Fast and Faithful Performance Prediction of MPI Applications: the HPL Case Study

International audienceFinely tuning MPI applications (number of processes, granularity, collectiveoperation algorithms, topology and process placement) is critical to obtain good performance onsupercomputers. With a rising cost of modern supercomputers, running parallelapplications at scale solely to optimize their performance is extremelyexpensive. Having inexpensive but faithful predictions of expected performancecould be a great help for researchers and system administrators.The methodology we propose captures the complexity of adaptive applications byemulating the MPI code while skipping insignificant parts. We demonstrate itscapability with High Performance Linpack (HPL), the benchmark used to ranksupercomputers in the TOP500 and which requires a careful tuning. We explain(1) how we both extended the SimGrid's SMPI simulator and slightly modified theopen-source version of HPL to allow a fast emulation on a single commodityserver at the scale of a supercomputer and (2) how to model the differentcomponents (network, BLAS, ...) of the system. We show that a carefulmodeling of both spatial and temporal node variability allows us to obtainpredictions within a few percents of real experiments (see Figure 1)