill % \>integer ndim, mx\_neigh \\ \>parameter (ndim=2, mx\_neigh=4) \\ \>integer gid, gid\_cart, dims(ndim), neigh\_def(2), neigh\_pid(mx\_neigh) \\ \>integer own\_position(ndim), period(ndim), i, j, num\_neigh, ret, rank \\ \>real$\ast$8 u(0:101,0:101), f(0:101,0:101) \\ \>data dims / ndim $\ast$ 0 / \\ C \hspace{5mm} Create a grid structure in ALL group and inquire about own position\\ \>gid = ALL \\ C \hspace{5mm} Set process grid size and periodicity \\ \>ret = MPI\_MAKDIM(gid, ndim, dims) \\ \>if (ret .lt. 0) call error\_handler(ret) \\ \>period(1) = MPI\_PER \\ \>period(2) = MPI\_PER \\ \>rank = MPI\_CART (gid, gid\_cart, ndim, dims, period) \\ \>ret = MPI\_INQCART (gid\_cart, ndim, dims, period, own\_position) \\ C \hspace{5mm} Look up the PID numbers for the neighbors. Own process coordinates are (i,j). \\ C \hspace{5mm} Neighbors are (i--1,j), (i+1,j), (i,j--1), (i,j+1) \\ \>i = own\_position(1) \\ \>j = own\_position(2) \\ \>neigh\_def(1) = i--1 \\ \>neigh\_def(2) = j \\ \>neigh\_pid(1) = MPI\_INQPID (gid, neigh\_def) \\ \>neigh\_def(1) = i+1 \\ \>neigh\_def(2) = j \\ \>neigh\_pid(2) = MPI\_INQPID (gid, neigh\_def) \\ \>neigh\_def(1) = i \\ \>neigh\_def(2) = j--1 \\ \>neigh\_pid(3) = MPI\_INQPID (gid, neigh\_def) \\ \>neigh\_def(1) = i \\ \>neigh\_def(2) = j+1 \\ \>neigh\_pid(4) = MPI\_INQPID (gid, neigh\_def) \\ \>num\_neigh = mx\_neigh \\ C \hspace{5mm} Initialize the grid functions and start the iteration \\ \>call init (u, f) \\ \>do 10 it=1,100 \\ \>\>call relax (u, f) \\ C \hspace{5mm} Exchange data with neighbor processes \\ \>\>call exchange (u, neigh\_pid) \\ 10 \>continue \\ \>call output (u) \\ \>end \\ \end{tabbing} \caption{Set--up of process structure for two--dimensional parallel Poisson solver.} \label{poisson} \end{figure} \end{document} From owner-mpi-ptop@CS.UTK.EDU Wed May 5 05:30:36 1993 Received: from CS.UTK.EDU by surfer.EPM.ORNL.GOV (5.61/1.34) id AA18716; Wed, 5 May 93 05:30:36 -0400 Received: from localhost by CS.UTK.EDU with SMTP (5.61+IDA+UTK-930125/2.8s-UTK) id AA01134; Wed, 5 May 93 05:30:18 -0400 