Site Update for Oakforest-PACS at JCAHPC

Transcription

1 Site Update for Oakforest-PACS at JCAHPC Taisuke Boku Vice Director, JCAHPC Uiversity of Tsukuba 1 Ceter for Computatioal Scieces, Uiv. of Tsukuba

2 Towards Exascale Computig PF Tier-1 ad tier-2 supercomputers form HPCI ad move forward to Exascale computig like two wheels Future Exascale Post K Computer -> Rike R-CCSAICS 10 OFP JCAHPC(U. Tsukuba ad U. Tokyo) 1 Tokyo Tech. TSUBAME2.0 T2K U. of Tsukuba U. of Tokyo Kyoto U Ceter for Computatioal Scieces, Uiv. of Tsukuba

3 Fiscal Year Hokkaido Tohoku Tsukuba Tokyo Tokyo Tech. Nagoya Kyoto Osaka Deploymet pla of 9 supercomputig ceter (Feb. 2017) HITACHI SR16000/M1(172TF, 22TB) Cloud System BS2000 (44TF, 14TB) Data Sciece Cloud / Storage HA8000 / WOS7000 (10TF, 1.96PB) NEC SX- 9 他 (60TF) HA-PACS (1166 TF) COMA (PACS-IX) (1001 TF) SX-ACE(707TF,160TB, 655TB/s) LX406e(31TF), Storage(4PB), 3D Vis, 2MW Fujitsu FX10 Reedbush PF 0.7 MW (1PFlops, 150TiB, 408 TB/s), Hitachi SR16K/M1 (54.9 TF, 10.9 TiB, 28.7 TB/s) TSUBAME 2.5 (5.7 PF, 110+ TB, 1160 TB/s), 1.4MW FX10(90TF) Fujitsu FX100 (2.9PF, 81 TiB) CX400(470T F) Fujitsu CX400 (774TF, 71TiB) Power cosumptio idicates maximum of power supply (icludes coolig facility) ~30PF, ~30PB/s Mem BW (CFL-D/CFL-M) ~3MW 50+ PF (FAC) 3.5MW 50+ PF (FAC/UCC + CFL-M) SGI UV2000 (24TF, 20TiB) 2MW i total up to 4MW Cray: XE6 + GB8K + XC30 (983TF) Cray XC30 (584TF) TSUBAME 2.5 (3~4 PF, exteded) 3.2 PF (UCC + CFL/M) 0.96MW 30 PF (UCC PF (Cloud) 0.36MW CFL-M) 2MW 5 PF(FAC/TPF) 1.5 MW PACS-X 10PF (TPF) 2MW Oakforest-PACS 25 PF (UCC + TPF) 4.2 MW TSUBAME PF, UCC/TPF 2.0MW 100+ PF 4.5MW (UCC + TPF) 200+ PF (FAC) 6.5MW TSUBAME 4.0 (100+ PF, >10PB/s, ~2.0MW) 100+ PF (FAC/UCC+CFL- M)up to 4MW PF (FAC/TPF + UCC) MW NEC SX-ACE NEC Express PB/s, 5-10Pflop/s, MW (CFL-M) (423TF) (22.4TF) 0.7-1PF (UCC) 25.6 PB/s, Pflop/s, MW Kyushu 3 HA8000 (712TF, 242 TB) SR16000 (8.2TF, 6TB) FX10 (272.4TF, 36 TB) CX400 (966.2 TF, 183TB) 2.0MW PF (UCC/TPF) FX10 (90.8TFLOPS) 2.6MW???? PF (FAC/TPF + UCC/TPF) 3MW K Computer (10PF, 13MW) Post-K Computer (??.??)

