On 4/16/23 21:19, Kirill A. Shutemov wrote: > On Mon, Apr 03, 2023 at 04:42:54PM +0200, Vlastimil Babka wrote: >> Hmm yeah it can be noisy. Did you try to only count events that have >> fragmenting=1 and/or MIGRATE_MOVABLE as fallback_migratetype? As those are >> the really bad events. > > I finally got around to retest it. > > total fragmenting movable fragmenting&&movable > base-1: 957 583 353 0 > base-2: 2715 2343 359 0 > base-3: 2033 1669 353 0 > patched-1: 1325 929 371 0 > patched-2: 2844 2451 371 0 > patched-3: 1304 917 361 0 > > fragmenting=1 is defined as fallback_order<pageblock_order which is most > of them. > > Patched kernel showed slightly elevated movable(fallback_migratetype=1) > cases. Is it critical? Maybe it's still not statistically significant anyway, also not as cricical as fragmenting&movable. > There's no allocations that is fragmenting and movable. Hm. It probably means your test wasn't stressfull enough to inflict a mix of rapid movable an unmovable allocations when memory is nearly full. But at that point the memory is all accepted, so we don't need such scenario. The important thing is that this kind of events didn't start happening during the gradual memory accepting phase.
