| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-Id: <20210513074027.543926-1-srikar@linux.vnet.ibm.com>
Date: Thu, 13 May 2021 13:10:19 +0530
From: Srikar Dronamraju <srikar@...ux.vnet.ibm.com>
To: Ingo Molnar <mingo@...nel.org>,
Peter Zijlstra <peterz@...radead.org>
Cc: LKML <linux-kernel@...r.kernel.org>,
Mel Gorman <mgorman@...hsingularity.net>,
Rik van Riel <riel@...riel.com>,
Srikar Dronamraju <srikar@...ux.vnet.ibm.com>,
Thomas Gleixner <tglx@...utronix.de>,
Valentin Schneider <valentin.schneider@....com>,
Vincent Guittot <vincent.guittot@...aro.org>,
Dietmar Eggemann <dietmar.eggemann@....com>,
Michael Ellerman <mpe@...erman.id.au>,
Gautham R Shenoy <ego@...ux.vnet.ibm.com>,
Parth Shah <parth@...ux.ibm.com>,
Aubrey Li <aubrey.li@...ux.intel.com>
Subject: [PATCH v3 0/8] sched/fair: wake_affine improvements
Changelog v2->v3:
v2: http://lore.kernel.org/lkml/20210506164543.90688-1-srikar@linux.vnet.ibm.com/t/#u
- Rebased to tip/sched/core
(Valentin)
- Update schedstat if target is current CPU
(Valentin)
- Search for idle-cores in LLC only if idle-core is -1
Changelog v1->v2:
v1 Link: http://lore.kernel.org/lkml/20210422102326.35889-1-srikar@linux.vnet.ibm.com/t/#u
- Fallback LLC domain has been split out as a subsequent patchset.
(suggested by Mel)
- Fix a panic due to two wakeups racing for the same idle-core
(Reported by Mel)
- Differentiate if a LLC surely has no idle-cores(-2) vs a LLC may or
may not have idle-cores(-1).
- Rebased to v5.12
Recently we found that some of the benchmark numbers on Power10 were lesser
than expected. Some analysis showed that the problem lies in the fact that
L2-Cache on Power10 is at core level i.e only 4 threads share the L2-cache.
One probable solution to the problem was worked by Gautham where he posted
http://lore.kernel.org/lkml/1617341874-1205-1-git-send-email-ego@linux.vnet.ibm.com/t/#u
a patch that marks MC domain as LLC.
Here the focus is on improving the current core scheduler's wakeup
mechanism by looking at idle-cores and nr_busy_cpus that is already
maintained per Last level cache(aka LLC)
Hence this approach can work well with the mc-llc too. It can help
other architectures too.
Request you to please review and provide your feedback.
Benchmarking numbers are from Power 10 but I have verified that we don't
regress on Power 9 setup.
# lscpu
Architecture: ppc64le
Byte Order: Little Endian
CPU(s): 80
On-line CPU(s) list: 0-79
Thread(s) per core: 8
Core(s) per socket: 10
Socket(s): 1
NUMA node(s): 1
Model: 1.0 (pvr 0080 0100)
Model name: POWER10 (architected), altivec supported
Hypervisor vendor: pHyp
Virtualization type: para
L1d cache: 64K
L1i cache: 32K
L2 cache: 256K
L3 cache: 8K
NUMA node2 CPU(s): 0-79
Hackbench: (latency, lower is better)
v5.12-rc5
instances = 1, min = 24.102529 usecs/op, median = usecs/op, max = 24.102529 usecs/op
instances = 2, min = 24.096112 usecs/op, median = 24.096112 usecs/op, max = 24.178903 usecs/op
instances = 4, min = 24.080541 usecs/op, median = 24.082990 usecs/op, max = 24.166873 usecs/op
instances = 8, min = 24.088969 usecs/op, median = 24.116081 usecs/op, max = 24.199853 usecs/op
instances = 16, min = 24.267228 usecs/op, median = 26.204510 usecs/op, max = 29.218360 usecs/op
instances = 32, min = 30.680071 usecs/op, median = 32.911664 usecs/op, max = 37.380470 usecs/op
instances = 64, min = 43.908331 usecs/op, median = 44.454343 usecs/op, max = 46.210298 usecs/op
instances = 80, min = 44.585754 usecs/op, median = 56.738546 usecs/op, max = 60.625826 usecs/op
v5.12-rc5 + mc-llc+
