Reducing Page Faults via Invalidation-Based Mapping Propagation in Multi-GPU Systems

Unified Virtual Memory in multi-GPU systems provides scalable hardware resources with simplified memory management to users. However, its performance is often restricted by high-latency page faults. When a page migrates, the migration mechanism invalidates old mappings on all GPUs but updates the new mapping only on the destination GPU. This leaves other GPUs' page table entries invalid, generating subsequent page faults when the non-destination GPUs attempt to access the migrated page. Additionally, we observe that naively updating all GPUs' entries incurs extra page table walks, resulting in significant latency overhead. To address this, we propose ShadowUpdate, a redesign of the migration handling mechanism to reduce page faults by leveraging the invalidation phase. ShadowUpdate exploits the existing invalidation broadcast to proactively propagate the new mapping simultaneously. By combining invalidation and mapping updates, this approach eliminates redundant page faults and the associated page table walks, accelerating the overall migration process. To ensure correctness, ShadowUpdate also includes a lightweight in-flight migration tracker that holds translation requests during migration, records the new mapping at invalidation, and releases the requests once the page copy completes to prevent invalid accesses. ShadowUpdate improves overall performance by ${1. 4 0} \times$ on average over a baseline UVM design across 14 representative multi-GPU UVM workloads.