X-Resent-To: mpi-ptop@CS.UTK.EDU ; Wed, 5 May 1993 05:30:16 EDT Errors-To: owner-mpi-ptop@CS.UTK.EDU Received: from gmdzi.gmd.de by CS.UTK.EDU with SMTP (5.61+IDA+UTK-930125/2.8s-UTK) id AA01101; Wed, 5 May 93 05:30:07 -0400 Received: from f1neuman.gmd.de (f1neuman) by gmdzi.gmd.de with SMTP id AA22153 (5.65c/IDA-1.4.4 for ); Wed, 5 May 1993 11:29:18 +0200 Received: by f1neuman.gmd.de id AA16964; Wed, 5 May 1993 11:29:34 GMT Date: Wed, 5 May 1993 11:29:34 GMT From: Rolf.Hempel@gmd.de Message-Id: <9305051129.AA16964@f1neuman.gmd.de> To: mpi-ptop@cs.utk.edu Subject: PostScript Cc: gmap10@f1neuman.gmd.de Here is the PostScript file of the newest version of the PTOP draft. I first compressed and then uuencoded it. Please send me your comments as soon as possible. Thanks! Rolf ------------------------ PostScript below ------------------------------ begin 664 PTOP.ps.Z M'YV0)4)`F=(B")DW8LJTD.$"AH(2)8;(*1.&SALY.D"0L9,&SAP0-5S0R`$# MQ)`WBP*9,1"I4G4+[`J/$EQPP7&],XA1+F3)DY&6/$@.E4R)LZ;GJZ.>,6 M3\:2)6ODL`$"!PT93HO$/=FF31DW=.8\)&(ER4#";<+$99+8-'#V"L(I5 M*U>O8#FV+7.F,A0Y;\9,*4,G(YTR>%S`0?TPA((75,I@(>+SI8P:)7N.>4FF MC!D0N7?W!I&PM!L0+YSL*6ZF#W7H0O:(J4Q&H_$^3J!/V0-[#!KPT+'LF0+" M21\%0J!3D;*'CAS)<]A4+(/^19K\'YD!U!Q%A?>"'7.DH4=1:H%P0TYMU!&4 M@4"LY-P>_[$QQQY;E-3"6FOA!0(,72@PQAMN_$1''VD<]Z!.9[W!1AUTI(&B M1FG8X6!.5D@1XXPUWMB3CI:=`<()A8AQP3(1:9 M?6G@H5$=<`Q:*`T@F/42:G!UAVAY:(``!XV&5EI#HJR!P&A<(($`J:0T*D`@ M'7_BV1\0^L6%9!APE+&'?P""8%\=!?81WPM`1.9&'4"944899.PQ!QT5U3&' M<"^]H"NO;/@*K`*S%C7IBK<"<2(<:9SUZ@C79OL1%K9"9T03%6T1XH@@?(@N MB08:(800':X[8A<&NO%<24[&&2T0FP8;K]K;%X[#M0>" MPLU5)BZ*=%"11ZL@S.#P"T903"Z@>9J!F!R"MDOQNV\(ZFY\X0V[TES0B1$& M@2!@`0)3]^6!753D:F;S"T+4<3`90Z#!U!X*%,U4T$.7(4=_@ITHEY'[.GQ8 M=V00^JH9;[QQ)&M<(]9'3CB'H?.U9]M1,@ALO,$3NO8J,"UT__:[1;@O#/SJ M',>M9:#(]JG\L<,#%QS&P7@/O*&6+_!=,W2`D[R%XPZ5Y'B1Z]*K$<$&LX%W M$7O`@=+#SV5=*'65&4?QUW2$32W=:+1P1QIDT('&'N:U0$9%82A@.MN'G6%[ MIW/4(0:FK[^0.QIEL.02[K'O3FRE4\TU/*+%'Z\HWKGC@909>\0@`PXF1L\[ M]<$/KW'VR!]I//>QY_$]]+J?_WOUPD>:$_N*PB3##:)ZWZUR1P8\T$]Z84"? M]2*U%OZQ!GXM2$-DS')`^Q%*5A#Z;B:Y0)`08ODP8 M/4&5I'O?8V,+Y&>&XRRO>2VA0P"/!