Workshop: 17th Workshop on Latest Advances in Scalable Algorithms for Large-Scale Heterogeneous Systems (ScalAH'26)
Authors: David Díez-Poza, Sergio Alonso-Pascual, and Arturo Gonzalez-Escribano (Universidad de Valladolid)
Abstract: Exascale computing requires the development and execution of applications with massive scalability capabilities. Iterative Stencil Loop (ISL) applications are an important class of scientific applications with high scalability potential. EPSILOD is a tool designed to simplify the development and execution of ISL applications in heterogeneous distributed environments. This work proposes and evaluates new techniques in EPSILOD to optimize data communication in distributed ISL applications running on multi-GPU, multi-node systems. Specifically, it explores multidimensional partitioning policies and techniques to optimize MPI communications, including both host-staging and GPU-aware MPI, by automatically creating data contiguity in memory based on the chosen partitioning policy. Experimental results are presented using up to 1,024 NVIDIA GPUs across 256 nodes and up to 8,192 AMD GPUs across 1,024 nodes, showing significant improvements in strong and weak scalability across various scenarios and ISL applications.
Back to 17th Workshop on Latest Advances in Scalable Algorithms for Large-Scale Heterogeneous Systems (ScalAH'26) Archive Listing
Back to Full Workshop Archive Listing