4 JCAHPC Joit Ceter for Advaced High Performace Computig ( Very tight collaboratio for post-t2k with two uiversities For mai supercomputer resources, uiform specificatio to sigle shared system Each uiversity is fiacially resposible to itroduce the machie ad its operatio -> uified procuremet toward sigle system with largest scale i Japa To maage everythig smoothly, a joit orgaizatio was established -> JCAHPC 4 Ceter for Computatioal Scieces, Uiv. of Tsukuba

5 Machie locatio: Kashiwa Campus of U. Tokyo U. Tsukuba Kashiwa Campus of U. Tokyo Hogo Campus of U. Tokyo 5

6 Oakforest-PACS (OFP) U. Tokyo covetio U. Tsukuba covetio Do t call it just Oakforest! OFP is much better 25 PFLOPS peak 8208 KNL CPUs FBB Fat-Tree by OmiPath HPL PFLOPS #1 i Japa #6 #7 HPCG #3 #5 Gree500 #6 #21 Full operatio started Dec Official Program started o April

7 Computatio ode & chassis Water coolig wheel & pipe Chassis with 8 odes, 2U size Computatio ode (Fujitsu ext geeratio PRIMERGY) with sigle chip Itel Xeo Phi (Kights Ladig, 3+TFLOPS) ad Itel Omi-Path Architecture card (100Gbps) 7

8 Water coolig pipes ad rear pael radiator Direct water coolig pipe for CPU Rear-pael idirect water coolig for others 8

9 Specificatio of Oakforest-PACS Total peak performace 25 PFLOPS Total umber of compute odes 8,208 Compute ode Itercoect Logi ode Product Processor Fujitsu Next-geeratio PRIMERGY server for HPC (uder developmet) Itel Xeo Phi (Kights Ladig) Xeo Phi 7250 (1.4GHz TDP) with 68 cores Memory High BW 16 GB, > 400 GB/sec (MCDRAM, effective rate) Product Lik speed Topology Product Low BW # of servers 20 Processor Memory 96 GB, GB/sec (DDR x 6ch, peak rate) Itel Omi-Path Architecture 100 Gbps Fat-tree with full-bisectio badwidth Fujitsu PRIMERGY RX2530 M2 server Itel Xeo E5-2690v4 (2.6 GHz 14 core x 2 socket) 256 GB, 153 GB/sec (DDR x 4ch x 2 socket) 9 Ceter for Computatioal Scieces, Uiv. of Tsukuba

10 Specificatio of Oakforest-PACS (I/O) Parallel File System Fast File Cache System Type Total Capacity Meta data Object storage Type Product Lustre File System 26.2 PB # of MDS 4 servers x 3 set MDT Total capacity Product Product # of OSS (Nodes) Aggregate BW DataDirect Networks MDS server + SFA7700X 7.7 TB (SAS SSD) x 3 set DataDirect Networks SFA14KE 10 (20) ~500 GB/sec # of servers (Nodes) 25 (50) Aggregate BW Burst Buffer, Ifiite Memory Egie (by DDN) 940 TB (NVMe SSD, icludig parity data by erasure codig) DataDirect Networks IME14K ~1,560 GB/sec 10 Ceter for Computatioal Scieces, Uiv. of Tsukuba

11 Full bisectio badwidth Fat-tree by Itel Omi-Path Architecture 12 of 768 port Director Switch (Source by Itel) Uplik: 24 2 Dowlik: Firstly, to reduce switches&cables, we cosidered : All the odes ito subgroups are coected with FBB Fat-tree Subgroups are coected with each other with >20% of FBB But, HW quatity is ot so differet from globally FBB, ad globally FBB is preferredfor flexible job maagemet. 362 of 48 port Edge Switch 2 Compute Nodes 8208 Logi Nodes 20 Parallel FS 64 IME 300 Mgmt, etc. 8 Total

12 Facility of Oakforest-PACS system Power cosumptio # of racks 102 Coolig system Compute Node Type Facility Others Type Air coolig 4.2 MW (icludig coolig) actually aroud 3.0 MW Warm-water coolig Direct coolig (CPU) Rear door coolig (except CPU) Coolig tower & Chiller Facility PAC 12

