On 9/30/19 8:56 AM, Stefan Priebe - Profihost AG wrote: > Hi, > > the current status is, that everything works well / fine since i > switched from CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS to > CONFIG_TRANSPARENT_HUGEPAGE_MADVISE Thanks, that indeed confirms the problem is related to THP's. > Am 27.09.19 um 14:45 schrieb Vlastimil Babka: >> On 9/19/19 12:21 PM, Stefan Priebe - Profihost AG wrote: >> >> Hi, could you try the patch below? I suspect you're hitting a corner >> case where compaction_suitable() returns COMPACT_SKIPPED for the >> ZONE_DMA, triggering reclaim even if other zones have plenty of free >> memory. And should_continue_reclaim() then returns true until twice the >> requested page size is reclaimed (compact_gap()). That means 4MB >> reclaimed for each THP allocation attempt, which roughly matches the >> trace data you preovided previously. >> >> The amplification to 4MB should be removed in patches merged for 5.4, so >> it would be only 32 pages reclaimed per THP allocation. The patch below >> tries to remove this corner case completely, and it should be more >> visible on your 5.2.x, so please apply it there. > > so i switched back to 4.19 LTS Kernel - as this is the kernel we run on > all our infrastructures. THP is now only in use von kvm host machines. > Your patch applies to 4.19 as well - but not sure if it is a good idea > to apply it to those machines. If you could try that, it would be great (and switch back hugepages to always after applying). The problem is older than 4.19. > Greets, > Stefan >
