High Performance Computing at CEA



Similar documents
How To Build A Supermicro Computer With A 32 Core Power Core (Powerpc) And A 32-Core (Powerpc) (Powerpowerpter) (I386) (Amd) (Microcore) (Supermicro) (

Supercomputing Resources in BSC, RES and PRACE

Cosmological simulations on High Performance Computers

Kriterien für ein PetaFlop System

PRACE hardware, software and services. David Henty, EPCC,

Appro Supercomputer Solutions Best Practices Appro 2012 Deployment Successes. Anthony Kenisky, VP of North America Sales

Interconnect Your Future Enabling the Best Datacenter Return on Investment. TOP500 Supercomputers, November 2015

HPC and Big Data. EPCC The University of Edinburgh. Adrian Jackson Technical Architect

PRACE the European HPC Research Infrastructure. Carlos Mérida-Campos, Advisor of Spanish Member at PRACE Council

Extreme Scaling on Energy Efficient SuperMUC

BSC - Barcelona Supercomputer Center

Evoluzione dell Infrastruttura di Calcolo e Data Analytics per la ricerca

Welcome to the. Jülich Supercomputing Centre. D. Rohe and N. Attig Jülich Supercomputing Centre (JSC), Forschungszentrum Jülich

HETEROGENEOUS HPC, ARCHITECTURE OPTIMIZATION, AND NVLINK

Agenda. HPC Software Stack. HPC Post-Processing Visualization. Case Study National Scientific Center. European HPC Benchmark Center Montpellier PSSC

Extreme Computing: The Bull Way

HPC Update: Engagement Model

Barry Bolding, Ph.D. VP, Cray Product Division

Mississippi State University High Performance Computing Collaboratory Brief Overview. Trey Breckenridge Director, HPC

International High Performance Computing. Troels Haugbølle Centre for Star and Planet Formation Niels Bohr Institute PRACE User Forum

Journée Mésochallenges 2015 SysFera and ROMEO Make Large-Scale CFD Simulations Only 3 Clicks Away

Extreme Scale Compu0ng at LRZ

Transforming the UL into a Big Data University. Current status and planned evolutions

Jean-Pierre Panziera Teratec 2011

OpenMP Programming on ScaleMP

Sun Constellation System: The Open Petascale Computing Architecture

Computational infrastructure for NGS data analysis. José Carbonell Caballero Pablo Escobar

Hadoop on the Gordon Data Intensive Cluster

YALES2 porting on the Xeon- Phi Early results

How Cineca supports IT

1 DCSC/AU: HUGE. DeIC Sekretariat /RB. Bilag 1. DeIC (DCSC) Scientific Computing Installations

HPC-related R&D in 863 Program

Jezelf Groen Rekenen met Supercomputers

CALMIP a Computing Meso-center for middle HPC academis Users

Stovepipes to Clouds. Rick Reid Principal Engineer SGI Federal by SGI Federal. Published by The Aerospace Corporation with permission.

JuRoPA. Jülich Research on Petaflop Architecture. One Year on. Hugo R. Falter, COO Lee J Porter, Engineering

Accelerating Simulation & Analysis with Hybrid GPU Parallelization and Cloud Computing

GPU System Architecture. Alan Gray EPCC The University of Edinburgh

Deploying and managing a Visualization Onera

Linux Cluster Computing An Administrator s Perspective

Performance, Reliability, and Operational Issues for High Performance NAS Storage on Cray Platforms. Cray User Group Meeting June 2007

A National Computing Grid: FGI

Relations with ISV and Open Source. Stephane Requena GENCI

MEGWARE HPC Cluster am LRZ eine mehr als 12-jährige Zusammenarbeit. Prof. Dieter Kranzlmüller (LRZ)


SGI High Performance Computing

How To Compare Amazon Ec2 To A Supercomputer For Scientific Applications

Estonian Scientific Computing Infrastructure (ETAIS)

Cluster Implementation and Management; Scheduling

Parallel Programming Survey

New Storage System Solutions

1 Bull, 2011 Bull Extreme Computing

PRIMERGY server-based High Performance Computing solutions

Intel Solid- State Drive Data Center P3700 Series NVMe Hybrid Storage Performance

LS-DYNA Best-Practices: Networking, MPI and Parallel File System Effect on LS-DYNA Performance

Current Status of FEFS for the K computer

Altix Usage and Application Programming. Welcome and Introduction

JUROPA Linux Cluster An Overview. 19 May 2014 Ulrich Detert


Cray s Storage History and Outlook Lustre+ Jason Goodman, Cray LUG Denver

Data management challenges in todays Healthcare and Life Sciences ecosystems

Building a Top500-class Supercomputing Cluster at LNS-BUAP

Purchase of High Performance Computing (HPC) Central Compute Resources by Northwestern Researchers