13 Software of Oakforest-PACS Compute ode Logi ode OS CetOS 7, McKerel Red Hat Eterprise Liux 7 Compiler MPI Library Applicatio Distributed FS Job Scheduler Debugger Profiler gcc, Itel compiler (C, C++, Fortra) Itel MPI, MVAPICH2 Itel MKL LAPACK, FFTW, SuperLU, PETSc, METIS, Scotch, ScaLAPACK, GNU Scietific Library, NetCDF, Parallel etcdf, Xabclib, ppope-hpc, ppope-at, MassiveThreads mpijava, XcalableMP, OpeFOAM, ABINIT-MP, PHASE system, FrotFlow/blue, FrotISTR, REVOCAP, OpeMX, xtapp, AkaiKKR, MODYLAS, ALPS, feram, GROMACS, BLAST, R packages, Biocoductor, BioPerl, BioRuby Globus Toolkit, Gfarm Fujitsu Techical Computig Suite Alliea DDT Itel VTue Amplifier, Trace Aalyzer & Collector 13

14 TOP500 list o Nov (#50) # Machie Architecture Coutry Rmax (TFLOPS) Rpeak (TFLOPS) MFLOPS/W 1 TaihuLight, NSCW MPP (Suway, SW26010) 2 Tiahe-2 (MilkyWay-2), Cluster (NUDT, NSCG CPU + KNC) 3 Piz Dait, CSCS MPP (Cray, XC50: CPU + GPU) 4 Gyoukou, JAMSTEC MPP (Exascaler, PEZY-SC2) 5 Tita, ORNL MPP (Cray, XK7: CPU + GPU) 6 Sequoia, LLNL MPP (IBM, BlueGee/Q) 7 Triity, NNSA/ MPP (Cray, LABNL/SNL XC40: MIC) 8 Cori, NERSC-LBNL MPP (Cray, XC40: KNL) 9 Cluster (Fujitsu, Oakforest-PACS, JCAHPC KNL) Chia 93, , Chia 33, , Switzerlad 19, , Japa 19, , Uited States 17, , Uited States 17, , Uited States 14, , Uited States 14, , Japa 13, , K Computer, RIKEN AICS MPP (Fujitsu) Japa 10, , Ceter for Computatioal Scieces, Uiv. of Tsukuba

15 Post-K Computer ad OFP OFP fills gap betwee K Computer ad Post-K Computer Post-K Computer is plaed to istall time frame K Computer will be shutdow aroud ?? Two system software developed i AICS RIKEN for Post-K Computer McKerel OS for May-core era, for a umber of thi-cores without OS jitter ad core bidig Primary OS (based o Liux) o Post-K, ad applicatio developmet goes ahead XcalableMP (XMP) (i collaboratio with U. Tsukuba) Parallel programmig laguage for directive-base easy codig o distributed memory system Not like explicit message passig with MPI 15 Ceter for Computatioal Scieces, Uiv. of Tsukuba

16 OFP resource sharig program (atio-wide) JCAHPC (20%) HPCI HPC Ifrastructure program i Japa to share all supercomputers (free!) Big challege special use (full system size) opportuity to use etire 8208 CPUs by just oe project for 24 hours, every ed of moth U. Tsukuba (23.5%) Iterdiscipliary Academic Program (free!) Large scale geeral use U. Tokyo (56.5%) Geeral use Idustrial trial use Educatioal use Youg & Female special use Ordiary job ca use up to 2048 odes/job 16 Ceter for Computatioal Scieces, Uiv. of Tsukuba

17 Research Area based o CPU Hours Oakforest-PACS i FY.2017 (TENTATIVE: ~2017.9) 17 Lattice QCD NICAM-COCO GHYDRA Seism3D SALMON/ ARTED Egieerig Earth/Space Material Eergy/Physics Iformatio Sci. Educatio Idustry Bio Social Sci. & Ecoomics Data Ceter for Computatioal Scieces, Uiv. of Tsukuba

18 Performace variat betwee odes Lower is Faster Normalized elapse time / Iteratio Hamiltoia Curret Misc. computatio Commuicatio Best Worst ormalized to best case most of time is cosumed for Hamiltoia calculatio ot icludig commuicatio time domai size is equal for all odes root cause of strog scalig saturatio performace gap exists o ay materials No-algorithmic load-imbalacig Ø Ø Ø Ø dyamic clock adjustmet (DVFS) o turbo boost is applied idividually o all processors it is observed o uder same coditio of odes o KNL, more sesitive tha Xeo serious performace degradatio o sychroized large scale system 18 IXPUG ME 2018 (Site Update) Ceter for Computatioal Scieces, Uiv. of Tsukuba

