On Wed, 5 Dec 2018, Michal Hocko wrote: > > It isn't specific to MADV_HUGEPAGE, it is the policy for all transparent > > hugepage allocations, including defrag=always. We agree that > > MADV_HUGEPAGE is not exactly defined: does it mean try harder to allocate > > a hugepage locally, try compaction synchronous to the fault, allow remote > > fallback? It's undefined. > > Yeah, it is certainly underdefined. One thing is clear though. Using > MADV_HUGEPAGE implies that the specific mapping benefits from THPs and > is willing to pay associated init cost. This doesn't imply anything > regarding NUMA locality and as we have NUMA API it shouldn't even > attempt to do so because it would be conflating two things.
This is exactly why we use MADV_HUGEPAGE when remapping our text segment to be backed by transparent hugepages, we want to pay the cost at startup to fault thp and that involves synchronous memory compaction rather than quickly falling back to remote memory. This is making the case for me. > > So to answer "what is so different about THP?", it's the performance data. > > The NUMA locality matters more than whether the pages are huge or not. We > > also have the added benefit of khugepaged being able to collapse pages > > locally if fragmentation improves rather than being stuck accessing a > > remote hugepage forever. > > Please back your claims by a variety of workloads. Including mentioned > KVMs one. You keep hand waving about access latency completely ignoring > all other aspects and that makes my suspicious that you do not really > appreciate all the complexity here even stronger. > I discussed the tradeoff of local hugepages vs local pages vs remote hugepages in https://marc.info/?l=linux-kernel&m=154077010828940 on Broadwell, Haswell, and Rome. When a single application does not fit on a single node, we obviously need to extend the API to allow it to fault remotely. We can do that without changing long-standing behavior that prefers to only fault locally and causing real-world users to regress. Your suggestions about how we can extend the API are all very logical. [ Note that is not the regression being addressed here, however, which is massive swap storms due to a fragmented local node, which is why the __GFP_COMPACT_ONLY patch was also proposed by Andrea. The ability to prefer faulting remotely is a worthwhile extension but it does no good whatsoever if we can encounter massive swap storms because we didn't set __GFP_NORETRY appropriately (which both of our patches do) both locally and now remotely. ]