Scientific Computing Data Management Visions

FLOW-3D Performance Benchmark and Profiling. September 2012

Summit and Sierra Supercomputers:

News about HPC and Inria

The PHI solution. Fujitsu Industry Ready Intel XEON-PHI based solution. SC Denver

GPFS und HPSS am HLRS

Trends in High-Performance Computing for Power Grid Applications

The Hartree Centre helps businesses unlock the potential of HPC

A Study on the Scalability of Hybrid LS-DYNA on Multicore Architectures

Information about Pan-European HPC infrastructure PRACE. Vít Vondrák IT4Innovations

Supercomputer Center Management Challenges. Branislav Jansík

Energy efficient computing on Embedded and Mobile devices. Nikola Rajovic, Nikola Puzovic, Lluis Vilanova, Carlos Villavieja, Alex Ramirez

Performance Evaluation of Amazon EC2 for NASA HPC Applications!

Hank Childs, University of Oregon

Intel Cluster Ready Appro Xtreme-X Computers with Mellanox QDR Infiniband

David Vicente Head of User Support BSC

IBM System Cluster 1350 ANSYS Microsoft Windows Compute Cluster Server

Access, Documentation and Service Desk. Anupam Karmakar / Application Support Group / Astro Lab

Supercomputing Status und Trends (Conference Report) Peter Wegner

HCERES report on research unit:

The K computer: Project overview

ALPS Supercomputing System A Scalable Supercomputer with Flexible Services

Transcription:

High Performance Computing at CEA Thierry Massard CEA-DAM Mascot 2012 Meeting 23/03/2012 1 1

www-hpc.cea.fr/en Paris CEA/DIF (Bruyères-Le-Châtel) Saclay CEA Supercomputing Complex Mascot 2012 Meeting 23/03/2012 2

Brief Presentation of CEA o CEA, namely Atomic and Alternative Energies Commission, is a French governmental research organization o 15 000+ employees o 1 000+ doctoral students o 3.9 B /year budget o 4 000+ scientific publications/year o 300+ European projects with CEA participation o Main focuses of CEA are: o Defense and global security o Energy o Information & Health technologies o Related and necessary fundamental research o Other outcomes o Teaching and knowledge dissemination o Technology transfer (patents, start-up/spin-off companies ) Mascot 2012 Meeting 23/03/2012 3

HPC a huge need for CEA Basic science and applied research for the Simulation program Applied research, industrial applications in the field of Energy and Health and Communication technology Fundamental research (Matter and Biology) ocomité Stratégique du Calcul Intensif 4/15 o22 octobre 2007 Mascot 2012 Meeting 23/03/2012 4

Fundamental Sciences of Matter Nanosciences Particles Physics Environmental risks tsunamis Global warming Fusion Structure of the Universe Mascot 2012 Meeting 23/03/2012 5

CEA Scientific Computing Complex Defense Computing Center TERA R&D on HPC Technologies Très Grand Centre de calcul du CEA TGCC CCRT Address the needs of CEA, French & European Research Communities and Industry in terms of High Performance Simulations While developing and assessing news technologies for more powerful and efficient HPC systems Joint Laboratories Extreme Computing Lab (BULL, CEA) Exascale Computing Research (INTEL, GENCI, CEA, UVSQ) Mascot 2012 Meeting 23/03/2012 6 6

Compute Complex Capacity Roadmap 100000 TERA-1000 10000 TERA-100 1000 Tflops 100 TERA-10 TGCC 10 TERA-1 CCRT-B 1 CCRT-A 0,1 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 Mascot 2012 Meeting 23/03/2012 7 7

TERA Computing Centre Classified computing resources for the CEA/DAM Simulation Program Mascot 2012 Meeting 23/03/2012 8 8

TERA 100 1 st european supercomputer in 2010 above the sustained Petaflops 1,25 Petaflop/s peak 4 300 nodes 140 000 coresintel Xeon Nehalem EX QDR Infiniband interconnect System software Open source 300 TB memory 20 PB storage disk 500 GB/s bandwidth toward the storage system High reliability machine to meet the strategic needs of CEA/DAM Mascot 2012 Meeting 23/03/2012 9