19 Operatio summary Memory model basically 50:50 for cache:flat modes started to watch the queue coditio for getly chagig the ratio ~ 15% plaig to itroduce dyamic o-demad switchig i job by job maer KNL CPU almost good ad failure rate is eough uder estimatio by Fujitsu eough stability to support up to 2048 ode job OPA etwork at first there was a problem at bootig up time, but ow it s fixed almost -> it was the mai reaso agaist to the dyamic memory mode chage hudreds of liks have bee chaged by iitial failure, but ow stable Special operatio every moth, 24hours operatio for just oe project to occupy etire system 19

20 New machie plaed at CCS, U. Tsukuba PACS-X with GPU+FPGA 20 Ceter for Computatioal Scieces, Uiv. of Tsukuba

21 CCS at Uiversity of Tsukuba Ceter for Computatioal Scieces Established i years as Ceter for Computatioal Physics Reorgaized as Ceter for Computatioal Scieces i 2004 Daily collaborative researches with two kids of faculty researchers (about 35 i total) Computatioal Scietists who have NEEDS (applicatios) Computer Scietists who have SEEDS (system & solutio) 21 Ceter for Computatioal Scieces, Uiv. of Tsukuba

22 PAX (PACS) series history i U. Tsukuba Started i 1977 (by Hoshio ad Kawai) 1 st geeratio PACS i 1978 with 9 CPUs 6 th geeratio CP-PACS awarded #1 i TOP d PAX-32 1st PACS-9 5 th QCDPAX th CP-PACS #1 i the world 2006 PACS-CS (7 th ) first PC cluster solutio 2012~2013 HA-PACS (8 th ) itroducig GPU/FPGA Year Name Performace 1978 年 PACS-9 7 KFLOPS 1980 年 PACS KFLOPS 1983 年 PAX MFLOPS 1984 年 PAX-32J 3 MFLOPS 1989 年 QCDPAX 14 GFLOPS 1996 年 CP-PACS 614 GFLOPS 2006 年 PACS-CS 14.3 TFLOPS 2012~13 年 HA-PACS (PACS-VIII) PFLOPS 2014 年 COMA (PACS-IX) PFLOPS 22 Ceter for Computatioal Scieces, Uiv. of Tsukuba

23 AiS AiS: Accelerator i Swtich Usig FPGA ot oly for computatio offloadig but also for commuicatio Combiig computatio offloadig ad commuicatio amog FPGAs for ultralow latecy o FPGA computig Especially effective o commuicatiorelated small/medium computatio (such as collective commuicatio) Coverig GPU o-suited computatio by FPGA OpeCL-eable programmig for applicatio users CPU PCIe GPU FPGA comp. comm. high speed itercoect 23 Ceter for Computatioal Scieces, Uiv. of Tsukuba

24 AiS computatio model ivoke GPU/FPGA kersls data trasfer via PCIe GPU (ivoked from FPGA) CPU CPU GPU PCIe FPGA comp. comm. > QSFP28 itercoect PCIe collective or specialized commuicatio with computatio FPGA comp. comm. Etheret Switch 24 Ceter for Computatioal Scieces, Uiv. of Tsukuba

25 Evaluatio test-bed Pre-PACS-X (PPX) CCS, U. Tsukuba comp. ode HCA: Mellaox IB/EDR PACS-X prototype IB/EDR : 100Gbps CPU:Xeo E v4 CPU:Xeo E v4 GPU: NVIDIA P100 x2 QSFP+ : 40Gbpsx2 FPGA: Bittware A10PL4 25 Ceter for Computatioal Scieces, Uiv. of Tsukuba

26 Time Lie Feb. 2018: Request for Iformatio Apr. 2018: Request for Commet (followigs are just requiremet) basic specificatio: AiS-based large cluster with up to 256 odes V100 class of GPU x2 Stratix10 or UltraScale class of FPGA x1 (25% of total cout of odes) OPA x2 or IfiiBad HDR class itercoectio Aug. 2018: Request for Proposal Biddig closed o begi of Sep Mar. 2019: Deploymet Apr. 2019: Startig official operatio 26 Ceter for Computatioal Scieces, Uiv. of Tsukuba