\$BUSK>M*:VF!"];21T,:CY!QW.3-4M@%^DFP+/PYI5DB,X>FL"[2 MH6=@S!1[ZL8>Q\5``<#\W?W2QT#W'>]W5XQ+%L68*<=I:5K#G%L:E0DP%*KP M/;=*`NYJZ$X0$`%O8D#)'HSE-&3%#"4M0$/77+D&-[SA#F[0#D`%^H8U#/,% M4TC"D9)8(!!4B"6548"(VO`&)5H$.E:P$\A,N*<^D>I.@:K46OKWNSE]JB0L MO>"A6N]I0,)UJ6$S)Z57]RS:#.>WLH*>\#;3SE;K&,M)V+/><%LWI"U MAMV!>1,)'4KDAIKZ$N>"^!O>#3"UAY(,6%&5>8T<[``40Z&`-W.`@W[RD((R MQ,%3>#`/NA"%8,0XC<%L4``*G*`;*E3XPA?IZ1E!(*"A@DFH!.IIBV,*]''0?ZQD.N$9`4H^4A&)C*+!]3DCT%9RBR:8)5G=N4LQY(_ M+[;QK7;:98I"ZSYNX-M%VJ"IMW#J4YOJSJ>B`FA!RX'0'/6HUY",+CFP(74? M+%@0<$R+)^*MS#L(=*E3IQ0IW?=SDV'4+&O">B3EG5TG<)"=P^>P"Y7L`>ZA!8AJ0N*>K3%IPP<'LK?>#-@2=CSZ([&^E9Y<#<(NT<(6I7Y>DB^ M:EG&YVJ!B:G!#'``@YOPA;"&G4'6#0L$ MM-957`F,`0#%MY:!\:"P-R@L#(Q0V!@4H;`S&`+:J;[VMK\=!C&0.PRF#G>Z MIQWM:O\[#`1_=Q@4_O"5,[S:$5_8P@O^\8%7.]_W3O>\VUWM;)\[W/4>@\S7 M'?`P&((,:*"`&]S$"#'`@0]H4!(B\`"ZA?5\:S=_^=@'GNZ3MSS>]:[XQB_^ M]Z`?O-I+'_G@0S[XC.]][W6O^\G/_O-NI_W@Z>YYV%=.]#2XP1!R@'K5)^'L MNF\[[W>O=O(7=@AY3_W:\SY[(^1@"$:@^@V,0/^VTY_W\R]__C6*]_NCG?Z9 MYW[P1W6H-P2SAW[H5WGC%W[F)WYJIQ8QD'U$D`,WD'H^\'V=5W@S8`0*$'__ M9P1ZAUMWIW7*%WPF6'\?:'\HF']_-WM_=W`,3.`,62('V!'Y&X'@HR(,>.`-O-WM\-P/\)W@F6()86'B^ M5WY:B(56*'SDQW>6!X5[)W=.J`!^AX+]QWDR,(1%>(0YD(3V1W=[YX%S6'Z3 MUW=X2(>4MX>5`X631W]"0'5.2'5QUWF&2(#RIWE_QW1&(G$ M-WU@*(6*6(:(V'\`&'L]:(A%L(%X5P/L5P0Y(`06^'W]IW\@J((C^'8[N'8`*'==^(5_-X:$*'=D2(5F"(!IZ'F=QX9NZ!6K"'Y` MZ(/*Z(OTYUI`.(KD%XF-N(B41X.VV(?#"(QV%WV>R'G7"'>^(K66'D"V72=-W_0J'H*0%@:XWI] M&'CCIWWZ-WYY6'X6>9$8^7^\YW\:.7P-^)'NN(`B67XA29(C^8`GZ8"%I5'A MIX.@F(XNB7=J003:]WX62%B7XGKS5WCS&'A+:)%=&)0>F85`691<:)3`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`RT'I"(2['(S[^ M4W9F0)EKAP,@"((WP'J`"@.L1W6NA77N2 MUZ50>H7"B)11*90H:9*8=TUZ-W7O!WC/]W9B:(R&B*K$B8DMZ(VU^'F?N86@ M:7N&MY!UEW=QNH.B)WQQ%Z=I5P-#P'I:AP/!:GNSUW2^BHK=YP-8)X>#9X:[ MFG:$]79J889QQY)\]S]49X")^:RT^GG@*GO%)W:%)0,D2`.IQW<225A#V)U- M-YEVIW55F)5T)XO%&8ZORJ"S"JNT2JW4:H;_BG78J2Y(56J$KV[)2.[6)Y[3E9P,D6%@S2X[-%0*/)EXMSJ9+-AWN0&Y\W^WA85P1Q:GB7:[F-=[D;NZ-Y&`3)RGV& M5[(XD(31-WOFFGIR9ZY!``/0-;/3*:CJ*KN+&)X+>KL.X:"LJ7L!JWB\2WEM MB*KH&KO#*WP7NWNZ%)\WF!ZX,$"'HU*(-PAW>59[L0^8#;^8,C"Z_>FGH\:Z@V MH)ZN*WS0I:B4BJ"&^(H9:'^!UX*99P,+?'K)?