instances = 1, min = 18.676505 usecs/op, median = usecs/op, max = 18.676505 usecs/op
instances = 2, min = 18.488627 usecs/op, median = 18.488627 usecs/op, max = 18.574946 usecs/op
instances = 4, min = 18.428399 usecs/op, median = 18.589051 usecs/op, max = 18.872548 usecs/op
instances = 8, min = 18.597389 usecs/op, median = 18.783815 usecs/op, max = 19.265532 usecs/op
instances = 16, min = 21.922350 usecs/op, median = 22.737792 usecs/op, max = 24.832429 usecs/op
instances = 32, min = 29.770446 usecs/op, median = 31.996687 usecs/op, max = 34.053042 usecs/op
instances = 64, min = 53.067842 usecs/op, median = 53.295139 usecs/op, max = 53.473059 usecs/op
instances = 80, min = 44.423288 usecs/op, median = 44.713767 usecs/op, max = 45.159761 usecs/op
v5.12-rc5 + this patchset
instances = 1, min = 19.240824 usecs/op, median = usecs/op, max = 19.240824 usecs/op
instances = 2, min = 19.143470 usecs/op, median = 19.143470 usecs/op, max = 19.249875 usecs/op
instances = 4, min = 19.399812 usecs/op, median = 19.487433 usecs/op, max = 19.501298 usecs/op
instances = 8, min = 19.024297 usecs/op, median = 19.908682 usecs/op, max = 20.741605 usecs/op
instances = 16, min = 22.209444 usecs/op, median = 23.971275 usecs/op, max = 25.145198 usecs/op
instances = 32, min = 31.220392 usecs/op, median = 32.689189 usecs/op, max = 34.081588 usecs/op
instances = 64, min = 39.012110 usecs/op, median = 44.062042 usecs/op, max = 45.370525 usecs/op
instances = 80, min = 43.884358 usecs/op, median = 44.326417 usecs/op, max = 48.031303 usecs/op
Summary:
mc-llc and this patchset seem to be performing much better than vanilla v5.12-rc5
DayTrader (throughput, higher is better)
v5.12-rc5 v5.12-rc5 v5.12-rc5
+ mc-llc + patchset
64CPUs/1JVM/ 60Users 6373.7 7520.5 7375.6
64CPUs/1JVM/ 80Users 6742.1 7940.9 7832.9
64CPUs/1JVM/100Users 6482.2 7730.3 7538.4
64CPUs/2JVM/ 60Users 6335 8081.6 8000.2
64CPUs/2JVM/ 80Users 6360.8 8259.6 8315.4
64CPUs/2JVM/100Users 6215.6 8046.5 8049.4
64CPUs/4JVM/ 60Users 5385.4 7685.3 8013.5
64CPUs/4JVM/ 80Users 5380.8 7753.3 7868
64CPUs/4JVM/100Users 5275.2 7549.2 7620
Summary: Across all profiles, this patchset or mc-llc out perform
vanilla v5.12-rc5
Not: Only 64 cores were online during this test.
schbench (latency: lesser is better)
======== Running schbench -m 3 -r 30 =================
Latency percentiles (usec) runtime 10 (s) (2545 total samples)
v5.12-rc5 | v5.12-rc5 + mc-llc | v5.12-rc5 + patchset
50.0th: 56 (1301 samples) | 50.0th: 49 (1309 samples) | 50.0th: 53 (1285 samples)
75.0th: 76 (623 samples) | 75.0th: 66 (628 samples) | 75.0th: 72 (635 samples)
90.0th: 93 (371 samples) | 90.0th: 78 (371 samples) | 90.0th: 88 (388 samples)
95.0th: 107 (123 samples) | 95.0th: 87 (117 samples) | 95.0th: 94 (118 samples)
*99.0th: 12560 (102 samples) *99.0th: 100 (97 samples) | *99.0th: 108 (108 samples)
99.5th: 15312 (14 samples) | 99.5th: 104 (12 samples) | 99.5th: 108 (0 samples)
99.9th: 19936 (9 samples) | 99.9th: 106 (8 samples) | 99.9th: 110 (8 samples)
min=13, max=20684 | min=15, max=113 | min=15, max=1433
Latency percentiles (usec) runtime 20 (s) (7649 total samples)
50.0th: 51 (3884 samples) | 50.0th: 50 (3935 samples) | 50.0th: 51 (3843 samples)
75.0th: 69 (1859 samples) | 75.0th: 66 (1817 samples) | 75.0th: 69 (1962 samples)
90.0th: 87 (1173 samples) | 90.0th: 80 (1204 samples) | 90.0th: 84 (1103 samples)
95.0th: 97 (368 samples) | 95.0th: 87 (342 samples) | 95.0th: 93 (386 samples)
*99.0th: 8624 (290 samples)| *99.0th: 98 (294 samples) | *99.0th: 107 (297 samples)
