Workshop: PDSW’26: The 11th International Parallel Data Systems Workshop
Authors: Baodi Shan (Stony Brook University), Mauricio Araya-Polo (TotalEnergies), and Barbara Chapman (Stony Brook University)
Abstract: Accelerator applications must move data between GPUs while computation continues, yet conventional host mediation delays transfer submission until control returns from an OpenMP target region. GiOMP moves the initiation point into GPU execution, so movement begins at data readiness and overlaps independent work. Using OpenMP target offloading, GiOMP provides device callable PUT, GET, and completion operations over HPE Slingshot. Its proxy path sends dynamic requests through a CPU worker, while its GPU trigger path releases prepared network operations without later CPU submission. On Tioga, a fused pipeline on two nodes achieves up to a 1.23× speedup over GPU aware MPI, with 5.6 to 6.8 microseconds attributed to eliminating a target relaunch. Jacobi weak scaling reaches 1.87× on Tioga and 1.82× on Delta AI. Remote GPU memory is the first concrete target of a broader GPU driven data movement layer, with future investigation of storage facing staging and in transit consumers.
Back to PDSW’26: The 11th International Parallel Data Systems Workshop Archive Listing
Back to Full Workshop Archive Listing