View
6
Download
0
Category
Preview:
Citation preview
Scientific Computing withIntel Xeon Phi Coprocessors
Andrey VladimirovColfax International
HPC Advisory Council Stanford Conference 2015
Compututing with Xeon Phi Welcome © Colfax International, 2014
Contents
§1 MIC Architecture, Developer’s Perspective§2 Case Studies
Ï Astrophysics (offload story)Ï N-body simulation (offload vs native in a cluster)Ï Finanical Monte Carlo (heterogeneous clustering)Ï Computational fluid dynamics (legacy code)
§3 Colfax Developer Training
Compututing with Xeon Phi Welcome © Colfax International, 2014
§1. MIC Architecture from Developer’sPerspective
Compututing with Xeon Phi MIC Architecture from Developer’s Perspective © Colfax International, 2014
Intel Xeon Phi Coprocessors and the MIC Architecture
PCIe end-point device
High Power efficiency
∼ 1 TFLOP/s in DP
Heterogeneous clustering
For highly parallel applications which reach the scaling limitson Intel Xeon processors
Compututing with Xeon Phi MIC Architecture from Developer’s Perspective © Colfax International, 2014
Examples of Solutions with the Intel MIC Architecture
Colfax’s CXP7450 workstation withtwo Intel Xeon Phi coprocessorsxeonphi.com/workstations
Colfax’s CXP9000 server with eightIntel Xeon Phi coprocessorsxeonphi.com/servers
Compututing with Xeon Phi MIC Architecture from Developer’s Perspective © Colfax International, 2014
Intel Xeon Phi Coprocessors and the MIC Architecture
≤18 cores/socket ≈3 GHz
2-way hyper-threading
Up to 768 GB of DDR3 RAM
256-bit AVX vectors
57 to 61 cores at ≈1 GHz
4 hardware threads per core
6–16 GB cached GDDR5 RAM
512-bit IMCI vectors
C/C++/Fortran; OpenMP/MPI
Linux OS (on host and on coprocessor)Compututing with Xeon Phi MIC Architecture from Developer’s Perspective © Colfax International, 2014
Linux µOS on Intel Xeon Phi coprocessors (part of MPSS)user@host% lspci | grep -i "co-processor"06:00.0 Co-processor: Intel Corporation Xeon Phi coprocessor 3120 series (rev 20)82:00.0 Co-processor: Intel Corporation Xeon Phi coprocessor 3120 series (rev 20)user@host% sudo service mpss statusmpss is runninguser@host% cat /etc/hosts | grep mic172.31.1.1 host-mic0 mic0172.31.2.1 host-mic1 mic1user@host% ssh mic0user@mic0% cat /proc/cpuinfo | grep proc | tail -n 3processor : 237processor : 238processor : 239user@mic0% ls /amplxe dev home lib64 oldroot proc sbin sys usrbin etc lib linuxrc opt root sep3.10 tmp var
Compututing with Xeon Phi MIC Architecture from Developer’s Perspective © Colfax International, 2014
Offload and Native modesExplicit offload mode:
Native mode:
Compututing with Xeon Phi MIC Architecture from Developer’s Perspective © Colfax International, 2014
Optimization Areas
Common methods for Intel Xeon CPUs and Intel Xeon Phi coprocessors:
1 Scalar optimization (compiler-friendly practices)
2 Vectorization (must use 16- or 8-wide vectors)
3 Multi-threading (must scale to 100+ threads)
4 Memory access (streaming access or tiling)
5 Communication (offload, MPI traffic control)
Compututing with Xeon Phi MIC Architecture from Developer’s Perspective © Colfax International, 2014
Getting Ready for the Future
Knights Landing (KNL) – next generation of Intel MIC architecture
3x the performance of current generation
Available as a stand-alone processor or as a coprocessor
Compututing with Xeon Phi MIC Architecture from Developer’s Perspective © Colfax International, 2014
Getting Ready for the Future
The best way to prepare applications for KNL is to optimize them forIntel Xeon Phi coprocessors based on KNC.
Compututing with Xeon Phi MIC Architecture from Developer’s Perspective © Colfax International, 2014
§2. Case Studies
Compututing with Xeon Phi Case Studies © Colfax International, 2014
Astrophysical Code HEATCODE: an Offload Story
xeonphi.com/papers/heatcode
Compututing with Xeon Phi Case Studies © Colfax International, 2014
Astrophysical Code HEATCODE: an Offload Story
xeonphi.com/papers/heatcode
Compututing with Xeon Phi Case Studies © Colfax International, 2014
N-body Simulation: Offload vs Native in a Cluster
xeonphi.com/papers/nbody-basic
Compututing with Xeon Phi Case Studies © Colfax International, 2014
N-body Simulation: Offload vs Native in a Cluster
Initial Multi-threaded
Vectorizedwith SoA
ScalarTuning
Tiled,Unrolled
0
500
1000
1500
2000
Sin
gle
Prec
isio
n G
FLO
P/s
5.3140 180
480 520
0.8120
220
870
1620
N-Body Simulation Performance
Processor: Intel Xeon E5-2697 v2 Coprocessor: Intel Xeon Phi 7120P
xeonphi.com/papers/sc14
Compututing with Xeon Phi Case Studies © Colfax International, 2014
N-body Simulation: Offload vs Native in a Cluster
0
5
10
15
20
25
30
35
1 2 3 4 8 12 16
Per
form
ance
, TFLO
P/s
Number of Nodes or Coprocessors (P)
92% eff
76% eff
Intel Xeon E5-2697 v2 CPUs (4 nodes)
Intel Xeon Phi 7120P coprocesors (4 per node)N=220 particles (strong scaling)
1 Xeon Phi/node
2 Xeon Phi/node
3 Xeon Phi/node
4 Xeon Phi/node
Xeon Phi,native MPI
Xeon Phi,MPI+Offload
CPU
xeonphi.com/papers/sc14
Compututing with Xeon Phi Case Studies © Colfax International, 2014
Asian Option Pricing: Heterogeneous Clustering
xeonphi.com/papers/heterogeneous
Compututing with Xeon Phi Case Studies © Colfax International, 2014
Computational Fluid Dynamics: Legacy Code
xeonphi.com/papers/shallowCompututing with Xeon Phi Case Studies © Colfax International, 2014
§3. Colfax Developer Training
Compututing with Xeon Phi Colfax Developer Training © Colfax International, 2014
Colfax Developer Training
Intel Xeon Phi Coprocessor ProgrammingFuture-Proofing Applications for Knights Landing (KNL)
xeonphi.com/trainingCompututing with Xeon Phi Colfax Developer Training © Colfax International, 2014
Free Training for HPCAC Stanford 2015 Participants
xeonphi.com/hpcac2015
Compututing with Xeon Phi Colfax Developer Training © Colfax International, 2014
Recommended