With the numa scan enhancements [1], only the threads which had previously accessed vma are allowed to scan. While this has improved significant system time overhead, there are corner cases, which genuinely needs some relaxation for e.g., concern raised by PeterZ where unfairness amongst the threadbelonging to disjoint set of VMSs can potentially amplify the side effects of vma regions belonging to some of the tasks being left unscanned. Currently that is handled by allowing first two scans at mm level (mm->numa_scan_seq) unconditionally. One of the test that exercise similar side effect is numa01_THREAD_ALLOC where allocation happen by main thread and it is divided into memory chunks of 24MB to be continuously bzeroed. (this is run default by LKP tests while numa01 is run default in mmtests which operate on full 3GB region by each thread) So to address this issue, proposal here is: 1) Have per vma scan counter that gets incremented for every successful scan (which potentially scans 256MB or sysctl_scan_size) 2) Do unconditional scan for first few times (To be precise, half of the window calculated for scanning normally otherwise) 3) Do reset of the counter when whole mm is scanned (this needs remembering mm->numa_scan_sequece) at vma level With this patch I am seeing good improvement in numa01_THREAD_ALLOC case, but please note that with [1] there was a drastic decrease in system time when benchmarks run, this patch adds back some of the system time. Your comments/Ideas are welcome. Result: SUT: Milan w/ 2 numa nodes 256 cpus Manaul run of numa01_THREAD__ALLOC Base 11-apr-next w/numascan w/o numascan numascan+patch real 1m33.579s 1m2.042s 1m11.738s user 280m46.032s 213m38.647s 231m40.226s sys 0m18.061s 6m54.963s 4m43.174s numa_hit 5813057 6166060 6146064 numa_local 5812546 6165471 6145573 numa_other 511 589 491 numa_pte_updates 0 2098276 1248398 numa_hint_faults 10 1768382 982034 numa_hint_faults_local 10 981824 625424 numa_pages_migrated 0 786558 356604 Below is the mmtest kernbench and autonuma performance kernbench =========== Base 11-apr-next w/numascan w/o numascan numascan+patch Amean user-256 23873.01 ( 0.00%) 23688.21 * 0.77%* 23948.47 * -0.32%* Amean syst-256 4990.73 ( 0.00%) 5113.32 * -2.46%* 4800.86 * 3.80%* Amean elsp-256 150.67 ( 0.00%) 150.52 * 0.10%* 150.63 * 0.03%* Duration User 71628.53 71074.04 71855.31 Duration System 14985.61 15354.33 14416.72 Duration Elapsed 472.69 473.24 473.72 Ops NUMA alloc hit 1739476674.00 1739443601.00 1739591558.00 Ops NUMA alloc local 1739534231.00 1739519795.00 1739647666.00 Ops NUMA base-page range updates 485073.00 673766.00 733129.00 Ops NUMA PTE updates 485073.00 673766.00 733129.00 Ops NUMA hint faults 107776.00 181920.00 186250.00 Ops NUMA hint local faults % 1789.00 6165.00 10889.00 Ops NUMA hint local percent 1.66 3.39 5.85 Ops NUMA pages migrated 105987.00 175755.00 175356.00 Ops AutoNUMA cost 544.29 917.66 939.71 autonumabench =============== w/numascan w/o numascan numascan+patch Amean syst-NUMA01 33.10 ( 0.00%) 571.68 *-1627.21%* 219.51 *-563.21%* Amean syst-NUMA01_THREADLOCAL 0.23 ( 0.00%) 0.22 * 4.38%* 0.22 * 5.00%* Amean syst-NUMA02 0.81 ( 0.00%) 0.75 * 7.76%* 0.76 * 6.00%* Amean syst-NUMA02_SMT 0.68 ( 0.00%) 0.73 * -7.79%* 0.65 * 3.58%* Amean elsp-NUMA01 299.71 ( 0.00%) 333.24 * -11.19%* 329.60 * -9.97%* Amean elsp-NUMA01_THREADLOCAL 1.06 ( 0.00%) 1.06 * 0.00%* 1.06 * -0.68%* Amean elsp-NUMA02 3.29 ( 0.00%) 3.23 * 1.95%* 3.18 * 3.51%* Amean elsp-NUMA02_SMT 3.75 ( 0.00%) 3.38 * 9.86%* 3.79 * -0.95%* Duration User 321693.29 437210.09 376657.80 Duration System 244.25 4014.23 1548.57 Duration Elapsed 2165.83 2395.53 2373.46 Ops NUMA alloc hit 49608099.00 62272320.00 55815229.00 Ops NUMA alloc local 49585747.00 62236996.00 55812601.00 Ops NUMA base-page range updates 1571.00 202868357.00 96006221.00 Ops NUMA PTE updates 1571.00 202868357.00 96006221.00 Ops NUMA hint faults 1203.00 204902318.00 97246909.00 Ops NUMA hint local faults % 981.00 187233695.00 81136933.00 Ops NUMA hint local percent 81.55 91.38 83.43 Ops NUMA pages migrated 222.00 10011134.00 6060787.00 Ops AutoNUMA cost 6.03 1026121.88 487021.74 Notes: Implementation considered/tried 1) Limit the disjoint set vma scan to 4 (hardcoded) = 1GB per whole mm scan 2) Current PID reset window = 4 * sysctl_scan_delay is changed to 8 * sysctl_scan_delay (to ensure some random overlapping overtime in scanning) links: [1] https://lore.kernel.org/lkml/cover.1677672277.git.raghavendra.kt@xxxxxxx/T/#t Raghavendra K T (2): sched/numa: Introduce per vma scan counter sched/numa: Introduce per vma numa_scan_seq include/linux/mm_types.h | 2 ++ kernel/sched/fair.c | 44 +++++++++++++++++++++++++++++++++++++--- 2 files changed, 43 insertions(+), 3 deletions(-) -- 2.34.1
