Workshop: ExaMPI26: Workshop on Extreme-Scale MPI
Authors: Kunting Qi (University of Illinois Chicago); Hui Zhou (Argonne National Laboratory); Ke Fan (Temple University); Michael Papka (Argonne National Laboratory, University of Illinois Chicago); Sidharth Kumar (University of Illinois at Chicago); and Rajeev Thakur (Argonne National Laboratory)
Abstract: Machine-learning autotuning selects MPI collective algorithms by benchmarking candidates and predicting the fastest for each execution configuration. Existing methods often assume that all candidates are valid throughout the configuration space and that measurements reflect their actual execution paths. These assumptions fail when restrictions cause silent fallback or compositional algorithms invoke nested collectives whose performance depends on the policy being trained. We present an execution-path- and restriction-aware framework for MPICH \texttt{MPI_Bcast}. It partitions the configuration space by algorithm restrictions, trains corresponding submodels, and combines them into a unified runtime policy. A multistage procedure first trains an intermediate policy using independent algorithms, then benchmarks compositional algorithms under that policy to build the final model. On Polaris, the intermediate policy improves \texttt{smp} performance by up to $2.275\times$; the final policy achieves speedups of up to $30.71\times$ over default MPICH and $2.76\times$ over ACCLAiM.
Back to ExaMPI26: Workshop on Extreme-Scale MPI Archive Listing
Back to Full Workshop Archive Listing