WIC - Hiding Producer-Consumer Synchronization Delays with Warp-Level Interrupt-based GPU Communications

Learn about a novel GPU communication synchronization framework through this 20-minute conference presentation from USENIX ATC '25. Discover how traditional synchronization methods create bottlenecks in GPU communication due to manual coordination requirements between producers and consumers, where consumers frequently poll for data availability leading to unnecessary computational resource occupation. Explore the Warp-level Interrupt-based Communication (WIC) framework, which introduces a fine-grained interruption mechanism at the warp level to replace repetitive polling by preemptively stalling warps engaged in frequent polling and releasing computational resources for other warps. Understand how WIC effectively overlaps producer-consumer synchronization with ongoing computations, achieving 1.13× performance improvement on average across various applications with diverse communication patterns compared to conventional polling methods. Gain insights into GPU communication challenges, synchronization bottlenecks, and innovative solutions for collaborative computation across multiple devices from researchers at Xi'an Jiaotong-Liverpool University, University of Liverpool, Beijing University of Posts and Telecommunications, and NingboTech University.