Performance enhancement of NUMA multiprocessor systems with on-demand memory migration

摘要

The quality of the scheduling has a strong impact on the overall application performance because of process and data affinities. However, this issue is now becoming critical due to the variable memory access latencies in NUMA (Non-Uniform Memory Access) architecture, because in NUMA architecture local data access being significantly faster than remote access, then data locality emerges as a critical criterion for scheduling threads and processes, and it becomes important to be able migrate memory together with their accessing tasks. To perform memory migration, we present memory migration on-demand policy to enable automatic dynamic migration of pages with low cost when they are actually accessed by a task. We use PTE flag setup with the help of madvise system call and the corresponding Copy-on-Touch code added in the page-fault handler which allocates the specific page near the accessing task.