General overview and first results.

Transcription

1 French Technological Watch Group: General overview and first results Fusion workshop at «Maison de la Simulation» 29/11/2016

2 FRENCH TECHNOLOGICAL WATCH GROUP Led by GENCI and its partners SC'16 l 01/12/2016 l 2

3 FRENCH TECHNOLOGICAL WATCH GROUP Led by GENCI and its partners Ø Share and mutualise expertise of partners at the national level Ø Anticipate upcoming Exascale architectures and provide orientations for future procurements ðmanycore/heterogeneous processors ðdeeper and more complex memory hierarchies ðfault Tolerance ðenergy optimisation Ø Organise code modernization for scientific communities ðease migration to these new platforms ðpreserve code legacy by using standards OpenMP ðenable potential specific optimisations with low level languages SC'16 l 01/12/2016 l 3

4 TECHNOLOGICAL WATCH GROUP Led by GENCI and its partners Link to developers of representative applications Climate : DYNAMICO, MesoNH Astro & Geophysics : RAMSES, EMMA, SPECFEM3D Combustion : YALES2, AVBP, TRUST Fusion : GYSELA5D Materials : METALWALLS Particle Physics : SMILEI, CMS-MEM Kernels : PATMOS, HYDRO, QR_MUMPS, QMC=Chem Deep/Machine Learning Link to developers of tools and libraries Core of the group : HPC expertise System (runtime, checkpoint/restart National centers: CINES, IDRIS and ) Profiling perf, energy Numercial solvers Data management/analysis TGCC Inria Maison de la Simulation «Groupe Calcul» GENCI Vendors support SC'16 l 01/12/2016 l 4

5 FRENCH TECHNOLOGICAL WATCH GROUP Led by GENCI and its partners q Deployment of 2 early technology platforms A Bull sequana platform at CINES based on Intel KNL 48 KNL nodes => 146 Tflop/s peak performance An IBM OpenPOWER platform at IDRIS based on P8+ Nvidia GPU 4*P100 GPU (latest generation) per node Nvlink between P8 and GPU 12 nodes => 254 Tflop/s peak performance (GPU only) More than 400GB/s bandwidth per node Software stack is not at its highest maturity q Tight collaboration between : Vendors, integrators, developers, HPC centres qevaluation of real time performance/energy profiling tools Allinea - Performance Report / MAQAO / Others Intel Vtune Amplifier and MPI Performance Snapshot (MPS) Nvprof for the OpenPOWER solution SC'16 l 01/12/2016 l 5

6 SOFTWARE ENVIRONMENT Programming Options for OpenPOWER qprogramming on Intel Knight Landing Straightforward: Intel Compiler, Intel MPI, OpenMP q Programming options for OpenPOWER Not as straightforward as Intel KNL General availibility for OpenMP is scheduled for the end of 2016 Full support is scheduled for 2017 SC'16 l 01/12/2016 l 6

7 KNL PORTING AND OPTIMISATION First results q Porting & optimisation workshops with Atos & Intel support Workshop on Colfax Ninja developer platforms (Intel processor Xeon Phi 7210) Workshop using the 48 Intel Xeon Phi 7250 nodes machine Frioul qporting efficiency: 2 1,8 1,6 1,4 Effective mean speed up obtained after 2 workshops (Haswell vs KNL) Speed up 1,2 1 0,8 0,6 0,4 0, Working time (hours) Performance vs Haswell 24c Energy efficiency vs Haswell 24c SC'16 l 01/12/2016 l 7

8 KNL PORTING AND OPTIMISATION First results Genci, code KNL speedup 4,5 4 3,5 3 2,5 2 1,5 1 0,5 0 CFD CFD CFD Climate Climate Geophysics Kernel Kernel Kernel Materials Speedup (Haswell node) Speedup (Broadwell node) Base SC'16 l 01/12/2016 l 8

9 OPENPOWER PORTING AND OPTIMISATION First results qmesonh: weather forecasting code 60% of the code has been ported using OpenACC Results with 16 MPI processes and MPS (multi-process service) PPOINT PROJET CELLULE DE VEILLE TECHNOLOGIQUE l 01/12/2016 l 9

10 OPENPOWER PORTING AND OPTIMISATION First results qvasp (an IBM ported app) Results reflect the first results, the whole application still have to be ported Speedup vs Full-Node, CPU-Only Execution 2,00 1,80 1,60 1,80 1,86 35% Faster 80% Faster 86% Faster 1,40 1,20 1,00 The Reference 1,00 1,35 Only 25% Slower 0,80 Only 50% Slower 0,75 0,60 0,46 0,40 0,20 0,00 Firestone Firestone Firestone Firestone Minsky Minsky Firestone 20 MPI Tasks 20 Tasks Firestone 1 MPI Task 1 Tasks Firestone 4 MPI Task 4 Tasks Firestone 20 MPI Task 20 Tasks Minsky 1 MPI 1 Tasks 1 Minsky 4 MPI 4 Tasks 4 0 GPUs 1 GPUs 4 GPUs 4 GPUs GPUs 1 4 GPUs PPOINT PROJET CELLULE DE VEILLE TECHNOLOGIQUE l 01/12/2016 l 10

11 FRENCH TECHNOLOGICAL WATCH GROUP Focus on Gysela qwork has just began on GPU with Altimesh Very first results qfirst lessons (= general good practices) Avoid buffers: That store addition/multiplication That are reused once àprefer recomputing than precomputing Use stack vs heap Malloc is very expensive (especially with many cores) Use zero-copy for data that you acces/write once SC'16 l 01/12/2016 l 11

12 TECHNOLOGICAL WATCH GROUP OPENING Phase 3 Q qopening to a wider panel of applications in May 2017 Applications will be reviewed through the «DARI» preparatory access Applications additionnelles Applications additionnelles PPOINT PROJET CELLULE DE VEILLE TECHNOLOGIQUE l 01/12/2016 l 12

13 FRENCH TECHNOLOGICAL WATCH GROUP Questions? Thank you for your attention PPOINT PROJET CELLULE DE VEILLE TECHNOLOGIQUE l 01/12/2016 l 13