Description of problem: Currently, page migration breaks 2MB pages into 4kB pages, which then have to be reassembled again at the destination. This negates a lot of the performance benefits of doing NUMA relocations in the first place. 2MB transparent hugepages should stay intact across page migration. We can use the same bit that is used to protect 2MB->4kB and 4kB->2MB transitions to block out page accesses during a 2MB->2MB transition.
THP-preserving migration is included in the NUMA balancing code.
Johannes, if you look unmap_and_move, which is indirectly called from sys_migrate_pages, you will see that it does NOT preserve THP on migration. There is a reason this bug was opened separately.
(In reply to Rik van Riel from comment #6) > Johannes, > > if you look unmap_and_move, which is indirectly called from > sys_migrate_pages, you will see that it does NOT preserve THP on migration. > There is a reason this bug was opened separately. My bad, I was under the assumption this was to track the side-project of adding THP-preserving migration to AutoNUMA. I'll dig up the patches.
This request was resolved in Red Hat Enterprise Linux 7.0. Contact your manager or support representative in case you have further questions about the request.
The comment above is incorrect. The correct version is bellow. I'm sorry for any inconvenience. --------------------------------------------------------------- This request was NOT resolved in Red Hat Enterprise Linux 7.0. Contact your manager or support representative in case you need to escalate this bug.