lkml.org 
[lkml]   [2018]   [Jun]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Subject[PATCH 00/19] Fixes for sched/numa_balancing
Date
This patchset based on v4.17-rc5, provides few simple cleanups and fixes in
the sched/numa_balancing code. Some of these fixes are specific to systems
having more than 2 nodes. Few patches add per-rq and per-node complexities
to solve what I feel are a fairness/correctness issues.


Here are the scripts used to benchmark this series
They are based on Andrea Arcangeli and Petr Holasek's
https://github.com/pholasek/autonuma-benchmark.git

# cat numa01.sh
#! /bin/bash
# numa01.sh corresponds to 2 perf bench processes each having ncpus/2 threads
# 50 loops of 3G process memory.

THREADS=${THREADS:-$(($(getconf _NPROCESSORS_ONLN)/2))}
perf bench numa mem --no-data_rand_walk $CPUS -p 2 -t $THREADS -G 0 -P 3072 -T 0 -l 50 -c -s 2000 $@


# cat numa02.sh
#! /bin/bash
# numa02.sh corresponds to 1 perf bench process having ncpus threads
# 800 loops of 32 MB thread specific memory.
THREADS=$(getconf _NPROCESSORS_ONLN)
perf bench numa mem --no-data_rand_walk -p 1 -t $THREADS -G 0 -P 0 -T 32 -l 800 -c -s 2000 $@



# cat numa03.sh
#! /bin/bash
# numa03.sh corresponds to 1 perf bench process having ncpus threads
# 50 loops of 3G process memory.

THREADS=$(getconf _NPROCESSORS_ONLN)
perf bench numa mem --no-data_rand_walk -p 1 -t $THREADS -G 0 -P 3072 -T 0 -l 50 -c -s 2000 $@


# cat numa04.sh
#! /bin/bash
# numa04.sh corresponds to nrnodes perf bench processes each having
# ncpus/nrnodes threads 50 loops of 3G process memory.

NODES=$(numactl -H |awk /available/'{print $2}')
INST=$NODES
THREADS=$(($(getconf _NPROCESSORS_ONLN)/$INST))
perf bench numa mem --no-data_rand_walk -p $INST -t $THREADS -G 0 -P 3072 -T 0 -l 50 -c -s 2000 $@


# cat numa05.sh
#! /bin/bash
# numa05.sh corresponds to nrnodes *2 perf bench processes each having
# ncpus/(nrnodes *2 ) threads 50 loops of 3G process memory.


NODES=$(numactl -H |awk /available/'{print $2}')
INST=$((2*NODES))
THREADS=$(($(getconf _NPROCESSORS_ONLN)/$INST))
perf bench numa mem --no-data_rand_walk -p $INST -t $THREADS -G 0 -P 3072 -T 0 -l 50 -c -s 2000 $@

Stats were collected on a 4 node/96 cpu machine.

# numactl -H
available: 4 nodes (0-3)
node 0 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
node 0 size: 32431 MB
node 0 free: 30759 MB
node 1 cpus: 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47
node 1 size: 31961 MB
node 1 free: 30502 MB
node 2 cpus: 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71
node 2 size: 30425 MB
node 2 free: 30189 MB
node 3 cpus: 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95
node 3 size: 32200 MB
node 3 free: 31346 MB
node distances:
node 0 1 2 3
0: 10 20 40 40
1: 20 10 40 40
2: 40 40 10 20
3: 40 40 20 10

Since we are looking for time as a metric, smaller numbers are better.

v4.17-rc5
Testcase Time: Min Max Avg StdDev
numa01.sh Real: 440.65 941.32 758.98 189.17
numa01.sh Sys: 183.48 320.07 258.42 50.09
numa01.sh User: 37384.65 71818.14 60302.51 13798.96
numa02.sh Real: 61.24 65.35 62.49 1.49
numa02.sh Sys: 16.83 24.18 21.40 2.60
numa02.sh User: 5219.59 5356.34 5264.03 49.07
numa03.sh Real: 822.04 912.40 873.55 37.35
numa03.sh Sys: 118.80 140.94 132.90 7.60
numa03.sh User: 62485.19 70025.01 67208.33 2967.10
numa04.sh Real: 690.66 872.12 778.49 65.44
numa04.sh Sys: 459.26 563.03 494.03 42.39
numa04.sh User: 51116.44 70527.20 58849.44 8461.28
numa05.sh Real: 418.37 562.28 525.77 54.27
numa05.sh Sys: 299.45 481.00 392.49 64.27
numa05.sh User: 34115.09 41324.02 39105.30 2627.68


