When "Grain Size" Doesn't Matter

Abstract

We describe insights gained from putting a quantum scattering problem on two very different parallel architectures: MasPar MP-I (massively parallel) and nCUBE 2 (moderately parallel). Our nearly trivial port from the SIMD MasPar to the MIMD nCUBE demonstrates that it is not categorically difficult to move software from one parallel architecture class to another. These machines show widely different processor and problem grain sizes. Their performance is strikingly similar on mal l problems, a fact not predicted by machine grain size, problem grain size, or peak speed comparisons. We introduce a new metric, fixed-time efficiency, that correlates very well with our experiments and has predictive value. Data and control decomposition and communication considerations are analyzed for each machine.