`N\)3Q[.-FGK4IW9$P'[NIXI)RXJ' MJ'=#8`,R$*<,F7V&.G5LIP`NS'8N_+IQ2[97F+:;>JBN:WFL5X""NGYR1W6G M*0-5+*@T@+5]%[AXAZ[0:X1!P*QQ:D]SRC$U)`-Q*#Z(8G93IWU+&,>+"IN+ M>`.,^K!)JP!;[*PHJGEW.)`O28=7VG^"W*6#?,B&;(94]F96^NGT%N'IQZ)`>^Y*;-\%I1X-1.)UH:JJO M7(=\"LNS3,CK&G]3V*4;.*2>!Y"!IX76JU&)C(=:6)%?2*W%"+]]J'[#"(D: M^HE.F+$NN9B?J(-HZI+(.G@3.,JL>)D^N9F#YZ\`N9[2[%JX''^(N,N(R)BR MC(.D6*G$/)&6?,`J.+8+V,N#O';C!X(*X,6S^9F8&8`!'7L#?9D?NI5K>*SH M?)4OB8ZY^"#AS'UYA\'=#)7F&)SQ_,N^*:$<+7G_B'8?':*,^]&I!T`C&,ZD MZIIW-YW(C-+JZIPH#8.R:)76['C`#,RGJ7\WK=,\_9*?.8]B%[K)%,[9.6WO=>8 M/;::#=J??=FEW=:5+=J9O=IM9Z8/Z-JM_=>P+8Z>-W5Y77Q,K8N,NLFH/#"7L,KGXCZ(&6-^')'*IV M*'<83G=R=[X)'GH@_N$4::F5RHOJ8+GZ:WK"9U MF'H+:0/V:XI"X`-(.*GZS9^Q9]2M..4`^8^DN*ZM[)O[G.A/2>:8I^CSJ]&0 MCM$2O,"Q1XYG38<_^<4Q8.>H:(HYH`!Z7LI,F^F7_N>&7N5C#J'^B+%\RNJ; MR:=%;GR_.C7G=&IEYTD*-TY6.M-:(J'N[D/N[FKN`TFNWB MKN#ESN[F?I;+6'E[.P/S![IJ`;B<*8M&7IAHRJ=D",>!&[@2^>_&>.@$OW<& M;ZK2B?#YV^_6F94=&*HLSN(R6HW1O-_13,>5WK&8^8`=;W?,V/%>O:BX9X`A M")M?3H"_R8.MW89W/M'?IX:3?9'929*][-6C7?./3/,YW_,\__,6J?,(;-*/ M#.?V-[?H**SH*(IW2*0;"H^2SNM0S_#]QW].'_5/C\!4+_5:?_5;G_6M^/62 M;N33^(I`[7\Q`+KTOMAM:`23V7S4&-8B^LX,2+46Z>(S[XM:GYU\']U[[_=A MW_=_/_A53_@]^9M",'K&;80W3-2'2?B"'_B`'_F=A_.0/_F8?_F:+_F;S\YA M?_B9GOAK;^])^Y5)N*1D:?3]A_3VI_1-#X]6[/6R/_6SW_6T?_NVG_N%+_:\ M7_MA;YEG;_9+R(-FR-C#;HINKQ8.P>=4R>ID:(X9.^T0"NTG/_WP%X*8B?U3 MG?W7+9"EQ_WAO/W6;_W`;\Z?W[4P.*&9?IO'#_..[\TK27]*39"K_OJ;WXE] M3XB>9^:-'O9IOO_GJ^;Q/ULG`-$<`8QN^N\.=2*Z]XJ(;@6YG,@T_@>:QW,[&^D=N!Z&YG?#DK\@@(0M< M.B\-:CD+V`0KH!KL;&K0";;!`V6-Z->F>WE+CA7)OU^GOBP4Z/&`).BDH:\; M^'^H$N[Y376PGGV[$+2M;E$?`H3K"PR]HM#$D0P0,U(+[8]DG:AVQ`E)TMN2 M3S3J^?@QH4-VU.W67[K3=NLMV*`J=41\D]>=B(2W<4R<0-E4J M+11_O)KX0%`231.^/WI6?^@9U_M]N`\9ZCZI-P-]WS%\AF`/&F*]:-X=)A&"S_/+V\I M,P26M0(=EYH)]]#Z/*@AL+*T3S^L=3_NQ,E!-NB4-%4,4$5VSOW4`(.E!6=3 MDX)UB^HA]C8.Y^V.U2S,7Q=Q7?D=EH:\T,[Y(@*F2J.!1+GU=HZ@]8%"J8?M M6!^UQ*="W.KC4\3JD0F!P3.J,,],)$)TA_^8P)#%T]Z91O.):,<`2I[N-9^Z M$(KB/;_A#BI$(K:;RN%,PW+3SF&1NW6XE"`4_NE=#*^/Q<.GY,XV8#W<4_H0 M+-JS?L@/1=O*PG%"+CO!P=&&A:+2^<%(4:N/31^Y"((J4EWD4CNL:KFNFNB3 M3%P4O(((R105`3*V%I:;9`-NU&HF0IU]--'$!Q4B`F?,"#B*E8.G'%\O`F[% M[U`Q*M'G`UH/^(E?L6T>JI\$M(%2X@SX=*,*!7DQ?Y:Q;@+\28FD432"1ME( MP4H4NEIRC_'LL$9/9-3\U#SB3\(M.