v4.17-rc5+patches
Testcase Time: Min Max Avg StdDev %Change
numa01.sh Real: 424.63 566.18 498.12 59.26 34.36%
numa01.sh Sys: 160.19 256.53 208.98 37.02 19.13%
numa01.sh User: 37320.00 46225.58 42001.57 3482.45 30.34%
numa02.sh Real: 60.17 62.47 60.91 0.85 2.528%
numa02.sh Sys: 15.30 22.82 17.04 2.90 20.37%
numa02.sh User: 5202.13 5255.51 5219.08 20.14 0.853%
numa03.sh Real: 823.91 844.89 833.86 8.46 4.543%
numa03.sh Sys: 130.69 148.29 140.47 6.21 -5.69%
numa03.sh User: 62519.15 64262.20 63613.38 620.05 5.348%
numa04.sh Real: 515.30 603.74 548.56 30.93 29.53%
numa04.sh Sys: 459.73 525.48 489.18 21.63 0.981%
numa04.sh User: 40561.96 44919.18 42047.87 1526.85 28.55%
numa05.sh Real: 396.58 454.37 421.13 19.71 19.90%
numa05.sh Sys: 208.72 422.02 348.90 73.60 11.10%
numa05.sh User: 33124.08 36109.35 34846.47 1089.74 10.89%


Even the perf bench o/p (not included here because its pretty verbose)
points to a better consolidation. Attaching perf stat data instead.

Performance counter stats for 'system wide' (5 runs): numa01.sh
v4.17-rc5 v4.17-rc5+patches
cs 196,530 ( +- 13.22% ) 117,524 ( +- 7.46% )
migrations 16,077 ( +- 16.98% ) 6,602 ( +- 9.93% )
faults 1,698,631 ( +- 6.66% ) 1,292,159 ( +- 3.99% )
cache-misses 32,841,908,826 ( +- 5.33% ) 27,059,597,808 ( +- 2.17% )
sched:sched_move_numa 555 ( +- 25.92% ) 8 ( +- 38.45% )
sched:sched_stick_numa 16 ( +- 20.73% ) 1 ( +- 31.18% )
sched:sched_swap_numa 313 ( +- 23.21% ) 278 ( +- 5.31% )
migrate:mm_migrate_pages 138,981 ( +- 13.26% ) 121,639 ( +- 8.75% )
migrate:mm_numa_migrate_ratelimit 439 ( +-100.00% ) 138 ( +-100.00% )
seconds time elapsed 759.019898884 ( +- 12.46% ) 498.158680658 ( +- 5.95% )

Numa Hinting and other vmstat info (Sum of 5 runs)

numa01.sh v4.17-rc5 v4.17-rc5+patches
numa_foreign 0 0
numa_hint_faults 7283263 5389142
numa_hint_faults_local 3689375 2209029
numa_hit 1401549 1264559
numa_huge_pte_updates 0 0
numa_interleave 0 0
numa_local 1401487 1264534
numa_miss 0 0
numa_other 62 25
numa_pages_migrated 693724 608024
numa_pte_updates 7320797 5410463
pgfault 8514248 6474639
pgmajfault 351 203
pgmigrate_fail 1181 171
pgmigrate_success 693724 608024

Faults and page migrations have decreased and that correlates with perf stat
numbers. We are achieving faster and better consolidation with lesser task
migrations. Esp number of failed numa task migrations have decreased.



