Accelerated Out of Core Shuffling
Summary
The article discusses RapidsMPF, an out-of-core shuffler designed for large-scale data analytics, and explains why shuffles are memory, transport, and synchronization bottlenecks. It presents benchmarking results on a DGX B200 setup, including simple and scaled-spilling tests, and highlights how host-side buffering and streaming can reduce spills while achieving high throughput; it also notes a throughput reporting quirk and outlines implications for building fast, scalable ETL pipelines.