Presentation
Phantom Walk: Reducing the Post-Migration Translation Path via Proactive Translation Dissemination in Multi-GPU Systems
DescriptionMulti-GPU systems rely on unified virtual memory to provide a global address space, which inevitably triggers page migrations. Our analysis shows that reducing migrations further is difficult even with advanced policies. These migrations invalidate page table entries (PTEs) across GPUs and generate migration-induced misses (M-misses), which constitute a large portion of all translation requests and significantly degrade performance. This paper introduces Phantom Walk, which accelerates M-miss address translation by reducing GPU-local page faults and page table walks through a shorter post-migration translation path. Our evaluation shows that Phantom Walk improves overall performance by 44.37% on average.
Event Type
Research Manuscript
TimeTuesday, July 283:30pm - 3:42pm PDT
LocationMtg Room 202C