June 16, 2026 · papers
Stride-aware GPU page prefetching presented at ISMM 2026
Unified memory makes GPUs easier to program and slower to run. Yunqi Shen presented our recovery of that lost performance at the ACM SIGPLAN International Symposium on Memory Management in Boulder: a prefetcher that predicts access patterns with a Markov model instead of reacting to page faults once the stall has already happened.