`I'_L2?E!KZ@4D;R@SA'(2%Z(`H(&#C$FI1;D@J[5^@%]QF@'+C?`]*&*4HN$C\HQP?E3ZHZ3LHGIRTGA5626Y MI*I\3;_-(9K)_&.0H$Z1Y)34RR$M(93G&_M4&?I7_TKPY"?]1XU*'NPI`I2P M\2RT7Z2,$-KZ^1\3J$.JA85HRL:6M9Q8V%)4[L?FQAZ56!);=-]R8F$=7F6X M1ICSD5\4BJN-.PADQ]S/PQJ4'(@<]J"9=L"FT'"REWT(7TZ=422TC``)B@G^ MLNVDHG])!(3`,"J8!Q,^'$R#^:GLEL/43N(#(4G+F&#&"*01.`Y[C(UMCF2' MICY@8MIV<6<#>;&T`XA"$"?ZD@P(*_JN\:,R6V8#0FLZ$4:MS*.4,C^2RZR9 M.)-EVDR\X^%\3\ZDF3KS9_X?,SDTGQS>86P\T29YI6LRJ8"B?LMA@&=D>B_? M]26;I%2"2I<1`'FXWL4US4_7_%U@DQUZS;$9-GVFV`159I.`@:>UBW*-VK$_AJ9.1JR/J32TE>*R/W4Q*)FPI!4XI*'/J&H M\,/!ZF;B#(TA#G0!R@F$M%;/\N,!O. MS),I[$[.QL+UOEHGE/J;;U%U(D[/21\EI_VJG-\G[?0[](2D=EW70EYI!XEM M!4>%$UF/8WM@R`[T/<7",A/Y#@X@"8,Q=^XI)R6_@`-K"5G&$U[=LHPU):%+ MW($N;F<'828")`%]$0%J7]\J99F?D6FXMM#C0C[(RQBIGQE`@A(B83ML1XU= M;:54]*Z`@[/Z7Q>)+0I0GN<72=P@PWDFCGZ>1?BH0"W2P^I"#]2!7K2WJ-EX MD46RH(C21FY.45B(*M6$8D^\!X0ZSN_E.2\:[#%':M$'"K]KYJ)T50R@20*H M\_B`U7+Z0-1^[&4^:"-YQ_V'>4)A?`Q$[1%<^;/M&/*^%?0KC^#JB%X>889$ MBV@2C54JB1EAK"!:VV*8'^.AM,H+[C95=PIEYD$\1:DH3LW0TC6I.A&F&WZ' M<.^<2[S3L#Z/,&M>F@<>WDA:-*0T:&RS0NH'GT5"+N4>]>*WJHX!-+^%JR7U M'<4CJZ)#T8E2A9_98[`&F.SQ.Q=QG,D$05K(K!3Z652H2/3$@,?&-"VC<;1& MGQ1]Z="I*4IW&W5,CJ9TE)[2E;1*22DJ;6Y!2OVYIU=:2FE5*SVE5D^5ZM)4 MRDMMZ2X%I55)_J$]QK8_BU8XM(R)$(N&J_BH3)LI,WVFWRKS,-%H"DV7*:VJ M9B@(`2&DF7A,]9OY]$7HR`Z](DS6_S29HQI!HG`D[IUTBD[I$/QQ6*+PG;JN M>+K`XFGIV5;Q]`#E4WSJ3@_1E%19_C2>!E1WRNP,Z26=6SDQ]!34>+IV.)8[ M;:@&%:(R5(DT4:?30Z6HE2=AI2=01,?@'3A]8^FH#8'1Q$>R[%@-Y:C?E"-1 M.)[D#F^@WM$HZI3WR%.9FGEDZ@%263?U_`0NF1I]>*J;:UIL1Z;6G2%`HRRJ M4/6'32ON"-4PV+28:NRL=1A+J-+`$$<=7^J;[&#@U/Q4PA`X4F5H-T.+B)#A M=:E$NH70U@YU3YI(/HE"@3>D8E>GPF?DZ_+`U;J#*./J]YJK>-5JK56+VA[W M*BDEJYBHD5)3$KK@0FAUFFQ=:,5]44SZ?L2H.$2.^BG^V=)(F+[DJF6]JY)N(JOU218[MJY*XL$J& MQNH-2"EL>\.ERM%G&]K):UKEXJZ-16#6G`JODZ6M77;KNG M6_2_&E(#2ZM$(8/]/*VEUW6Q.O@]ZR`]JUKK"OW0)*[D52F36B-P?#5CN55= M!4!K:N^955/T?W6XSP-=4FR+I57Z<(`&TAE[6FLL:JU:,M;&IM8=*UI[[&C= ML8.-J'K0@@1;MVD6;'++[2E&,R5+?