Am 10.09.19 um 11:02 schrieb Michal Hocko: > On Tue 10-09-19 10:38:25, Stefan Priebe - Profihost AG wrote: >> Am 10.09.19 um 10:29 schrieb Michal Hocko: >>> On Tue 10-09-19 07:56:36, Stefan Priebe - Profihost AG wrote: >>>> >>>> Am 09.09.19 um 14:56 schrieb Stefan Priebe - Profihost AG: >>>>> Am 09.09.19 um 14:49 schrieb Michal Hocko: >>>>>> On Mon 09-09-19 14:37:52, Stefan Priebe - Profihost AG wrote: >>>>>>> >>>>>>> Am 09.09.19 um 14:28 schrieb Michal Hocko: >>>>>>>> On Mon 09-09-19 14:10:02, Stefan Priebe - Profihost AG wrote: >>>>>>>>> >>>>>>>>> Am 09.09.19 um 14:08 schrieb Michal Hocko: >>>>>>>>>> On Mon 09-09-19 13:01:36, Michal Hocko wrote: >>>>>>>>>>> and that matches moments when we reclaimed memory. There seems to be a >>>>>>>>>>> steady THP allocations flow so maybe this is a source of the direct >>>>>>>>>>> reclaim? >>>>>>>>>> >>>>>>>>>> I was thinking about this some more and THP being a source of reclaim >>>>>>>>>> sounds quite unlikely. At least in a default configuration because we >>>>>>>>>> shouldn't do anything expensinve in the #PF path. But there might be a >>>>>>>>>> difference source of high order (!costly) allocations. Could you check >>>>>>>>>> how many allocation requests like that you have on your system? >>>>>>>>>> >>>>>>>>>> mount -t debugfs none /debug >>>>>>>>>> echo "order > 0" > /debug/tracing/events/kmem/mm_page_alloc/filter >>>>>>>>>> echo 1 > /debug/tracing/events/kmem/mm_page_alloc/enable >>>>>>>>>> cat /debug/tracing/trace_pipe > $file >>>>>>>> >>>>>>>> echo 1 > /debug/tracing/events/vmscan/mm_vmscan_direct_reclaim_begin/enable >>>>>>>> echo 1 > /debug/tracing/events/vmscan/mm_vmscan_direct_reclaim_end/enable >>>>>>>> >>>>>>>> might tell us something as well but it might turn out that it just still >>>>>>>> doesn't give us the full picture and we might need >>>>>>>> echo stacktrace > /debug/tracing/trace_options >>>>>>>> >>>>>>>> It will generate much more output though. >>>>>>>> >>>>>>>>> Just now or when PSI raises? >>>>>>>> >>>>>>>> When the excessive reclaim is happening ideally. >>>>>>> >>>>>>> This one is from a server with 28G memfree but memory pressure is still >>>>>>> jumping between 0 and 10%. >>>>>>> >>>>>>> I did: >>>>>>> echo "order > 0" > >>>>>>> /sys/kernel/debug/tracing/events/kmem/mm_page_alloc/filter >>>>>>> >>>>>>> echo 1 > /sys/kernel/debug/tracing/events/kmem/mm_page_alloc/enable >>>>>>> >>>>>>> echo 1 > >>>>>>> /sys/kernel/debug/tracing/events/vmscan/mm_vmscan_direct_reclaim_begin/enable >>>>>>> >>>>>>> echo 1 > >>>>>>> /sys/kernel/debug/tracing/events/vmscan/mm_vmscan_direct_reclaim_end/enable >>>>>>> >>>>>>> timeout 120 cat /sys/kernel/debug/tracing/trace_pipe > /trace >>>>>>> >>>>>>> File attached. >>>>>> >>>>>> There is no reclaim captured in this trace dump. >>>>>> $ zcat trace1.gz | sed 's@.*\(order=[0-9]\).*\(gfp_flags=.*\)@\1 \2@' | sort | uniq -c >>>>>> 777 order=1 gfp_flags=__GFP_IO|__GFP_FS|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_NOMEMALLOC >>>>>> 663 order=1 gfp_flags=__GFP_IO|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_NOMEMALLOC >>>>>> 153 order=1 gfp_flags=__GFP_IO|__GFP_NOWARN|__GFP_RETRY_MAYFAIL|__GFP_NORETRY|__GFP_COMP|__GFP_NOMEMALLOC >>>>>> 911 order=1 gfp_flags=GFP_KERNEL_ACCOUNT|__GFP_ZERO >>>>>> 4872 order=1 gfp_flags=GFP_NOWAIT|__GFP_IO|__GFP_FS|__GFP_NOWARN|__GFP_COMP|__GFP_ACCOUNT >>>>>> 62 order=1 gfp_flags=GFP_NOWAIT|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_NOMEMALLOC >>>>>> 14 order=2 gfp_flags=GFP_ATOMIC|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP >>>>>> 11 order=2 gfp_flags=GFP_ATOMIC|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_RECLAIMABLE >>>>>> 1263 order=2 gfp_flags=__GFP_IO|__GFP_FS|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_NOMEMALLOC >>>>>> 45 order=2 gfp_flags=__GFP_IO|__GFP_FS|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_NOMEMALLOC|__GFP_RECLAIMABLE >>>>>> 1 order=2 gfp_flags=GFP_KERNEL|__GFP_COMP|__GFP_ZERO >>>>>> 7853 order=2 gfp_flags=GFP_NOWAIT|__GFP_IO|__GFP_FS|__GFP_NOWARN|__GFP_COMP|__GFP_ACCOUNT >>>>>> 73 order=3 gfp_flags=__GFP_IO|__GFP_FS|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_NOMEMALLOC >>>>>> 729 order=3 gfp_flags=__GFP_IO|__GFP_FS|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_NOMEMALLOC|__GFP_RECLAIMABLE >>>>>> 528 order=3 gfp_flags=__GFP_IO|__GFP_NOWARN|__GFP_RETRY_MAYFAIL|__GFP_NORETRY|__GFP_COMP|__GFP_NOMEMALLOC >>>>>> 1203 order=3 gfp_flags=GFP_NOWAIT|__GFP_IO|__GFP_FS|__GFP_NOWARN|__GFP_COMP|__GFP_ACCOUNT >>>>>> 5295 order=3 gfp_flags=GFP_NOWAIT|__GFP_IO|__GFP_FS|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP >>>>>> 1 order=3 gfp_flags=GFP_NOWAIT|__GFP_IO|__GFP_FS|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_NOMEMALLOC >>>>>> 132 order=3 gfp_flags=GFP_NOWAIT|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_NOMEMALLOC >>>>>> 13 order=5 gfp_flags=GFP_KERNEL|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_ZERO >>>>>> 1 order=6 gfp_flags=GFP_KERNEL|__GFP_NOWARN|__GFP_NORETRY|__GFP_COMP|__GFP_ZERO >>>>>> 1232 order=9 gfp_flags=GFP_TRANSHUGE >>>>>> 108 order=9 gfp_flags=GFP_TRANSHUGE|__GFP_THISNODE >>>>>> 362 order=9 gfp_flags=GFP_TRANSHUGE_LIGHT|__GFP_THISNODE >>>>>> >>>>>> Nothing really stands out because except for the THP ones none of others >>>>>> are going to even be using movable zone. >>>>> It might be that this is not an ideal example is was just the fastest i >>>>> could find. May be we really need one with much higher pressure. >>>> >>>> here another trace log where a system has 30GB free memory but is under >>>> constant pressure and does not build up any file cache caused by memory >>>> pressure. >>> >>> So the reclaim is clearly induced by THP allocations >>> $ zgrep vmscan trace2.gz | grep gfp_flags | sed 's@.*\(gfp_flags=.*\) .*@\1@' | sort | uniq -c >>> 1580 gfp_flags=GFP_TRANSHUGE >>> 15 gfp_flags=GFP_TRANSHUGE|__GFP_THISNODE >>> >>> $ zgrep vmscan trace2.gz | grep nr_reclaimed | sed 's@nr_reclaimed=@@' | awk '{nr+=$6+0}END{print nr}' >>> 1541726 >>> >>> 6GB of memory reclaimed in 1776s. That is a lot! But the THP allocation >>> rate is really high as well >>> $ zgrep "page_alloc.*GFP_TRANSHUGE" trace2.gz | wc -l >>> 15340 >>> >>> this is 30GB worth of THPs (some of them might get released of course). >>> Also only 10% of requests ends up reclaiming. >>> >>> One additional interesting point >>> $ zgrep vmscan trace2.gz | grep nr_reclaimed | sed 's@.*nr_reclaimed=\([[0-9]*\)@\1@' | calc_min_max.awk >>> min: 1.00 max: 2792.00 avg: 965.99 std: 331.12 nr: 1596 >>> >>> Even though the std is high there are quite some outliers when a lot of >>> memory is reclaimed. >>> >>> Which kernel version is this. And again, what is the THP configuration. >> >> This is 4.19.66 regarding THP you mean this: > > Do you see the same behavior with 5.3? I rebootet with 5.3.0-rc8 - let's see what happens it might take some hours or even days. >> /sys/kernel/mm/transparent_hugepage/defrag:always defer [defer+madvise] >> madvise never >> >> /sys/kernel/mm/transparent_hugepage/enabled:[always] madvise never >> >> /sys/kernel/mm/transparent_hugepage/hpage_pmd_size:2097152 >> >> /sys/kernel/mm/transparent_hugepage/shmem_enabled:always within_size >> advise [never] deny force >> >> /sys/kernel/mm/transparent_hugepage/use_zero_page:1 >> >> /sys/kernel/mm/transparent_hugepage/enabled was madvise until yesterday >> where i tried to switch to defer+madvise - which didn't help. > > Many processes hitting the reclaim are php5 others I cannot say because > their cmd is not reflected in the trace. I suspect those are using > madvise. I haven't really seen kcompactd interfering much. That would > suggest using defer. You mean i should set transparent_hugepage to defer? Stefan