Grand Challenges on TERA 100 o We have taken advantage of the pre-production phase to «invite» a selection of frontier extreme computing cases : Part of the V&V process for the global system Reach the limit of the machine To observe the behavior of other software than those in CEA-DAM s portfolio o Sampling from the nov 2010 feb.2011 campaign : Massive parallellcalculation for the 3D transport of a sodium fast reactor using APOLLO3 R. Baron, C. Calvin, J. Dubois, J-J. Lautard, S. Salmons CEA/DEN/DANS/DM2S/SERMA up to 30000+ cores Simulation of graphene growth on Silicon Carbibe (BigDFT) Thierry Deutsch, Luigi Genovese, Pascal Pochet, Eduardo Machado-Charry CEA/DSM/INAC Grenoble Classical Molecular Dynamics Simulations (STAMP) L. Soulard et al., CEA/DAM/DPTA up to 10000 cœurs Design of artiificial enzymes (STAMPS) G. Debret, M. Masella, P. Cuniasse, G. Collet CEA/DSV/iBiTecS Mascot 2012 Meeting 23/03/2012 10 10

Mascot 2012 Meeting 23/03/2012 11 11

Mascot 2012 Meeting 23/03/2012 12 12

Simulations de dynamique moléculaire TERA-100 La dynamique moléculaire résout l équation fondamentale de la dynamique pour un ensemble de N particules en interaction mutuelle. Pourvu que N soit suffisamment grand, elle peut permettre de faire le lien entre le calcul quantique et la mécanique des milieux continus. Performances du code Stamp sur TERA-100 entre 250 millions et un milliard d atomes, et jusqu à 10 000 processeurs. Ejection de matière lors de la réflexion d un choc sur une surface libre. 125 millions d atomes, 4000 cœurs, potentiel interatomique assez simple mais temps physique de simulation très long (pour la dynamique moléculaire) : de l ordre de la nanoseconde. Plastification du diamant au passage d une onde de choc. Relativement peu d atomes (1,3 millions, 6400 cœurs) mais potentiel interatomique extrêmement complexe. Mascot 2012 Meeting 23/03/2012 13 13

TERA-100: Co-designed by BULL and CEA/DAM TERA-100 has been the first European computer to reach the Petaflops. General Purpose production system with a tremendous effectiveness and reliability (1.05 Pflops, 83,7% of the peak performance in a 22 hours run) Tera100 key figures : 220 racks in 650 m² 4370 servers interconnected with IB-QDR 17480 Intel X7560 processors 140 000 compute cores 300 TB of memory 500 GB/s I/O bandwidth System software mostly based on Open source developments (Linux, LUSTRE, SLURM..) Mascot 2012 Meeting 23/03/2012 14 14

Supercomputers... A roadmap to Exaflops Joint Bull/CEA R&D to insure availability of supercomputers for the needs of the Simulation program EXA 1 N? Monde N? EU 1 Ef 2019 Ambitious goal : to maintain a european capabillity to design and build Supercomputers in Europe x30 TERA 1000 High ROI in technology and economy for the civilian applications x30 N? Monde N? EU 30Pf 2001 Closed DAM machines TERA 1996 1 N 5 monde N 1 EU 4,8 TF x10 TERA 10 2001 N 5 monde N 1 EU 63 TF x20 2005 TERA 100 N 6 monde N 1 EU 1Pf TGCC Open CEA computing centers Mascot 2012 Meeting 23/03/2012 Direction 15 des

CCRT Computing Centre for Research and Technology A partnership approach to gather academic research and industrials high end simulation in the same computing centre Mascot 2012 Meeting 23/03/2012 16 16

Computing Center for Research and Technology o A successful business model since 2003 o CCRT major goals o Meet CEA & partners simulation needs for research o Synergy between research organizations, universities and industry Promoting exchanges and scientific collaborations between partners Usages of HPC, code scaling and optimization Centralized application support, workshops o Industrial partners o Multi-year agreements for a share based on total cost of ownership (sharing risk and opportunities) Mascot 2012 Meeting 23/03/2012 17

CCRT Partnership Overview (2011) ONERA 1,0% Techspace Aero 0,8% Snecma 8,9% INERIS* 1.5% Areva* Turbomeca 3% 8% Astrium 1,0% Cerfacs 0,7% DAM 12,2% DEN 17,2% EDF 22% DSV 5,5% DSM 18,2% Mascot 2012 Meeting 23/03/2012 18 18

