Predicting Price/Performance Trade-offs for Whitney: A Commodity Computing Cluster

We couple simple performance models with pricing to optimize the design of clusters built from commodity components for scientific computing. We apply this technique using the NAS Parallel Benchmarks as a representative workload. We develop models of the BT, LU, and SP benchmarks. The models consist of closed form expressions based on problem size, number of processors, and three measured quantities (single processor performance, network latency, and network bandwidth). These models predict benchmark performance to within 30%. This technique was used in the design of Whitney, a commodity computing cluster at NASA Ames Research Center. In particular, for systems costing less than $1,000,000, the performance characteristics of Intel Pentium processors are better matched to the slower (and less expensive) Fast Ethernet, than to the faster (and more expensive) Myricom Myrinet.