Application Performance on Commodity-class Computers: The Impact of Cluster Architecture.

Martyn F. Guest and C.A. Kitchen

Computational Science and Engineering Department, CCLRC Daresbury Laboratory, Daresbury, Warrington WA4 4AD, UK

In many application areas, commodity-based clusters provide an established, cost effective solution to the provision of High Performance Computing. The ability of such clusters to address the differing demands from the fields of Capability (high-end) and
Capacity (mid-range) computing is, however, very much a function of both application domain and detailed cluster architecture - node and interconnect.

In this presentation we compare the performance of a variety of clusters with current high-end hardware, such as the IBM p5-575 series characterizing the HPCx system at Daresbury, the SGI Altix 3700, HP's Superdome and the Cray XT3. We overview performance from a variety of commodity-based systems (CS1-CS36), featuring IA32, x86-64 and IA64 architectures, coupled to a variety of interconnect technologies - Gbit Ethernet, Myrinet, SCI, Infiniband, Infinipath and Quadrics QSNet. Such systems dominate current mid-range, University-based expenditure in the UK through HEFCE's Scientific Research Investment Fund (SRIF) - expenditure that is comparable with that associated with high-end systems such as HPCx and HECToR.

Our focus lies in applications and concentrates on the areas of Chemistry, Materials, Engineering, Environmental modelling and Astrophysics. Based on simple metrics, we consider the performance of variety of codes from each of these disciplines, including NWChem and GAMESS-UK, CPMD, DLPOLY, ANGUS and PDNS3D, GADGET and POLCOMS. We identify the challenges in scaling to large processor counts, the associated bottlenecks as a function of discipline, and provide a guide to the optimum resource that might be deployed in a given application domain.

back