Don't miss out! Join us at the next Open Source Summit in Amsterdam, Netherland (August 25-29); Seoul, South Korea (November 4-5). Join us at the premier vendor-neutral open source conference, where developers and technologists come together to collaborate, share knowledge, and explore the latest innovations and advancements in open source technology. Learn more at https://events.linuxfoundation.org/
Overview of Linux Page Migration and Accelerating It With Multi-threading and DMA Offloading - Shivank Garg, AMD
Page migration is a vital mechanism in Linux NUMA systems, enabling the movement of memory pages across NUMA nodes to optimize performance and resource utilization.
However, its current sequential copying approach often results in performance bottlenecks especially in tiered-memory systems underutilizing modern hardware.
I'll start with an overview of current page migration implementation and its use cases such as NUMA balancing, memory compaction, and demotion. I'll then present my proposed enhancements to accelerate page migration using three key techniques: leveraging batching operations, multi-threading, and DMA offloading, for which RFCs have been posted to the community (https://lore.kernel.org/linux-mm/20250319192211.10092-1-shivankg@amd.com).
My measurements reveal that just copying folios can eat up to 96.7% of migration time for large pages - a massive opportunity for improvement. The microbenchmarks from my approach show throughput improvement by up to 5x for large page migrations.