_@>#!*NYDC"^K)>9_4VE+=[9ZUJGE*= M"[5E9P^7U;(`-'`YR%"EX78J[8%-&NXYN3E_6HZPT>($0FXVH^XZ.&M=S5D] M#)AKY\[VNGIX`V9B0^ULSVZGP4`(6;:4]1VO-PLLF7-R_((VZ\Y13N35CVVQC;9 MGLWN1;`"8)G".^2IM4$H(01IX2:^`T78*&12.]R"+IWI8'4(CU3S8"VY$T%= M%--[EA)3?.`O1/8);]3;HE$Y"F[%6X07B-Q%\WI"&C2M=B/T15FM6&?EMP#W MW^[;M*JN$-:5*DO7S6?J3!WV@XU>T%&[N-547(G;P%5N MORGPBYBP5GR`NFZFT(S1E*I`R,G2$T^G[;3K M)VG*G4=5T9S/'`-%7RID$H'S-7$9H3D203<(!F&COM-\T.RO\VAC*ENZSE\G M9*W38'-5*[+"X:*M2W[V;67U1$RH\@0OA-2&%J(6K+6Z%M>JW;3+MJ;MT&U" M_NCRB*&$18S"4#([5U-F9.I0=8CI#U_"&GK.+Z&XMVV64GW=2NE8#9@FA%9%* M0.(G:^6=E5@6>VU'BU@HZ?E$7`7T)HFJSNI@AXG[S#3=VVI?J.$UDH1EP-#: M&QAY*>]++;ZH,_DBW^4[>9,O_;F2^@WS:M[82M006D@S`KZ6/!&DKC6(F/6WMZXI@MUIP=RL8Z:MZ/6\H->;>E,A!"A# M+@!QN+'0_1C<)\1779`&Y5`:*@MW8"V,A<51/2K!B71\'5Q[FW#;;8U*4?/6 M$PT?EL>$T^V@'(G0S`B((!&4O9HO\TV=*"RKOL/* M5Q=&W.M%BD>QVX$]4XW#I3V$Q'3@ZW&$OL?1`??=!2R+:3$"GL4-.`Q=2>9* M0:E/#PI."K<2-UQQZ,V"6\`Z5L;H]_Q-NZD.(P_^<8D5*(1.4G7$RPK90+R: MRJ=(M>`ZUQ!CTT\2IO`G8#U;7=01TT_&6CP'$8`:GH.(#QL/.[:;NLDT#1ZM MA'_H<>!IL8;5W=56>6C(W.J]ZE2"YXN^6@50)*.GAX5U!.Q#B]MPK>R(6+33[206,]B M5$47S#%63'%A)$J76BA=9J>U]*GKYZ>"W-AJ=5.Y*C\Q:-F*ARZD(J.6$4@R MMP,858'@#^R!'6HC$:"-Q(OPCY"M=1()+C]`C2:7*:!XNHIV>2[GNKISM@MY9S(/8CRS?40^DWM>1:N=)O&*-RZ)0I'L;%ZZ3K7#2P47$5 MNL)V5SXG,OBZD>K')/L>#62'()!B*Y(7[*:XGBB[H5KS"Q2LC3>`/C[.1YLI M7^:3?#,P\BE/YNEV$!(1&%TXZ:0*/R.$Y:(8)4-G'79S1K=0<9AZ'N12I)I3B%B7_N/*OG6=@+ M35T,(Y5H4)P-4=E;D$Q1']Y-G2=D(53PN8_@%S+S/$;.3'DRWHB@3AZT@E;` M$G[ALM`CD(#0:58+?VM:)L_RA)#!4W6^T-0Y0TMGZ#R=-[1U5HL<&CIWHI0W M?1A8RFNRKJ4/YSFU('PCH_[@"\_3=^0IA,H[_X[U9%2LYU'ML>CXP

ZP1A M>7.$&1_N1L)HKI$FG*!X$Y<\:R2+>C-1S0$S;,F-A.K%C*A=N2*A.\H'7=,M M[0/=X-,ZM![)F&DJ$A>,KO*M>T\#;_&Z(%XD?]@T#1*NL\>*[JB[X\/FIB@, M:4)(#D^@*`VI/N35&+U4KMF"LZM+4`*M?R6I/'A6DM M*"5_O)HWPCI_YW+]G1U5@^(K$\/&4,MJJ:"<)H08*Y]ND__R-X52+I>#D*)% M&J_C=4>I'5FMM5R2V:[]TDCT\C0B-EV^_I"E9W??FUAL:*&:ET.L._Z?/@UW9'5SP=@ M4\3%VY[B44@Z[OI13K38$_N*KEM'L MJ[ES;ENZ]D@DEC2".$\LJCTH5'VF;.D394-5R99TJQ'1`A!$:Y;P-75J;LQH MALGAH^58I[495<8H2@P):7`+1PLK;AM2HU#*OB^=E[V*&=6[TO"1!