Performance counter stats for 'system wide' (5 runs): numa02.sh
v4.17-rc5 v4.17-rc5+patches
cs 33,541 ( +- 2.20% ) 33,472 ( +- 2.58% )
migrations 2,022 ( +- 2.36% ) 1,742 ( +- 4.36% )
faults 452,697 ( +- 6.29% ) 400,244 ( +- 3.14% )
cache-misses 4,559,889,977 ( +- 0.40% ) 4,510,581,926 ( +- 0.17% )
sched:sched_move_numa 27 ( +- 40.26% ) 2 ( +- 32.39% )
sched:sched_stick_numa 0 0
sched:sched_swap_numa 9 ( +- 41.81% ) 8 ( +- 23.39% )
migrate:mm_migrate_pages 23,428 ( +- 6.91% ) 19,418 ( +- 9.28% )
migrate:mm_numa_migrate_ratelimit 238 ( +- 61.52% ) 315 ( +- 66.65% )
seconds time elapsed 62.532524687 ( +- 1.20% ) 60.943143605 ( +- 0.70% )

Numa Hinting and other vmstat info (Sum of 5 runs)

numa02.sh v4.17-rc5 v4.17-rc5+patches
numa_foreign 0 0
numa_hint_faults 1797406 1541180
numa_hint_faults_local 1652638 1423986
numa_hit 447642 427011
numa_huge_pte_updates 0 0
numa_interleave 0 0
numa_local 447639 427011
numa_miss 0 0
numa_other 3 0
numa_pages_migrated 117142 97088
numa_pte_updates 1812907 1557075
pgfault 2273993 2011485
pgmajfault 112 119
pgmigrate_fail 0 0
pgmigrate_success 117142 97088

Again, lesser page faults, lesser page migrations but hitting page
migrations ratelimits more often.




Performance counter stats for 'system wide' (5 runs): numa03.sh
v4.17-rc5 v4.17-rc5+patches
cs 184,615 ( +- 2.83% ) 178,526 ( +- 2.66% )
migrations 14,010 ( +- 4.68% ) 9,511 ( +- 4.20% )
faults 766,543 ( +- 2.55% ) 835,876 ( +- 6.09% )
cache-misses 34,905,163,767 ( +- 0.75% ) 35,979,821,603 ( +- 0.30% )
sched:sched_move_numa 562 ( +- 6.38% ) 4 ( +- 22.64% )
sched:sched_stick_numa 16 ( +- 16.42% ) 1 ( +- 61.24% )
sched:sched_swap_numa 268 ( +- 4.88% ) 394 ( +- 6.05% )
migrate:mm_migrate_pages 53,999 ( +- 5.89% ) 51,520 ( +- 8.68% )
migrate:mm_numa_migrate_ratelimit 0 508 ( +- 76.69% )
seconds time elapsed 873.586758847 ( +- 2.14% ) 833.910858522 ( +- 0.51% )

Numa Hinting and other vmstat info (Sum of 5 runs)

numa03.sh v4.17-rc5 v4.17-rc5+patches
numa_foreign 0 0
numa_hint_faults 2962951 3275731
numa_hint_faults_local 1159054 1215206
numa_hit 702071 693754
numa_huge_pte_updates 0 0
numa_interleave 0 0
numa_local 702042 693722
numa_miss 0 0
numa_other 29 32
numa_pages_migrated 269918 256838
numa_pte_updates 2963554 3305006
pgfault 3853016 4193700
pgmajfault 202 281
pgmigrate_fail 77 764
pgmigrate_success 269918 256838

Seeing more faults and cache misses but lesser task migrations. Increase in
migration ratelimits is a worry.




Performance counter stats for 'system wide' (5 runs): numa04.sh

v4.17-rc5 v4.17-rc5+patches
cs 203,184 ( +- 6.67% ) 141,653 ( +- 3.26% )
migrations 17,852 ( +- 12.84% ) 6,837 ( +- 5.14% )
faults 3,650,884 ( +- 3.15% ) 2,910,839 ( +- 1.36% )
cache-misses 34,362,104,705 ( +- 2.26% ) 30,064,624,934 ( +- 1.18% )
sched:sched_move_numa 923 ( +- 21.36% ) 8 ( +- 30.22% )
sched:sched_stick_numa 10 ( +- 23.89% ) 1 ( +- 46.77% )
sched:sched_swap_numa 350 ( +- 21.32% ) 261 ( +- 7.80% )
migrate:mm_migrate_pages 288,410 ( +- 4.10% ) 296,726 ( +- 3.33% )
migrate:mm_numa_migrate_ratelimit 0 162 ( +-100.00% )
seconds time elapsed 778.519948731 ( +- 4.20% ) 548.606652462 ( +- 2.82% )

