Thoughts on system software for next-generation hardware !"#"$%"&'()*$ $ +,-"&#.-/$01)2&)3"$4"&5*.3.67$)*8$9.(:;<*6$=*2<#;#"$>049,?$ $$$$$$@-6.**"$A)<.*)3$B)C.-)#.-7$ $D"*,.-$E"33.F/$9.(:;#)<.*$=*2<#;#"$ $$$$$$G*,H"-2,#7$.I$95,&)6.$ $
What’s Happening in Exascale? (do we care?) Pete Beckman Argonne National Laboratory 2
!"#$%&'()*%+*,'(*-+,(.+/01+/$*21".+/$*13*4%5'* !(.31.6/+7(*8169"0+5*:99$%7/01+&* $ ;"%$)*/+*%+,(.+/01+/$*9$/+*31.* 711.)%+/0+5*.(&(/.7'*31.*,'(*+(<,* 5(+(./01+*19(+*&1".7(*&1=>/.(*31.* &7%(+0?7*'%5'@9(.31.6/+7(*7169"0+5* Pete Beckman Argonne National Laboratory 3
EU Announced Funding… Pete Beckman Argonne National Laboratory 4 J$
Three Exascale Platform Projects Started in Oct-2011 to Explore European Prototype Architectures K.)3L$M;(:2#)-#$"1)2&)3"$:3)N.-(2$I.-$0;-.:"$ ! ! O.,*#$I;*8,*6L$09$P$>2.("?$("(C"-$2#)#"2$ ! ! =(("8,)#"$,*H"2#("*#$(.8"2#Q$RSTU$#.#)3$)&-.22$T$7")-2$>RVWUX7")-?$ ! ! Y ! A1+,@;$/+7* !-.M"&# * >WJZ[U\$#.#)3?$ ] ! 0;-.:")*L$$@^U$>G_?/$D4U,&-.$>E-)*&"X=#)37?/$%GBB$>E-)*&"?$ ] ! P$-"2")-&5$#")(2$I-.($3)C2$X$;*,H"-2,<"2$ Y ! BCC! $!-.M"&#$>W`Z[U\$#.#)3?$ ] ! 0G$X$GDL$$0a4bBB>K"-()*?/$=*#"3$>GD?$ ] ! P$-"2")-&5$#")(2$I-.($3)C2$X$;*,H"-2,<"2$ Y ! 8DCEF: $!-.M"&#$>WVU\$#.#)3?$ ] ! c)(:,-$>K"-()*?/$9-)7$>G_?/$@33,*")$>G_?$ ] ! P$-"2")-&5$#")(2$I-.($3)C2$X$;*,H"-2,<"2$ 00D=$!3)*$-"d;"2#2$2,6*,e&)*#/$2;2#),*"8$,*H"2#("*#2$,*$V$.-$T$#-)&'2$I.-$VfWV$ ! ! Y ! [ffU\$g$WfffU\$.H"-$Wf$7")-2$ Pete Beckman Argonne National Laboratory 5 [$
Kobe Japan: Advanced Instituted for Computational Science Pete Beckman Argonne National Laboratory 6 S$
Japan: Current #1: The “K” Computer The heart of the K computer consists of 80,000 Fujitsu’s SPARC64 VIIIfx CPUs � 864 Cabinets 10PFlops 1PB � �������� 24 Boards / Cabinet � 2060 mm � An amazing accomplishment, with unique and advanced system software $ D":#$VfWWL$A"F$&5,:$)**.;*&"8$ Pete Beckman Argonne National Laboratory 7
Pete Beckman Argonne National Laboratory 8 `$
Pete Beckman Argonne National Laboratory 9 h$
Pete Beckman Argonne National Laboratory 10 W $
Pete Beckman Argonne National Laboratory 11 W $
