Simulation-based optimization and sensibility analysis of MPI applications: Variability matters

Finely tuning MPI applications and understanding the influence of key parameters (number processes, granularity, collective operation algorithms, virtual topology, process placement) is critical to obtain good performance on supercomputers. With high consumption running at scale, doing so solely optimize their particularly costly. Having inexpensive but faithful predictions expected could be a great help for researchers system administrators. The methodology we propose decouples complexity platform, which captured through statistical models its main components (MPI communications, BLAS operations), from adaptive by emulating application skipping regular non-MPI parts code. We demonstrate capability our method with High-Performance Linpack (HPL), benchmark used rank supercomputers in TOP500, requires careful tuning. briefly present (1) how open-source version HPL can slightly modified allow fast emulation single commodity server scale supercomputer. Then (2) an extensive (in)validation study that compares simulation real experiments demonstrates ability predict within few percent consistently. This allows us identify modeling pitfalls (e.g., spatial temporal node variability or network heterogeneity irregular behavior) need considered. Last, show (3) “surrogate” studying several subtle parameter optimization problems while accounting uncertainty platform.