The problem
When does moving a page to where it’s used improve locality, and when does migration become the work?
What current research shows
CUDA exposes managed memory and a unified address model, while device performance guidance emphasizes minimizing host-device transfers and improving overlap. A page that moves across the interconnect adds latency and bandwidth demand.
Where the evidence stops
Migration behavior depends on access pattern, residency, prefetching, page size, interconnect, and concurrent users.
What Valen Systems is testing
Report page faults, migration volume, transfer time, residency, end-to-end work, and energy as workload behavior changes.