TEMPI: An Interposed MPI Library with a Canonical Representation of CUDA-aware Datatypes

MPI Derived Datatypes are an abstraction that simplifies handling of non-contiguous data in MPI applications. These datatypes are recursively constructed at runtime from primitive Named Types defined in the MPI standard. More recently, the development and deployment of CUDA-aware MPI implementations has encouraged the transition of distributed high-performance MPI codes to use GPUs. These implementations allow MPI functions to directly operate on GPU buffers, easing integration of GPU compute into MPI codes. Despite substantial attention to CUDA-aware MPI implementations, they continue to offer cripplingly poor GPU performance when manipulating derived datatypes on GPUs. This work presents a new MPI library, TEMPI, to address this issue. TEMPI first introduces a common datatype to represent equivalent MPI derived datatypes. TEMPI can be integrated into existing MPI deployments as an interposed library. This library can be used regardless of MPI deployment and without modifying application code. Furthermore, this work presents a performance model of GPU derived datatype handling, demonstrating that previously preferred "one-shot" methods are not always fastest. Ultimately, the interposed-library model of this work demonstrates MPI_Pack speedup of up to 720,400x and MPI\_Send speedup of up to 59,000x compared to the MPI implementation deployed on a leadership-class supercomputer. This yields speedup of more than 1000x in a 3D halo exchange at 192 ranks.