第3回戦略シンポジウム緑川公開用

Size: px

Start display at page:

Download "第3回戦略シンポジウム緑川公開用"

としなりしんまつ
6 years ago
Views:

2 (SDSM) SMS MpC DLM

3 Top500 Top 500 list of Supercomputers (

4 Top 500 list of Supercomputers (

5 SMP Symmetric Multiprocessor CPU CPU CPU CPU CPU CPU CPU CPU cluster CPU CPU

8 send a recv b recv x send y

9 a = 10 b = a data

10 CPU 1. send p0 a; receive p1 b; ) MPI ( 2. b=a; OpenMP ( )

11 P0 P1 int a; int a; a=10; send P1 a; recv P0 a; P0 P1 int a; a=10; a a

12 2 b a P0 P1 int a=10; int b send P1 a; recv P0 b; P0 P1 int a=10; int b; b=a; a,b a,b

14 SMS SDSM b = a x = y

15 (SDSM) SMS MpC DLM

16 (SDSM) SMS MpC DL

17 PC 0 PC 1 PC 2 P0 P1 P2 PC

19 MpC C shared auto register static extern typedef shared

20 MpC P0 shared int a; int b; main( ){ int c=10; a=20; } P1 shared int a; int b; main( ){ int c=10; b=a; } PC

21 mpcc prg.mpc -osms_prog use sms MpC MpC MpC MpC MpC C C SMS pthread SDSM SMS, TreadMarks, JIAJIA pthread SDSM pthread

22 MpC vs. OmniOpenMP (Score : RWC floyd : Shortest path search 1Gbps Ethernet2Gbps Myrinet2000 (sec) MpC OmniOpenMP 2 Myrinet PM H. Midorikawa, et al.: "The Performance Analysis of Portable Parallel Programming Interface MpC for SDSM and pthread", Proc. of IEEE/ACM Inter. Symp. on Cluster Computing and the Grid (CCGrid2005),Vol.2, pp

23 (sec) ep (sec) laplace

24 (sec) mm(blocking) (sec) mm(nonblocking) MpC

25 NPB3.0 API 74% MpC 7% H. Midorikawa, et al.: "The Performance Analysis of Portable Parallel Programming Interface MpC for SDSM and pthread", Proc. of IEEE/ACM Inter. Symp. on Cluster Computing and the Grid (CCGrid2005),Vol.2, pp , (2005) MpC,,Vol.46 No.SIG4(ACS9), pp.69-85, 2005)

26 (SDSM) SMS MpC DLM

27 64bitOS x86_6448bit 256TB) 48bit (256tebibytes) 56bit (64pebibytes) 64bit (16exbibytes)

28 DLM ( Distributed Large Memory) Cal Thread calhost Cal Process Com Thread DLM memhost1 Memserv Process memhost2 Memserv Process memhost3 Memserv Process usr_prog args -- -n 4 f hostfile hostfile calhost 2048 // 2GB memhost // 8GB memhost // 4GB memhost // 4GB memhost // 4GB : DLM OS

DLM 1 matv.c #include <stdio.h> #include <dlm.

main(int argc, char *argv[]) { int i,j; double temp; //

= i; // x for (i = 0; i < N; i++) x[i] = i; //

29 DLM 1 matv.c #include <stdio.h> #include <dlm.h> #define N // total memory 2 31 B B, 2GiB dlmcc matv.c -omatv dlm double a[n][n], x[n], y[n]; // DLM int main(int argc, char *argv[]) { int i,j; double temp; // a for ( i = 0; i< N; i++) for ( j = 0; j<n; j++) a[i][j] = i; // x for (i = 0; i < N; i++) x[i] = i; // a[n][n]*x[n]=y[n] for(i = 0; i < N; i++){ temp = 0; for(j = 0; j<n; j++) temp += a[i][j]*x[j]; y[i] = temp; } return 0; } dlm DLM

30 10GbEthe CSLM 10GbEther (CSLM) swap 10GB Cluster Node CPU Node Memory PCI bus OS HP DL585 G2 x 5 Nodes Opteron 2.8GHz x 4 (8Cores) 64GByte 64GiB) 64bit/100MHz PCI-X, PCI-Expressx4 PCI-Expressx8 Linux kernel x86_64 TCP socket Compiler gcc version Network NIC Switch Hard Disk 10GbEthernet protocol Myri-10G Fujitsu XG1200(10GbE Switch) SAS 147GB 10krpm 2 RAID1 Smart array 5i HP B21 (SAS 147GB, 10krpm, TransRate 300MBps, seektime 4(Ave),8.1(Max)ms) op // 60GB op hostfile

31 SwapDisk DLM 10GbEthernet, 64GBMemory,10GBswap matv.c swap 15%, DLM 10 Disk DLM GB 64GiB swap3%, DLM : DLMDLM, Vol.102, No.398, pp.29-34, 2007

32 SwapDisk DLM 1GbEthernet, 1GBMem, 4GBswap matv.c swap 160%, DLM 9.5 swap DLM : DLMDLM, Vol.102, No.398, pp.29-34, 2007

33 DLM (STREAM Benchmark) 380MB/sec ~ 40MB/ sec DLM 1MB ~ 4KB DLM Panda[2005] InfiniBand RDMA, 119MB/s [2006] 10Gb Ethernet, NIC,RDMA 131MB/S 204MB/s

) DLM Himeno Benchmark Large) 35 DLM 4KB DLM 5 10

Remote Memory Swapping over Cluster Nodes", Proc.