New at Argonne: BLUE GENE/Q ! ! !"#$%&% %3;"$K"*"Xi$D72#"($ ] ! J`$-)&'2$ ] ! J`_$WZS$Kjk$*.8"2$ ] ! lS`_$&.-"2$m$l`S4%$^@U$ ] ! T`J$=Xb$*.8"2$ ] ! !")'L$Wf!E$ ! ! D#.-)6"$ Y ! nT[$!%$&):)&,#7/$VJfK%X2$C)*8F,8#5$>K!ED?$ Y ! +,2'$2#.-)6"$;:6-)8"$:3)**"8$,*$VfW[$ ] ! +.;C3"$&):)&,#7$)*8$C)*8F,8#5$ ! ! A"F$c,2;)3,k)<.*$D72#"(2 $ Y ! =*,<)3$272#"($,*$VfWV$ Y ! @8H)*&"8$H,2;)3,k)<.*$272#"($,*$VfWJ$ ] ! D#)#"g.Ig#5"g)-#$2"-H"-$&3;2#"-$$ F,#5$3)#"2#$K!G$)&&"3"-)#.-2$ ] ! !-.H,2,.*"8$F,#5$#5"$C"2#$)H),3)C3"$:)-)33"3$)*)372,2$)*8$H,2;)3,k)<.*$2.oF)-"$ $ Pete Beckman Argonne National Laboratory 12
BG/Q installed and running! A GREEN Solution: Co-Designed with IBM Pete Beckman Argonne National Laboratory 13 W $
USA: Exascale RFI: Deep NDAs with Companies to Explore Computing Technology for 2020 Pete Beckman Argonne National Laboratory 14
What Did We Learn? Maybe the Obvious… CPUs are Changing… ! ! !/./$$($%&6 $F,#5,*$)$*.8"$,2$8-)()<&)337$,*&-")2,*6$ Y ! D72#"($2.oF)-"$F,33$&5)*6"$ ! ! BG+/6%7*91>(.*6/+/5(6(+, $,2$&-,<&)3$#.$:"-I.-()*&"$ Y ! D72#"($2.oF)-"$F,33$&5)*6"$ ! ! B%&,.%#",()*6(61.G L$&)&5"$&.5"-"*&"$*.#$:.F"-$"p&,"*#$ Y ! D72#"($2.oF)-"$F,33$&5)*6"$ ! ! B((9*6(61.G*'%(./.7'%(& L$T+$3.&)3$^@U$)*8$Ac^@U $ Y ! D72#"($2.oF)-"$F,33$&5)*6"$ ! ! H/"$,& $()7$,*&-")2"$ Y ! D72#"($2.oF)-"$F,33$&5)*6"$ '()*+,%-+$./%.+,01)2,%3)--)45% Pete Beckman Argonne National Laboratory 15
Parallelism Pete Beckman Argonne National Laboratory 16
Parallelism Has Suddenly Exploded “The core is the new transistor” (new Moore’s law) R*T(/.&*13*;$%&&* ^)2:C"--7$!,L$RV[$ ! ! lffUjk$@^UWW$ ! ! RV[$ JIIKIII* SIKIII* RIKIII* QIKIII* D.;-&"L$+@^!@$01)2&)3"$^":.-#$ PIKIII* OIKIII* NIKIII* MIKIII* LIKIII* JIKIII* I* Pete Beckman Argonne National Laboratory 17
With Intranode Parallelism Exploding, How Do We Write Programs? Pete Beckman Argonne National Laboratory 18
In-Socket Parallel Programming is a Mess: EG&,(6*E1=>/.(*8'/$$(+5(&U* q"$8.$*.#$7"#$5)H"$)$6..8$,*g2.&'"#$ ! ! :)-)33"3$:-.6-)((,*6$(.8"3$ A"F$!-.6-)((,*6$U.8"32$m$B)*6;)6"2$ ! ! A""8"8$>b:"*U!$,2$)$("22?$ U"(.-7$(6(#$I.-$8"":"-$5,"-)-&5,"2$>T+$ ! ! 2&-)#&5:)8/$&)&5"/$("(.-7?$ bD$#5)#$&.*#-.32$#5-")82/$#)2'2/$)*8$:.F"-$ ! ! j.F$8.$F"$-":-"2"*#$5"#"-.6"*".;2$jqr$ ! ! Pete Beckman Argonne National Laboratory 19
Rethinking the sequential abstract machine…. Pete Beckman Argonne National Laboratory 20 V $
