Perseus- Hardware

Perseus is a high-performance Beowulf cluster.

The Beowulf cluster was built entirely from inexpensive commodity components. It consists of 116 dual-Pentium PCs (i.e. 232 processors) connected on a private subnet with 100 Mbit/s Fast Ethernet using Intel 510 switches. An initial prototype system made up of 16 nodes, 8 with dual 400MHz Pentium IIs and 8 with dual 450MHz Pentium IIIs, was constructed in early 1999. The additional 100 nodes, which were purchased at the end of 1999, contain dual 500MHz Pentium IIIs, 256 MBytes of memory and 10 GBytes of local disk. Some of the nodes have been given additional memory to handle large problem sizes. An additional 3 nodes are used as front ends to interface to the cluster, and to host the user file systems and provide backup capabilities.

Designing the Cluster

A range of economic as well as technical issues are involved in the design of a Beowulf cluster for a particular set of applications. This requires some analysis to find the most cost-effective system that will satisfy the usage requirements. There are many trade-offs in the various configuration parameters of a Beowulf cluster, including the type of processor, processor speed, memory, disk, single/dual/quad processor, network connectivity, etc.

To help with the design analysis and provide a proof of concept, we used a prototype 486 Beowulf cluster and some Pentium PCs to experiment with Beowulf hardware and software, and to run some performance benchmarks using Gaussian and some standard benchmark codes.

We analysed the memory, disk and CPU usage requirements of the applications to be used on the cluster, and the price/performance tradeoffs, to come up with the choice of configuration for the PCs in the cluster. We also found that a Fast Ethernet network provided adequate latency and bandwidth for the user requirements, giving good speedups for multi-processor GAMESS or NAMD jobs and allowing high throughput of multiple single or dual-processor Gaussian jobs.

Further information on the design process and some benchmark results for computational chemistry codes can be found in technical report DHPC-073.

The DHPC group also put together the cluster from the commodity components, which is shown in the mandatory Beowulf installation photo gallery. The racks holding the cluster were built by the chemistry departmental workshop.

Technical Reports

1. DHPC-061: Beowulf - A New Hope for Parallel Computing?, K.A. Hawick, D.A. Grove and F.A. Vaughan, January 1999.

2. DHPC-073: Commodity Cluster Computing for Computational Chemistry, K.A. Hawick, D.A. Grove, P.D. Coddington, and M.A. Buntine, January 2000.

For more information, contact Paul Coddington/A> in the Computer Science Department or Mark Buntine/A> in the Chemistry Department.

Perseus User's Guide