Performance Comparison of Scientific Applications on Linux and Windows HPC Server Clusters

Abstract

The TOP500 project ranks the 500 th. most powerful computer systems in the world. This rank is based on High-Performance LINPACK (HPL), a portable implementation of the LINPACK benchmark for distributed-memory computers. According to the TOP500’s list published in June, 2010, most of the systems are Linux based clusters. However, the number of Windows based systems tends to grow in the next lists. In this paper the performance of a Linux and a Windows HPC Server cluster using the same hardware and scientific applications is evaluated. To assess performance, the HPL and NAS Parallel Benchmarks (NPB), as well as a real-world multi-physics application, named EdgeCFD, are used. This application is developed at NACAD-COPPE/UFRJ and it is an implicit edge-based coupled fluid flow and transport solver for large-scale problems optimized for modern clusters. EdgeCFD adopts peer-to-peer non-blocking communication pattern among processes. Tests were conducted for three MPI distributions MVAPICH2 (based on MPICH2 with Infiniband support), OpenMPI (High Performance Message Passing Library) and MS-MPI (Microsoft MPI based on MPICH2) on all benchmarks and applications of this study. All performance measurements show that Windows HPC Server is a viable option.