99.5th: 11344 (37 samples) | 99.5th: 102 (37 samples) | 99.5th: 110 (39 samples)
99.9th: 18592 (31 samples) | 99.9th: 106 (30 samples) | 99.9th: 1714 (27 samples)
min=13, max=20684 | min=12, max=113 | min=15, max=4456
Latency percentiles (usec) runtime 30 (s) (12785 total samples)
50.0th: 50 (6614 samples) | 50.0th: 49 (6544 samples) | 50.0th: 50 (6443 samples)
75.0th: 67 (3059 samples) | 75.0th: 65 (3100 samples) | 75.0th: 67 (3263 samples)
90.0th: 84 (1894 samples) | 90.0th: 79 (1912 samples) | 90.0th: 82 (1890 samples)
95.0th: 94 (586 samples) | 95.0th: 87 (646 samples) | 95.0th: 92 (652 samples)
*99.0th: 8304 (507 samples)| *99.0th: 101 (496 samples) | *99.0th: 107 (464 samples)
99.5th: 11696 (62 samples) | 99.5th: 104 (45 samples) | 99.5th: 110 (61 samples)
99.9th: 18592 (51 samples) | 99.9th: 110 (51 samples) | 99.9th: 1434 (47 samples)
min=12, max=21421 | min=1, max=126 | min=15, max=4456
Summary:
mc-llc is the best option, but this patchset also helps compared to vanilla v5.12-rc5
mongodb (threads=6) (throughput, higher is better)
Throughput read clean update
latency latency latency
v5.12-rc5 JVM=YCSB_CLIENTS=14 68116.05 ops/sec 1109.82 us 944.19 us 1342.29 us
v5.12-rc5 JVM=YCSB_CLIENTS=21 64802.69 ops/sec 1772.64 us 944.69 us 2099.57 us
v5.12-rc5 JVM=YCSB_CLIENTS=28 61792.78 ops/sec 2490.48 us 930.09 us 2928.03 us
v5.12-rc5 JVM=YCSB_CLIENTS=35 59604.44 ops/sec 3236.86 us 870.28 us 3787.48 us
v5.12-rc5 + mc-llc JVM=YCSB_CLIENTS=14 70948.51 ops/sec 1060.21 us 842.02 us 1289.44 us
v5.12-rc5 + mc-llc JVM=YCSB_CLIENTS=21 68732.48 ops/sec 1669.91 us 871.57 us 1975.19 us
v5.12-rc5 + mc-llc JVM=YCSB_CLIENTS=28 66674.81 ops/sec 2313.79 us 889.59 us 2702.36 us
v5.12-rc5 + mc-llc JVM=YCSB_CLIENTS=35 64397.51 ops/sec 3010.66 us 966.28 us 3484.19 us
v5.12-rc5 + patchset JVM=YCSB_CLIENTS=14 67604.51 ops/sec 1117.91 us 947.07 us 1353.41 us
v5.12-rc5 + patchset JVM=YCSB_CLIENTS=21 63979.39 ops/sec 1793.63 us 869.72 us 2130.22 us
v5.12-rc5 + patchset JVM=YCSB_CLIENTS=28 62032.34 ops/sec 2475.89 us 869.06 us 2922.01 us
v5.12-rc5 + patchset JVM=YCSB_CLIENTS=35 60152.96 ops/sec 3203.84 us 972.00 us 3756.52 us
Summary:
mc-llc outperforms, this patchset and upstream almost give similar performance.
Cc: LKML <linux-kernel@...r.kernel.org>
Cc: Michael Ellerman <mpe@...erman.id.au>
Cc: Gautham R Shenoy <ego@...ux.vnet.ibm.com>
Cc: Parth Shah <parth@...ux.ibm.com>
Cc: Ingo Molnar <mingo@...nel.org>
Cc: Peter Zijlstra <peterz@...radead.org>
Cc: Valentin Schneider <valentin.schneider@....com>
Cc: Dietmar Eggemann <dietmar.eggemann@....com>
Cc: Mel Gorman <mgorman@...hsingularity.net>
Cc: Vincent Guittot <vincent.guittot@...aro.org>
Cc: Rik van Riel <riel@...riel.com>
Cc: Aubrey Li <aubrey.li@...ux.intel.com>
Srikar Dronamraju (8):
sched/fair: Update affine statistics when needed
sched/fair: Maintain the identity of idle-core
sched/fair: Update idle-core more often
sched/fair: Prefer idle CPU to cache affinity
sched/fair: Use affine_idler_llc for wakeups across LLC
sched/idle: Move busy_cpu accounting to idle callback
sched/fair: Remove ifdefs in waker_affine_idler_llc
sched/fair: Dont iterate if no idle CPUs
include/linux/sched/topology.h | 2 +-
kernel/sched/fair.c | 220 ++++++++++++++++++++++++++-------
kernel/sched/features.h | 1 +
kernel/sched/idle.c | 33 ++++-
kernel/sched/sched.h | 6 +
kernel/sched/topology.c | 9 ++
6 files changed, 222 insertions(+), 49 deletions(-)
base-commit: 2ea46c6fc9452ac100ad907b051d797225847e33
--
2.18.2
Powered by blists - more mailing lists