A PERFORMANCE COMPARISON USING HPC BENCHMARKS: WINDOWS HPC SERVER 2008 AND RED HAT ENTERPRISE LINUX 5 R. Henschel, S. Teige, H. Li, J. Doleschal , M. S. Mueller October 2010
Contents • HPC at Indiana University • FutureGrid • Benchmarking Setup • Hardware • Results • SPEC MPI2007 medium • NSF Acceptance Test Suite • HPCC • Conclusion Robert Henschel
HPC at Indiana University • Indiana University • Founded in 1820 • 8 campuses • Bloomington and Indianapolis • 80,000 students, 7,000 faculty • HPC systems are operated by Research Technologies of University Information Technology Services • Open access to graduate students, faculty and staff Robert Henschel
HPC at Indiana University • IU is one of 11 TeraGrid resource providers Robert Henschel
HPC at Indiana University • IU hosts the Open Science Grid (OSG) operations center Robert Henschel
FutureGrid • $15 Mio. NSF Track 2 D award (2009) • Support the research on the future of distributed, grid, and cloud computing. • Build a robustly managed simulation environment and testbed to support the development and early use in science of new technologies at all levels of the software stack: from networking to middleware to scientific applications. • Mimic TeraGrid and/or general parallel and distributed systems. • FutureGrid is a (small 5600 core) Science Cloud but it is more accurately a virtual machine based simulation environment. Robert Henschel
FutureGrid Robert Henschel
Benchmarking Setup • SPEC MPI2007 medium • NSF Acceptance Test Suite • HPCC • Windows HPC Server 2008 • RedHat Enterprise Linux 5.4 • IBM System x iDataPlex dx340 cluster • Intel Compiler Version 11.1 (incl. MKL) • OpenMPI 1.3.1 / MS MPI 2008 Robert Henschel
Hardware • IBM System x iDataPlex dx340 – 84 nodes – 64 used for benchmarking • Intel Xeon L5420 at 2.5 GHz • 32 GByte of memory per node • Mellanox MT26418 DDR Infiniband • Cisco SFS 7024D switch Robert Henschel
Benchmarks • SPEC MPI2007 (medium) • NSF Acceptance Test Suite • HPCC Robert Henschel
SPEC MPIM2007 • Developed by the SPEC High Performance Group • Includes 13 MPI parallel applications • Computational fluid dynamics, molecular dynamics Electromagnetism, geophysics, ray tracing, and hydrodynamics • Version 1.1, released October 2008 • Results were published on the SPEC website, after having been reviewed by the HPG • http://www.spec.org/mpi2007/results/res2010q1/ • First published results running SPEC MPIM2007 on Windows (HPC Server 2008) Robert Henschel
SPEC MPIM2007 • Overall performance SPEC MPIM2007 Score from 32 to 512 Cores 30 25 SPEC MPIM2007 Score 20 15 10 5 0 32 64 128 256 512 Cores WinHPC RHEL5 Robert Henschel
SPEC MPIM2007 • No Difference in performance, 6 applications SPEC MPIM2007 Score for TACHYON 60 50 SPEC MPIM2007 Score 40 30 20 10 0 32 64 128 256 512 Cores WinHPC RHEL5 Robert Henschel
SPEC MPIM2007 • RHEL5 outperforms WinHPC in 5 applications SPEC MPIM2007 Score for SOCORRO 25 SPEC MPIM2007 Score 20 15 10 5 0 32 64 128 256 512 Cores WinHPC RHEL5 Robert Henschel
SPEC MPIM2007 • RHEL5 outperforms WinHPC in 5 applications SPEC MPIM2007 Score for FDS4 SPEC MPIM2007 Score for SOCORRO 70 25 60 SPEC MPIM2007 Score 50 SPEC MPIM2007 Score 20 40 15 30 10 20 10 5 0 32 64 128 256 512 0 Cores 32 64 128 256 512 Cores WinHPC RHEL5 WinHPC RHEL5 Robert Henschel
