Posted to the Linux kernel mailing list this week was the newest revision of a patch series originally started in early 2025 by a NVIDIA engineer for accelerating page migration. Now being worked on by AMD engineers, this accelerated page migration via batch copies and hardware offloading continues to show promising results.

AMD engineer Shivank Garg posted the latest request for comments (RFC) patches on speeding up page migration in the Linux kernel by batching folio-copy operations and providing accelerated hardware offloading using the DMA offload interface. The intent is for overcoming the current Linux kernel behavior of single-threaded, folio-by-folio copies in the kernel. As part of the patch series is an example/reference DMA offload driver "DCBM"

With the new RFC v5 patches, various code improvements have been made and the patches re-based against the Linux 7.1-rc1 Git state. The performance is reportedly unchanged from the v4 patches, which showed much potential from this patch series when tested on an EPYC Zen 3 server:

Those interested can find these latest accelerated page migration patches on the [1]Linux kernel mailing list .



[1] https://lore.kernel.org/lkml/20260428155043.39251-2-shivankg@amd.com/