Rethinking the parallel abstract machine…. V* V* V* Pete Beckman Argonne National Laboratory 21
Returning to our Roots: Graphs JSSS $ Pete Beckman Argonne National Laboratory 22
Reinventing Programming Models? A(61.G* A(61.G* !* !* !* !* !* !* !* !* ! ! =*$#5,2$*"F$F.-38/$F"$(;2#$-",*H"*#$.;-$)C2#-)&#$()&5,*"$ Y ! !-.6-)(("-2$5)H"$I.&;2"8$.*$s&.-"2t/$8,H,8,*6$F.-'$)&-.22$&.-"2$ q"$&)*u#$:-.6-)($#.$)*$"1:.*"*<)337$&5)*6,*6$&.(:.*"*#ZZZ$>*;($&.-"2?$ ! ! Y ! b*37$#-""2$5)*83"$"1:.*"*<)337$6-.F,*6$-"2.;-&"2v$ q"$(;2#$-"#;-*$#.$5,65"-g3"H"3$(.8"32$ ! ! Y ! 9.5"-"*&"$8.(),*2/$2")$.I$@BG2$$ !-.6-)((,*6$(.8"3$&)**.#$C"$C)2"8$.*$:)-)33"3,2($)o"-$#5"$I)&#$>.:"*U!?$ ! ! Y ! 95)-(PP/$9=B_r$9.*&;--"*#$9.33"&<.*2r$$E;*&<.*)3$!-.6-)((,*6r$ D72#"($D.oF)-"$95)33"*6"L$ ! ! Y ! 01:3.-"$*"F$)C2#-)&#$()&5,*"$)*8$:-.6-)((,*6$3)*6;)6"2/$)*8$-;*g<("$272#"(2$ Pete Beckman Argonne National Laboratory 23
Intranode Power Constraints and Cache Coherence Pete Beckman Argonne National Laboratory 24
Within the Node, What Else is Changing? How Will System Software Manage CPUs? How Will They Be Programmed? !1>(.*81+&,./%+()*A(61.G*81+&%&,(+7G* $ :)/90[/U* $ -+,($U*W+%5',X&*H(..G $ ;1._/.*`*8'%(+ $ a $ -+,($U*E88 $ ^JR $ -;AU*;Y\] $ F%$(./U*YZ $ B/$$GU*C7'($1+ $ Pete Beckman Argonne National Laboratory 25
Power, Parallelism, Coherence, Fault, Storage EG&,(6*E1=>/.(*8'/$$(+5(&U* Y ! !.F"-$(;2#$C"$)$()*)6"8$-"2.;-&"$ ] ! +)-'$D,3,&.*L$U.-"$I;*&<.*)3$;*,#2$#5)*$&)*$-;*$)#$I;33$2:""8$ ] ! c)-,)C3"$2:""8$2;C&.(:.*"*#2$ ] ! A"FL$b:<(,k"$:"-I$$I.-$45"-()3$+"2,6*$!.,*#$>4+!?$ Y ! ^"2#-;&#;-"8$*.8"$)-&5,#"&#;-"$ ] ! U)22,H"$3"H"32$.I$,*g:)&')6"$:)-)33"3,2($ ] ! c)-,)C3"$&.5"-"*&"$8.(),*2$)*8$,*#-)2.&'"#$("22)6,*6$ ] ! j"#"-.6"*".;2$(;3<g&.-"$>6-):5,&2/$&.(:-"22,.*/$"#&?$ Y ! !-.6-)((,*6$(.8"3$I.-$#5,2r$ Y ! 9.(:3"1$I);3#$C"5)H,.-$ ] ! D,*63"$&.-"$&.;38$"1:"-,"*&"$I);3#$ ] ! A""8$I.-$I);3#$8.(),*2$ Pete Beckman Argonne National Laboratory 26
BG/P & BG/Q Power Experiments Comparison between CNK and Linux on sleep() CNK Linux % KWatt 14.935 13.809 7.75 Pete Beckman Argonne National Laboratory 27
Exploring Power on Intel Knights Ferry Pete Beckman Argonne National Laboratory 28
Recommend
More recommend