Current CCRT Architecture (~200 Tflops) NEC Vector system Hybride (CPU/GPU) Bull Cluster BULL Itanium Cluster Extension CCRT-2010 40 Tflops 47.7 Tflops Mémoire 23.2 To Disques 420 To 6 * 10 Gb/s HP Post-processing Cluster 8 * 1Gb/s SX8 SX9 8 nodes 3 nodes 2 Tflops 4.8 Tflops 0.9 To 3 To 40 TB of Disk 8 * 10 Gb/s 8 * 10 Gb/s 1 092 nœuds Bull 8736 Intel/Nehalem cores 3 GB/core 103 TFlops SGI Altix 450 DMF Level 1: 1 Po of DDN Sata disk 500 TB of disk 48serveurs Tesla S1070 46 080 NVIDIA cores 4 GB/GPU 192 Tflops-SP 38 graphic nodes 3.2 TB of memory, 100 TB of disk Backbone Ethernet 10Gb/s Users 2 Gb/s NFS Server Level 2: SUN-SL8500 (9940 & T10000 tapes) Storage System Mascot 2012 Meeting 23/03/2012 19 19

TGCC Très Grand Centre de calcul du CEA A completely new computing centre built by CEA to host the French contribution to the PRACE Research Infrastructure and other high end HPC resources Mascot 2012 Meeting 23/03/2012 20 20

TGCC Overview Designed to host : o2nd system of the PRACE Research Infrastructure (European Tier0) o3rd generation of the CCRT (French - Tier1) ohigh speed network node (10+ GIGE links) Main characteristics of the infrastructure onew power line 225 kv / 60 MW oinitially configured for ~12 MW (7.5 MW of IT equipments) o2600 m² machine rooms floor space o Conference room for 200 people ooperated by CEA/DAM-Ile-de-France site teams Mascot 2012 Meeting 23/03/2012 21 21

Très Grand Centre de Calcul du CEA «in one glance» Mascot 2012 Meeting 23/03/2012 22

The European strategy for HPC Gouvernance européenne 21 pays membres dont la France représentée par GENCI Tier-0 European Centres Tier-1 National Centres > 1 Pflop/s Gouvernance nationale Tier-2 Regional Centres (Universities) < 50 Tflop/s Gouvernances régionales Equip@meso Mascot 2012 Meeting 23/03/2012 23

PRACE supercomputers GSC@Juelich JUGENE : IBM BG/P 1 PF GENCI@TGCC CURIE : BULL Bullx, 2 PF GSC@Juelich & CINECA IBM BG/Q systems (Fermi) GSC@HLRS Hermit : Cray XE6, 1 PF Mascot 2012 Meeting 23/03/2012 GSC@LRZ SuperMUC : IBM idataplex, 3 PF BSC IBM 24

Mai JUGENE : 360Mh PRACE : calls in progress 2011 2012 2013 Novembre 1st regular call JUGENE : 360Mh CURIE : 40Mh (nœuds larges) Mai Novembre Mai Novembre 2nd regular call JUGENE : 360 Mh CURIE : 30 Mh (NL), 125 Mh (NF) HERMIT : 160 Mh 3rd regular call JUGENE : 33 % PRACE CURIE : 80 % PRACE HERMIT : 33 % PRACE SUPERMUC : 50% PRACE FERMI :? MARENOSTRUM :? 1 er nov à mi janvier JUGENE : 300 Mh, HERMIT : 160 Mh, SuperMuc (LRZ) : 200Mh CURIE : 28Mh (NL), 188 Mh (NF), Marenostrum (BSC) : 135 Mh, FERMI (CINECA) : 300 Mh 4 th regular call Preparatory Access Call Appel à projets Mascot 2012 Meeting 23/03/2012 Allocation JUGENE : 10 Mh/an - CURIE : 5 Mh/an (NL), 20 Mh/an (NF) et 1.4 Mh/an (230kh en 2011) (NH) - HERMIT (tbd) Appel continu relevé tous les 3 mois 25

Curie: The BULL Tier0 System Mascot 2012 Meeting 23/03/2012 26 26

Curie configuration Fat nodes 360 BullX S6010 Intel NH EX 2,26 GHz 11520 cœurs 32 =>128 core/node 128 GB => 512 GB/node 4 GB/core 105 TFlops Thin nodes 5040 BullX B510 Intel Sandy Bridge 2,7 GHz 80 640 cœurs 16 core/node 4 GB/core 1 740 TFlops Hybrid nodes 144 Bullx B505 288 Nvidia M2090 192 + 12 TFlops Interconnect Infiniband QDR 5 Po - 150 GB/s Lustre file system Mascot 2012 Meeting 23/03/2012 27 27

CEA Supercomputing Complex Summary A single operator mutualizing expertise developed at CEA/DAM A rich ecosystem connected from regional to worldwide scale Collaborations with academics and industry ranging from design of HPC system and their environment (I/O, storage..) to applications and usages through system and management software, algorithms and middleware TERATEC TGCC TERA CCRT Mascot 2012 Meeting 23/03/2012 28

ENJOY YOUR STAY AT TGCC! Mascot 2012 Meeting 23/03/2012 29 29