Measurement and Analysis of Application Performance on Exascale GPU-accelerated Systems

This talk from the SPCL_Bcast series features John Mellor-Crummey from Rice University discussing the extension of HPCToolkit performance tools for exascale GPU-accelerated supercomputers. Learn how the HPCToolkit project team influenced hardware support for instruction-level performance measurement in AMD, Intel, and NVIDIA GPUs. Discover the innovative techniques employed, including PC sampling, binary instrumentation, wait-free data structures, parallel analysis of large binaries, and distributed-memory parallelism to handle terabytes of performance data. Explore how these strategies have enabled efficient measurement and analysis of applications running on up to 64K MPI ranks and 64K GPU tiles on ORNL's Frontier supercomputer. The presentation covers key aspects of HPCToolkit, successful application analyses, and future challenges in performance measurement for exascale computing. Recorded on March 13, 2025, as part of the SPCL_Bcast #56 series from ETH Zurich's Scalable Parallel Computing Lab.