GPHash: An Efficient Hash Index for GPU with Byte-Granularity Persistent Memory

This conference talk presents GPHash, an efficient hash index designed specifically for GPU systems with byte-granularity persistent memory (GPM). Learn how researchers from Huazhong University of Science and Technology address the inefficiencies of conventional hash indexes in GPM systems by tackling three key challenges: warp-agnostic execution, high-overhead consistency guarantees, and the significant bandwidth gap between persistent memory and GPU. Discover how GPHash implements a lock-free, warp-cooperative execution model for all index operations, ensures consistency with minimal overhead using CAS primitive and slot states, and bridges the bandwidth gap by intelligently caching hot items in GPU memory. The presentation highlights evaluation results showing GPHash outperforming state-of-the-art CPU-assisted data management approaches and existing GPM hash indexes by up to 27.62× on YCSB and real-world workloads.