Reliable performance prediction for multigrid software on distributed memory systems

摘要: We propose a model for describing and predicting the parallel performance of broad class numerical software on distributed memory architectures. The purpose this is to allow reliable predictions be made large numbers processors given system, by only benchmarking code small processors. Having described methods used, emphasized simplicity their implementation, approach tested range engineering applications that are built upon use multigrid algorithms. Despite simplicity, models demonstrated provide both accurate robust across different architectures, partitioning strategies codes. In particular, effectiveness predictive methodology shown practical implementation an elastohydrodynamic lubrication solver.