@+:J&6] M<%V[.W=8&Q5F[>[(>C:JVK&C]]HGWU:L$WZ^:^'YKM8GA%+K3I58;]X+=7E" M@/NT(>WIL(GBU*QMC-"/L4_I(WN5[@#V8T=W`B91&Y0#+:!4,Z*$NQ5=-()5 M>/J9.,-H$?MQU^!V^%=OJ\>>FBW;3:[8YT,J-0]K1(0^;&>O(W-%M_4C#7P^ MJW%)ST.OC:)$(<4.VX;T=2C`?'9>VNSI^12-G6O'HVPFOT0=X_VZK M);S3-6'%.M01>5LI8ZVHW=/EHHX03JK*WKSMK8`;(=+9M[/#\LEK5(6H=^IF MV:^[=7.M42B[RS?YMMV=34-61]Q==[YBW/V`9MF=O-M/M2Y'>G_UFQ%7&.?K22.6T_K M!&WB@/2)2[8H#JRI(\E^3YD;B2K:EZW%*9(0GYK=<6_#VR^^MX-H=V3BJNZ, MGBD[U,,*N.62`=J3[%GM[)@=SS@1O^.%U#.;UAL+;G7L7,7?[!I8HT=TE[S!NI;CVV=9L-."86Z1D3J-H`#(O/,Y\F1>>ECBOI[EFDU]J MX5P/HR04,\RYWC*I^L M/UU?\\X[79L/U6FP[<3R#C_W8;\YL#9O",N'CB5Y-.@_UL2ZNYYT'(\<8]/E M,F",L2(`8N9J7;,K9KE(1F7T>>V8EGE'G]?':IF']$4'N_+UK$:#%LE*C]>E MN]+GM_$V8BF][JSTZ1.G$.7^4M];;W_48W'.^YEBKZ%#-OM(5RSP>[JJ,0BH9L MG5C*Q:>UT^"X;)^%)[5>M=$36U?KPZP](["X;HUU60IG>'[U3OM(2G5+3?:( MPJJ2")'^T4O+I2H4<1,[<=L(D.Z)7K=):7Q./!);/(9;=H>Q)?OKKMF#50#V MUY[Y8!,W($./H+V4AW923LH#EXNKY-,QM??O3+YXX?<"PMY'J_*5;H!(4<$T M"[?;M1Q\A=N:I]`4*8#.[6!/MW9V^IH=Z6"5LUIYB&"]U!]<'L,/XVZ/[2G@ M::""!]VG^].6[DZ,.FF4KHWFG'E5@I//3[%'-,NU29^.LR)%B2KT\&KOQ8#_ M,9U3QYHVOBL?34M:KQTP(W7S#TA%*)ZF7*VF:975;P?`3S'?(^#3-@"521/)/WCCOJHD5Y:%Y[MCN4O_*D5'7/[BSO MGK2\EX]/X]*VI7*O%9]@N=Z1U;H'S>L=/D2MK&58S&=OFR:N(R)VU9$IA1ID M'>I/CU9ZO:R[5/@.W_OZ*?7K1R>KO5BA7_-!M+GB[N=7OR]/\68]RA6GLQT3 MAWH"*/M>.^,5TP?0#\[:IP_>^T)7R,G;'U_EX]^X4+YT;'C_*5[>7N2GJ7C, M6I.]R;,[66^'%/@?L_7."]%[.XBGZS&1N#;IOAZE4ZJ2'EYG#[$OI%::5L=I M,B2KWW2[WO*I['L%\7?-OGM5!I:/K,?MC->XD^TUUZ$E1!V(5B,\ZB/N"V`& MQ-H=&SSF^%'JD>X67&SJQGM9*^IH5>K9'+X3?OC'!V'TO@,'!2$2S?291]Q_ M'A^'SLQC!//@8=&GX[UQ7ZW9(A/DX^LNXK>=JJVH^]E\+.D7Z9)B)`R^KD?; MY1IM6=K$??$&W]K'?$0JX9;L^H[#W]3&';E5OT"G25RD`1!``YRRN=J8H9'P MJ*0?K?/=)U:-V7Z,<^NXRNYMK:G1;Z:O6^CK;43KB2]:\NE8>3/\1%UE1MQH M$A%@;!'(M*EL_NO:B[/SZ;:5DT$_Z:'_N5*'+CJ^][GZ$J M;_;\Y1/W@\^*+2Q5GPAA?1[>S2J49DM;R,E>92-]ADDODS/'#$@CK2W0>\,;",>]VY#.@>9+7_^7F2AI2+U1W M;.VLG@'%^=>Z')M-&B_MV+$A@`_=WA;SY?#^:H+>-*RBO.4Q`TIP\`BC83/L M+3]UJ(Z4PQ]N^7Z*=?S7[KR]UMV2^5_0H+>WK+[FAU=.[:P-O>XEW=73H5-S M0AC-F;E9AXC&O_E/1N3?2:'_\T_(^%G[7__PWY:%?WN$YHZ1?([_VM^M&:0- M^WX6TN:/D"H/9I2832BQ60'H1QF`"&!PU$)]6,*-%=/#K!W-40^3`\Q0?AIX MXFFQDR`6`ONYQ8Q)031-=8#"D0P3I##P%AD5TMZI2R%1Z%' MDH13C7*`3M#3!.