Presentation
Z-Paraswap: Co-Optimizing ZNS Parallelism and Swapping for Heterogeneous Graph Neural Network
DescriptionZoned Namespace (ZNS) solid-state drives (SSDs) have been widely adopted in database systems, file systems, and large-scale data centers because they expose part of the physical storage layout to the host system. By leveraging this visibility, host software can access correlated data in parallel and eliminate valid-data copying during garbage collection, thereby improving I/O efficiency. To further integrate ZNS SSDs into memory management, prior work developed a Linux-based ZNS swapping mechanism that reserves multiple zones as swap space, extending virtual memory capacity. However, when deploying Heterogeneous Graph Neural Networks (HetGNNs), this mechanism fails to fully exploit ZNS parallelism due to HetGNN's highly irregular and non-sequential access patterns, which lead to chip-level congestion and long-tail latency. To address this issue, we propose Z-ParaSwap, a ZNS-based Parallelism-Aware Swapping Management framework for HetGNN applications. Z-ParaSwap identifies access correlations in graph data and distributes highly correlated pages across different chips to enhance parallelism and mitigate I/O contention. Experimental results show that Z-ParaSwap reduces average swapping latency by 34% and tail latency by 44%, significantly improving overall HetGNN execution efficiency on ZNS-based systems.
Event Type
Research Manuscript
TimeWednesday, July 2911:15am - 11:30am PDT
LocationMtg Room 202C