34 ) DLM Himeno Benchmark Large) 35 DLM 4KB DLM % DLM 1MB 5 DLM [2006] GbE NIC RDMA KB H.Midorikawa et al. : "DLM: A Distributed Large Memory System using Remote Memory Swapping over Cluster Nodes", Proc. of IEEE Cluster2008 pp , ( ) " DLM10GbEthernet ",, Vol.1, No.3, pp

35 DLM-MPI MPI DLM MPI Ethernet, InfiniBand, Myrinet ) 2 T2K-Tokyo Myri-10G 4 40Gbps) T2K-Tsukuba InfiniBand 64Gbps)

36 !"#$%&'($)*&'+,-.&*/'+ ੩ᄢቇ ᚑ 䋨ᣣ 䋩 Top500 45位日本3位 Nov インターコネクト Myri-10G x 4本 /node Node間 40Gbps x双方向 J?$7-K&"J8'?&6 /"L:M""/)E:F'&> A$7"/*B"#$%&'C D01E:F'&>"G"/"H 9*1E:F'&>"G"9!"#$%&'! ()* +,&-.! /01234$,' 5&6$78! 9/"2: 4CPUs(16コア /node AMD Opteron8356 メモリ 32GB/node 8nodes メモリ 128GB/node 3;44<:='&>?=$#"@#?&7>$##&>? A$7")/*"#$%&'C"*I)2:F'&>!"#䉲䊮䊘䉳䉡䊛䈧䈒䈳"$$%!"#䉲䊮䊘䊘䉳䉡䊛䈧䈒䈳"$$% システム図朴泰祐 T2Kシンポジウムつくば2008 資料

37 STREAM Kernel COPY SCALE ADD TRIAD Code a(i) = b(i) a(i) = q*b(i) a(i) = b(i) + c(i) a(i) = b(i) + q*c(i) DLM-MPI: MPI-MX 493MB/s Myri-10G 2 613MB/s Myri-10G 4 DLM-socket: TCP/IP EthernetonMyri-10G x 1 380MB/s

38 DLM Himeno benchmark XLARGE (112GB) 179.4MFLOPS, Relative Time 2.32 ( based on the time in Elarge,15GB) float 1025 x 1025 x GB/node x 6 nodes Local memory ratio 17.4% Bonding = 4 XLARGE-d (241GB)88.8 MFLOPS, Relative Time 4.68 ( based on the time in Elarge,15GB) double1025 x 1025 x GB/node x12nodes Localmemory ratio 8.1% Bonding = 4

39 !"#$%&'($)*&'+,-.&*/'+ 01&&+- 2$3+45$64&4(7 Top Nov. 2009!"#$%&'! ()* +,&-.! /0"123$,' 4&5$67! 89"1: InfiniBand 4x DDR 4 /node Node 64Gbps x 2C33D:E'&B<E$#"F#<&6B$##&B< 0G8"1:A' 4CPUs(16 /node AMD Opteron GB/node ;<$6-=&";7'<&5 *99"1:>""?8@:A'&B T2K

40 DLM-M DLM TCP/IP

41 Clients DLM fora Cluster on LAN Memory Servers

42 WANInTrigger WAN 17, WAN /848 Intrigger :, Vol.49, No.8, pp , Aug.2008

43 DLM forclusters on WAN Client User User Program Cluster(LAN) DLM-LAN Admin DLM-WAN Admin Cluster(LAN) DLM-LAN Admin Group of Clusters (WAN) Cluster(LAN) DLM-LAN Admin Memory Server Calculate Node Memory Server Calculate Node Memory Server Calculate Node

44 Thank you!

( CUDA CUDA CUDA CUDA ( NVIDIA CUDA I

( CUDA CUDA CUDA CUDA ( NVIDIA CUDA I GPGPU (II) GPGPU CUDA 1 GPGPU CUDA(CUDA Unified Device Architecture) CUDA NVIDIA GPU *1 C/C++ (nvcc) CUDA NVIDIA GPU GPU CUDA CUDA 1 CUDA CUDA 2 CUDA NVIDIA GPU PC Windows Linux MaxOSX CUDA GPU CUDA NVIDIA