Numa Hinting and other vmstat info (Sum of 5 runs)

numa04.sh v4.17-rc5 v4.17-rc5+patches
numa_foreign 0 0
numa_hint_faults 16506833 12815656
numa_hint_faults_local 10237480 7526798
numa_hit 2617983 2647923
numa_huge_pte_updates 0 0
numa_interleave 0 0
numa_local 2617962 2647914
numa_miss 0 0
numa_other 21 9
numa_pages_migrated 1441453 1481743
numa_pte_updates 16519819 12844781
pgfault 18274350 14567947
pgmajfault 264 180
pgmigrate_fail 595 1889
pgmigrate_success 1441453 1481743

Lesser faults, page migrations and task migrations but increasing hitting
migrate ratelimits.




Performance counter stats for 'system wide' (5 runs): numa05.sh
v4.17-rc5 v4.17-rc5+patches
cs 149,941 ( +- 5.30% ) 119,881 ( +- 9.39% )
migrations 10,478 ( +- 13.01% ) 4,901 ( +- 6.53% )
faults 6,457,542 ( +- 3.07% ) 5,799,805 ( +- 1.62% )
cache-misses 31,146,034,587 ( +- 1.40% ) 29,894,482,788 ( +- 0.73% )
sched:sched_move_numa 667 ( +- 27.46% ) 6 ( +- 21.28% )
sched:sched_stick_numa 3 ( +- 27.28% ) 0
sched:sched_swap_numa 173 ( +- 20.79% ) 113 ( +- 17.60% )
migrate:mm_migrate_pages 419,446 ( +- 4.94% ) 325,522 ( +- 13.88% )
migrate:mm_numa_migrate_ratelimit 1,714 ( +- 66.17% ) 338 ( +- 45.02% )
seconds time elapsed 525.801216597 ( +- 5.16% ) 421.186302929 ( +- 2.34% )

Numa Hinting and other vmstat info (Sum of 5 runs)

numa05.sh v4.17-rc5 v4.17-rc5+patches
numa_foreign 0 0
numa_hint_faults 29575825 26294424
numa_hint_faults_local 21637356 21808958
numa_hit 4246286 3771867
numa_huge_pte_updates 0 0
numa_interleave 0 0
numa_local 4246270 3771854
numa_miss 0 0
numa_other 16 13
numa_pages_migrated 2096896 1625671
numa_pte_updates 29620565 26399455
pgfault 32309072 29013170
pgmajfault 285 255
pgmigrate_fail 334 1937
pgmigrate_success 2096896 1625671

Faults and page migrations have decreased . We are achieving faster and
better consolidation with lesser task migrations. Esp ratio of swap/move numa
task migrations has increased.

Srikar Dronamraju (19):
sched/numa: Remove redundant field.
sched/numa: Evaluate move once per node
sched/numa: Simplify load_too_imbalanced
sched/numa: Set preferred_node based on best_cpu
sched/numa: Use task faults only if numa_group is not yet setup
sched/debug: Reverse the order of printing faults
sched/numa: Skip nodes that are at hoplimit
sched/numa: Remove unused task_capacity from numa_stats
sched/numa: Modify migrate_swap to accept additional params
sched/numa: Stop multiple tasks from moving to the cpu at the same time
sched/numa: Restrict migrating in parallel to the same node.
sched:numa Remove numa_has_capacity
mm/migrate: Use xchg instead of spinlock
sched/numa: Updation of scan period need not be in lock
sched/numa: Use group_weights to identify if migration degrades locality
sched/numa: Detect if node actively handling migration
sched/numa: Pass destination cpu as a parameter to migrate_task_rq
sched/numa: Reset scan rate whenever task moves across nodes
sched/numa: Move task_placement closer to numa_migrate_preferred

include/linux/mmzone.h | 4 +-
include/linux/sched.h | 1 -
kernel/sched/core.c | 11 +-
kernel/sched/deadline.c | 2 +-
kernel/sched/debug.c | 4 +-
kernel/sched/fair.c | 328 +++++++++++++++++++++++-------------------------
kernel/sched/sched.h | 6 +-
mm/migrate.c | 8 +-
mm/page_alloc.c | 2 +-
9 files changed, 178 insertions(+), 188 deletions(-)

--
1.8.3.1

\
 
 \ /
  Last update: 2018-06-04 12:02    [W:0.556 / U:0.580 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site