Workshop: 2026 International Workshop on Performance, Portability, and Productivity in HPC
Authors: Avitay Geltman (Nuclear Research Center - Negev, Israel); Shahar Maimon (Faculty of Computer Science, Technion - Israel Institute of Technology, Israel); Yehonatan Fridman (Faculty of Computer and Information Science, Ben-Gurion University of the Negev, Israel); Ron Wagner (School of Physics and Astronomy, Tel Aviv University, Israel); Shay Shani (School of Computer Science and AI, Tel Aviv University, Israel); and Re'em Harel (Nuclear Research Center - Negev, Israel; Ben-Gurion University of the Negev)
Abstract: Directive-based offload enables portable GPU execution from one source, yet evaluations usually use microbenchmarks or compact kernels. We extend SCALSALE, a Modern Fortran full multi-material Lagrangian--Eulerian hydrodynamics proxy application, to multi-node GPU execution using MPI+OpenMP without vendor-specific paths. This required restructuring branch-heavy Eulerian advection and replacing host-staged halo exchange with GPU-aware MPI over device-resident buffers. Tests on NVIDIA A100 and Intel Max 1550 GPUs and an Intel Xeon Platinum 8480+ CPU yield weak-scaling efficiencies of 84%-91% within a GPU node, 88%-91% acrosss three GPU nodes, and higher 95%-96% across CPU nodes because longer computation amortizes communication. Using Pennycook's metric with hierarchical-Roofline efficiencies, performance-portability scores are 34.9% (Eulerian) and 37.9% (Lagrangian) over all platforms, and 44.1% and 42.2% over the GPUs, respectively. Streaming kernels attain 63%-83% of their ceilings across architectures, whereas the dominant advection sweep ranges from 84% to 13%, identifying the optimization target.
Back to 2026 International Workshop on Performance, Portability, and Productivity in HPC Archive Listing
Back to Full Workshop Archive Listing