What Size Cluster Equals a Dedicated Chip

摘要

Poisson-Boltzmann calculations within the framework of the Boundary Element Method are parallelized with MPI. The approach is shown to be numerically accurate and of reasonable computational cost. Parallel performance is investigated on two different types of architectures, a cluster of Alphas ES40, 667 MHz EV67, and an IBM Regatta 690, 1.3 GHz Power 4. The obtained scaling characteristics deviate only slightly from the theoretical scaling due to Amdahl's Law. An alternative version is also implemented that makes use of a specialized chip for the core-calculational part. The raw-performance of this specialized ASIC, MD-GRAPE-2, is compared to the parallel performance gained via MPI. It is estimated that for a medium-sized test-case one would have to operate a parallel machine composed of roughly 330 CPUs to equal the performance of 1 single specialized device.