SPEC MPIM2007 • RHEL5 outperforms WinHPC in 5 applications SPEC MPIM2007 Score for POP2 12 SPEC MPIM2007 Score for FDS4 SPEC MPIM2007 Score for SOCORRO 70 10 SPEC MPIM2007 Score 25 60 SPEC MPIM2007 Score 8 50 SPEC MPIM2007 Score 20 6 40 15 4 30 2 10 20 10 0 5 32 64 128 256 512 0 Cores 32 64 128 256 512 0 Cores WinHPC RHEL5 32 64 128 256 512 Cores WinHPC RHEL5 WinHPC RHEL5 Robert Henschel
SPEC MPIM2007 • WinHPC outperforms RHEL5 in 2 applications SPEC MPIM2007 Score for WRF 60 50 SPEC MPIM2007 Score 40 30 20 10 0 32 64 128 256 512 Cores WinHPC RHEL5 Robert Henschel
SPEC MPIM2007 • WinHPC outperforms RHEL5 in 2 applications SPEC MPIM2007 Score for LESLIE3D 30 SPEC MPIM2007 Score for WRF 25 60 SPEC MPIM2007 Score 20 50 SPEC MPIM2007 Score 40 15 30 10 20 5 10 0 32 64 128 256 512 0 Cores 32 64 128 256 512 WinHPC RHEL5 Cores WinHPC RHEL5 Robert Henschel
NSF Acceptance Test Suite • NSF Track 2 program • Sun constellation at TACC, Cray XT5 at University of Tennessee, FutureGrid • Benchmarking Information Referenced in NSF 05-625 “High Performance Computing System Acquisition: Towards a Petascale Computing Environment for Science and Engineering” (2005) • http://www.nsf.gov/pubs/2006/nsf0605/nsf0605.jsp • Initially comprised of HPCC, 6 applications, SPIOBENCH • Now, just HPCC and 4 applications • PARATEC, HOMME, MILC, WRF Robert Henschel
NSF Acceptance Test Suite • MILC, PARATEC, HOMME • WRF not included Ratio of WinHPC to RHEL 1.2 1 0.8 Ratio 0.6 0.4 0.2 0 32 64 128 256 512 Cores RHEL MILC Medium MILC Large WinHPC WinHPC PARATEC HOMME WinHPC WinHPC Robert Henschel
HPCC • Innovative Computing Laboratory at the University of Tennessee • Version 3.1.3 • 3 categories: – Floating Point – Memory – Interconnect Robert Henschel
HPCC • Floating point tests (HPL, G-FFT) Ratio of WinHPC to RHEL 1.2 1 0.8 Ratio 0.6 0.4 0.2 0 8 16 32 64 128 256 512 Cores RHEL HPL WinHPC GFFT WinHPC Robert Henschel
HPCC • Memory performance tests (Random Access, Stream) Ratio of WinHPC to RHEL 3 2.5 2 Ratio 1.5 1 0.5 0 8 16 32 64 128 256 512 Cores RHEL G-Ran- EP-Stream domAccess WinHPC WinHPC Robert Henschel
HPCC • Interconnect performance tests (G-PTRANS, RR- Bandwidth and Latency) Ratio of WinHPC to RHEL 2 1.8 1.6 1.4 1.2 Ratio 1 0.8 0.6 0.4 0.2 0 8 16 32 64 128 256 512 Cores RHEL G-PTRANS RR-Bandwidth RR-Latency WinHPC WinHPC WinHPC Robert Henschel
Conclusion • Overall performance of WinHPC and RHEL5 is almost identical • Certain applications scale better on RHEL5 than on WinHPC for larger core counts, while they perform very similar on smaller core counts • When applications scale better on WinHPC, they do so across all core counts • Building and running the benchmarks is more challenging on WinHPC Robert Henschel
Acknowledgements This document was developed with support from the National Science Foundation (NSF) under Grant No. 0910812 to Indiana University for " FutureGrid: An Experimental, High- Performance Grid Test-bed. " Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the NSF. This material is based upon work supported by the National Science Foundation under Grants No. ACI-0338618l, OCI- 0451237, OCI-0535258, OCI-0504075, OCI-0910812 and CNS- 0521433. This research was supported in part by the Indiana METACyt Initiative. The Indiana METACyt Initiative of Indiana University is supported in part by Lilly Endowment, Inc. Robert Henschel
Thank You Questions? Robert Henschel
Recommend
More recommend
