Section: New Results

Predictive Simulation of HPC Applications

Finely tuning MPI applications (number of processes, granularity, collective operation algorithms, topology and process placement) is critical to obtain good performance on supercomputers. With a rising cost of modern supercomputers, running parallel applications at scale solely to optimize their performance is extremely expensive. Using SimGrid, we work toward providing a methodology allowing to provide inexpensive but faithful predictions of expected performance.

The methodology we propose relies on SimGrid/SMPI and captures the complexity of adaptive applications by emulating the MPI code while skipping insignificant parts. In [18] we demonstrate its capability with High Performance Linpack (HPL), the benchmark used to rank supercomputers in the TOP500 and which requires a careful tuning. We explain (1) how we both extended the SimGrid's SMPI simulator and slightly modified the open-source version of HPL to allow a fast emulation on a single commodity server at the scale of a supercomputer and (2) how to model the different components (network, BLAS, ...) of the system. We show that a careful modeling of both spatial and temporal node variability allows us to obtain predictions within a few percents of real experiments. The modeling of BLAS operations is particularly important and we have thus started investigating in the context of simulating a sparse direct solver how to automatically performance models for commonly used BLAS kernels [33]. A key difficulty remains the acquisition of faithful performance measurements as modern processors are often quite unstable. This effort is therefore particularly related to the aforementioned "Design of Experiments" line of research.