Workshop: PDSW’26: The 11th International Parallel Data Systems Workshop
Authors: Matthew Gardner, James Harris, and Revanth Munugala (Northern Arizona University); Tasmia Jannat and Satish Puri (Missouri University of Science and Technology); and Michael Gowanlock (Northern Arizona University)
Abstract: Sorting is fundamental in many data-intensive applications, and its memory-bound nature limits parallel efficiency on parallel processors. Processing-in-Memory (PIM) architectures co-locate DIMM Processing Units (DPUs) with memory and are designed to mitigate the memory bottleneck, but their distributed memory model, lack of peer-to-peer communication between DPUs, and lightweight cores yield several challenges. We present PIM-Sort, a comparison-based sorting algorithm designed for large-scale execution on the UPMEM PIM platform, the first commercially produced PIM system. We design PIM-Sort to address the architectural challenges above, and evaluate it on a UPMEM system with 2546 operational DPUs across uniform and skewed inputs up to 155 GiB. PIM-Sort outperforms a parallel CPU baseline on a separate, higher-bandwidth platform by up to 1.24x on 79 GiB of uniformly distributed data (the largest input that fits in a single DPU wave), while the multi-core CPU remains faster on heavily duplicated inputs where cache reuse yields a performance advantage. We find that PIM-Sort is more energy efficient (up to 35%) than the parallel CPU algorithm even when the latter strongly benefits from cache reuse.
Back to PDSW’26: The 11th International Parallel Data Systems Workshop Archive Listing
Back to Full Workshop Archive Listing