Workshops Archive

A Multi-Layer Workflow for Diagnosing Performance Portability in RAJA Applications


Workshop: 2026 International Workshop on Performance, Portability, and Productivity in HPC

Authors: Befikir Bogale and Ian Lumsden (University of Tenessee), Stephanie Brink and Olga Pearce (Lawrence Livermore National Laboratory), and Michela Taufer (University of Tennessee)

Abstract: Performance-portable frameworks like RAJA enable a single source to run on diverse HPC architectures, but productivity relies on the compiler removing abstraction overhead. When it does not, runtime data show where performance diverges, but not why. We present a multi-layer workflow that diagnoses abstraction-induced divergence by combining runtime measurements with compiler-visible evidence, including GPU resource usage, optimization remarks, and LLVM IR metrics. We apply this workflow to 55 RAJA Performance Suite kernels using HIP and OpenMP backends, three optimization levels, and three problem sizes (1,980 configurations). For selected kernels, the workflow distinguishes optimization-agnostic from optimization-sensitive behavior and links performance losses to the compiler's excessive loop unrolling, register pressure, reduced occupancy, scratch spills, and control-flow complexity. These diagnostics guide source changes that restore native-level performance for four kernels and an optimization-specific transformation that eliminates compiler-induced register pressure in a fifth, providing guidance for understanding and improving performance portability beyond runtime analysis.


Back to 2026 International Workshop on Performance, Portability, and Productivity in HPC Archive Listing

Back to Full Workshop Archive Listing