sched/numa: Fix mm numa_scan_seq based unconditional scan
authorRaghavendra K T <raghavendra.kt@amd.com>
Fri, 20 Oct 2023 15:57:46 +0000 (21:27 +0530)
committerPeter Zijlstra <peterz@infradead.org>
Wed, 15 Nov 2023 08:57:46 +0000 (09:57 +0100)
commit84db47ca7146d7bd00eb5cf2b93989a971c84650
tree26b980bd690bb11d37e4379ebc070bcc20535697
parentd6111cf45c5787282b2e20d77bdb6b28881d516a
sched/numa: Fix mm numa_scan_seq based unconditional scan

Since commit fc137c0ddab2 ("sched/numa: enhance vma scanning logic")

NUMA Balancing allows updating PTEs to trap NUMA hinting faults if the
task had previously accessed VMA. However unconditional scan of VMAs are
allowed during initial phase of VMA creation until process's
mm numa_scan_seq reaches 2 even though current task had not accessed VMA.

Rationale:
 - Without initial scan subsequent PTE update may never happen.
 - Give fair opportunity to all the VMAs to be scanned and subsequently
understand the access pattern of all the VMAs.

But it has a corner case where, if a VMA is created after some time,
process's mm numa_scan_seq could be already greater than 2.

For e.g., values of mm numa_scan_seq when VMAs are created by running
mmtest autonuma benchmark briefly looks like:
start_seq=0 : 459
start_seq=2 : 138
start_seq=3 : 144
start_seq=4 : 8
start_seq=8 : 1
start_seq=9 : 1
This results in no unconditional PTE updates for those VMAs created after
some time.

Fix:
 - Note down the initial value of mm numa_scan_seq in per VMA start_seq.
 - Allow unconditional scan till start_seq + 2.

Result:
SUT: AMD EPYC Milan with 2 NUMA nodes 256 cpus.
base kernel: upstream 6.6-rc6 with Mels patches [1] applied.

kernbench
========== base                  patched %gain
Amean    elsp-128      165.09 ( 0.00%)      164.78 *   0.19%*

Duration User       41404.28    41375.08
Duration System      9862.22     9768.48
Duration Elapsed      519.87      518.72

Ops NUMA PTE updates           1041416.00      831536.00
Ops NUMA hint faults            263296.00      220966.00
Ops NUMA pages migrated         258021.00      212769.00
Ops AutoNUMA cost                 1328.67        1114.69

autonumabench

NUMA01_THREADLOCAL
==================
Amean  elsp-NUMA01_THREADLOCAL   81.79 (0.00%)  67.74 *  17.18%*

Duration User       54832.73    47379.67
Duration System        75.00      185.75
Duration Elapsed      576.72      476.09

Ops NUMA PTE updates                  394429.00    11121044.00
Ops NUMA hint faults                    1001.00     8906404.00
Ops NUMA pages migrated                  288.00     2998694.00
Ops AutoNUMA cost                          7.77       44666.84

Signed-off-by: Raghavendra K T <raghavendra.kt@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Mel Gorman <mgorman@suse.de>
Link: https://lore.kernel.org/r/2ea7cbce80ac7c62e90cbfb9653a7972f902439f.1697816692.git.raghavendra.kt@amd.com
include/linux/mm_types.h
kernel/sched/fair.c