Workshop: 21st Workshop on Workflows in Support of Large-Scale Science (WORKS26)
Authors: Kevin Assogba (Rochester Institute of Technology), Nigel Tan (Los Alamos National Laboratory), M. Mustafa Rafique (Rochester Institute of Technology), Michela Taufer (University of Tennessee Knoxville), and Bogdan Nicolae (Argonne National Laboratory)
Abstract: Data-intensive workloads in scientific computing, e.g., in-situ and ad-hoc data processing for numerical reproducibility analysis, exhibit I/O patterns dominated by small, non-contiguous reads from large files. Although many HPC frameworks overlap I/O with computation, existing interfaces, e.g., POSIX and io_uring, are primarily optimized for sequential or large-block access. Workloads issuing many random reads therefore incur high system-call overhead, inefficient batching, and underutilized bandwidth. We present SCISSOR, an I/O runtime for workloads dominated by fine-grained random access. SCISSOR applies gap-aware offset coalescing to merge nearby requests into contiguous segments, reducing kernel interactions and increasing effective I/O size. A regression-based performance model selects batching and coalescing parameters that balance I/O and computation throughput. Finally, SCISSOR leverages io_uring to issue large asynchronous batches with minimal context-switch overhead. On HPC reproducibility analytics over large-scale cosmology simulations (HACC), SCISSOR achieves 3.6× and 42× higher end-to-end throughput on average over io_uring and POSIX memory-mapped baselines, respectively.
Back to 21st Workshop on Workflows in Support of Large-Scale Science (WORKS26) Archive Listing
Back to Full Workshop Archive Listing