(A0HZ\DL\(']F?YY1>H6(GH!(XQ4$Y,D_0AL^T2D"1ZJ7+ M6#+J6/878WE]\(]^LILQ:0Y!Z@+1J1X,V[3VY#QOGPA`LW>,1#[13Q+2U%96 MRC0BR(PB-X_A`12%*)H-I^-`*8$,3'@$`A$T`90BZ`C>>(O@H-63S5.%X""( MM510^8Q@ILSL@88(:I,"8457ERN"CAPDE),,0++8.344R):8123IC_[A`0TR MUQ=>MB3M@=4-/+,D<3(2BM^1"U8.>-KV<]=(*(<97`,,WH*^H##XR'@XO6`P MJ,GT@KI@RD>:13H^44@C#39N=9DY]>SL/VH-SN:NM3VWT_@@E%$V$A)*P\H` M04<7XQ'27"A1UQZX))$@_XCWE.NX@RB-/D3TW#03E)_SS="#.\U.(P^F9&]1 M.]@/:B%"%IXF?V`G`^$&Q8X,6BX(-@0#D2+"S0"8@>UL!)FI]UB!@S4-H44. MRC3`QSDX"ZZ#2U(_\PYV1/X@.\@/:C3XH"L"`MF#UTT]B)QH-!L+3[,/JF.2 MX)(T"<:$-*$26!-Z0/6:&M>/N7%S6T4S!]HX]0U#@IE MR$PX1LD@V`KJ>8B@3`A`'389223("+I'2*"25-CQ0;+(A+/4W#15B*`5`S*% MD`Y01-*,52Q5I#,-VH+@EE#8?Q0SRD@P\HK`'FY@4*!E$<)`JB%K>!JVAK#A M:R@;\H*SH6LX`%9Z;ES6EQ>22=L2;SAH<5J-6Q^C#D8Z+^%'\M$8,W='6/31 M]$,M#7_W%:$T\XT?.,FA13_($#C).4M]D6F3?PR"1(^.]YFP@E?A:@@$F86P M3Z13#3(ZXV!E***X)(/-MM0>CAX\X01X`RPMIMM^1,0U;C.A3:@?AD?[H67# M'_Z'_J$?U1].<@$B@6@@]C0A3DBC#(8@KV`PJ'N0-,&@SR6Y9&9@D.5E*>E' M:8\;QY@T1&W;&I[8YG]+QK1C(@1"8E!8H]XC/@_WE-(I!;] M+S@(^-0C-HE(XI\C%*%B,9&0J,TD0)X33A0EUB$S4)@%^'B)3=`)-?5`,[!, MFI)-#3E/!VRWCNPM%(WC$R4E0HG094/+=2E^4ZWC-VF$#HM/Y#'TT1A.[XSAU(D8BG/A")4H>DY"D.$$D#B*B`BD"((XBI<)I5@=-G4; M"*5X8?4O&PV\1^H4AS!A>R*_%1YA$6BEYQEWL`<1(QIV@Z2A"G+ML%;(63,( M`PDD5DYIQJ'].4[3Z+71)$D_"*\8*MV#)N%]=Q*&52CA6:.%Y"?#XK#H*UXV M#M%&DRL*26#.]L7H8%D/UUGHX-5K-@\;UW_8A?_#-V@.R3IFV6E(+3XGBDEY M>!9R1(++N0@#J3+P"+K8B5B+"Z$/X@%E.NK-,D+J#(@(8GX((!Z(?F&!F"]* MA3%A?!5?N240'=_'M*@^,9,:@HC@@F89AJ/)+(RVXA.HSS"+$=&N>"S^(,FB ML*@Q!HL<([#H,7X@]]U'$Q8NBS@4((CH[#+:1S-2_DF'V%?7@F4A9[`/GF?< M,28F4%'BOL4MQEW^M]^Y(FV;MQ;"M8&'2GA2"L9-X&*])O.@-%5@:S@>PH!D M8:ZCGJ"%`=%Q"#6^.%)C3T,>RH058D3R^>5'Z16A$XR,.3V-I1+MW(>$T*S( M>Z2'M)T=".F2P!X,1%RP0B MZ%)K`?55454:@)8273L1"5RFJED\SLU^LEIY4Q9@.57R.01*U5`$P\%P^'!3,9."8/2@/MY#JK%Q"4%\%`8@D05`0$ M@S8`[6@[.HWJ&+ZH+_J+NR._"#`2/;YC3;@O_HO"(U#(D51XR8C!8K\0-P8C M-@/Z%$/I#4#8D\F"L9,D2#UJ%-0C-"8]RH398_-`CI#5VU[7ZZ!=`%$*V>+1=N`ARO>_O#/HWX.H,!8D M]8N[LN3H\D>D9JP.A2$!+Y+03WID_%N`\4G_L,BB:4/-"87_W(I!DI3UW4`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