[Problem Statement]
select_idle_cpu() might spend too much time searching for an idle CPU,
when the system is overloaded.
The following histogram is the time spent in select_idle_cpu(),
when running 224 instances of netperf on a system with 112 CPUs
per LLC domain:
@usecs:
[0] 533 | |
[1] 5495 | |
[2, 4) 12008 | |
[4, 8) 239252 | |
[8, 16) 4041924 |@@@@@@@@@@@@@@ |
[16, 32) 12357398 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[32, 64) 14820255 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[64, 128) 13047682 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[128, 256) 8235013 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[256, 512) 4507667 |@@@@@@@@@@@@@@@ |
[512, 1K) 2600472 |@@@@@@@@@ |
[1K, 2K) 927912 |@@@ |
[2K, 4K) 218720 | |
[4K, 8K) 98161 | |
[8K, 16K) 37722 | |
[16K, 32K) 6715 | |
[32K, 64K) 477 | |
[64K, 128K) 7 | |
netperf latency usecs:
=======
case load Lat_99th std%
TCP_RR thread-224 257.39 ( 0.21)
The time spent in select_idle_cpu() is visible to netperf and might have a negative
impact.
[Symptom analysis]
The patch [1] from Mel Gorman has been applied to track the efficiency
of select_idle_sibling. Copy the indicators here:
SIS Search Efficiency(se_eff%):
A ratio expressed as a percentage of runqueues scanned versus
idle CPUs found. A 100% efficiency indicates that the target,
prev or recent CPU of a task was idle at wakeup. The lower the
efficiency, the more runqueues were scanned before an idle CPU
was found.
SIS Domain Search Efficiency(dom_eff%):
Similar, except only for the slower SIS
patch.
SIS Fast Success Rate(fast_rate%):
Percentage of SIS that used target, prev or
recent CPUs.
SIS Success rate(success_rate%):
Percentage of scans that found an idle CPU.
The test is based on Aubrey's schedtests tool, including netperf, hackbench,
schbench and tbench.
Test on vanilla kernel:
schedstat_parse.py -f netperf_vanilla.log
case load se_eff% dom_eff% fast_rate% success_rate%
TCP_RR 28 threads 99.978 18.535 99.995 100.000
TCP_RR 56 threads 99.397 5.671 99.964 100.000
TCP_RR 84 threads 21.721 6.818 73.632 100.000
TCP_RR 112 threads 12.500 5.533 59.000 100.000
TCP_RR 140 threads 8.524 4.535 49.020 100.000
TCP_RR 168 threads 6.438 3.945 40.309 99.999
TCP_RR 196 threads 5.397 3.718 32.320 99.982
TCP_RR 224 threads 4.874 3.661 25.775 99.767
UDP_RR 28 threads 99.988 17.704 99.997 100.000
UDP_RR 56 threads 99.528 5.977 99.970 100.000
UDP_RR 84 threads 24.219 6.992 76.479 100.000
UDP_RR 112 threads 13.907 5.706 62.538 100.000
UDP_RR 140 threads 9.408 4.699 52.519 100.000
UDP_RR 168 threads 7.095 4.077 44.352 100.000
UDP_RR 196 threads 5.757 3.775 35.764 99.991
UDP_RR 224 threads 5.124 3.704 28.748 99.860
schedstat_parse.py -f schbench_vanilla.log
(each group has 28 tasks)
case load se_eff% dom_eff% fast_rate% success_rate%
normal 1 mthread 99.152 6.400 99.941 100.000
normal 2 mthreads 97.844 4.003 99.908 100.000
normal 3 mthreads 96.395 2.118 99.917 99.998
normal 4 mthreads 55.288 1.451 98.615 99.804
normal 5 mthreads 7.004 1.870 45.597 61.036
normal 6 mthreads 3.354 1.346 20.777 34.230
normal 7 mthreads 2.183 1.028 11.257 21.055
normal 8 mthreads 1.653 0.825 7.849 15.549
schedstat_parse.py -f hackbench_vanilla.log
(each group has 28 tasks)
case load se_eff% dom_eff% fast_rate% success_rate%
process-pipe 1 group 99.991 7.692 99.999 100.000
process-pipe 2 groups 99.934 4.615 99.997 100.000
process-pipe 3 groups 99.597 3.198 99.987 100.000
process-pipe 4 groups 98.378 2.464 99.958 100.000
process-pipe 5 groups 27.474 3.653 89.811 99.800
process-pipe 6 groups 20.201 4.098 82.763 99.570
process-pipe 7 groups 16.423 4.156 77.398 99.316
process-pipe 8 groups 13.165 3.920 72.232 98.828
process-sockets 1 group 99.977 5.882 99.999 100.000
process-sockets 2 groups 99.927 5.505 99.996 100.000
process-sockets 3 groups 99.397 3.250 99.980 100.000
process-sockets 4 groups 79.680 4.258 98.864 99.998
process-sockets 5 groups 7.673 2.503 63.659 92.115
process-sockets 6 groups 4.642 1.584 58.946 88.048
process-sockets 7 groups 3.493 1.379 49.816 81.164
process-sockets 8 groups 3.015 1.407 40.845 75.500
threads-pipe 1 group 99.997 0.000 100.000 100.000
threads-pipe 2 groups 99.894 2.932 99.997 100.000
threads-pipe 3 groups 99.611 4.117 99.983 100.000
threads-pipe 4 groups 97.703 2.624 99.937 100.000
threads-pipe 5 groups 22.919 3.623 87.150 99.764
threads-pipe 6 groups 18.016 4.038 80.491 99.557
threads-pipe 7 groups 14.663 3.991 75.239 99.247
threads-pipe 8 groups 12.242 3.808 70.651 98.644
threads-sockets 1 group 99.990 6.667 99.999 100.000
threads-sockets 2 groups 99.940 5.114 99.997 100.000
threads-sockets 3 groups 99.469 4.115 99.977 100.000
threads-sockets 4 groups 87.528 4.038 99.400 100.000
threads-sockets 5 groups 6.942 2.398 59.244 88.337
threads-sockets 6 groups 4.359 1.954 49.448 87.860
threads-sockets 7 groups 2.845 1.345 41.198 77.102
threads-sockets 8 groups 2.871 1.404 38.512 74.312
schedstat_parse.py -f tbench_vanilla.log
case load se_eff% dom_eff% fast_rate% success_rate%
loopback 28 threads 99.976 18.369 99.995 100.000
loopback 56 threads 99.222 7.799 99.934 100.000
loopback 84 threads 19.723 6.819 70.215 100.000
loopback 112 threads 11.283 5.371 55.371 99.999
loopback 140 threads 0.000 0.000 0.000 0.000
loopback 168 threads 0.000 0.000 0.000 0.000
loopback 196 threads 0.000 0.000 0.000 0.000
loopback 224 threads 0.000 0.000 0.000 0.000
According to the test above, if the system becomes busy, the
SIS Search Efficiency(se_eff%) drops significantly. Although some
benchmarks would finally find an idle CPU(success_rate% = 100%), it is
doubtful whether it is worth it to search the whole LLC domain.
[Proposal]
It would be ideal to have a crystal ball to answer this question:
How many CPUs must a wakeup path walk down, before it can find an idle
CPU? Many potential metrics could be used to predict the number.
One candidate is the sum of util_avg in this LLC domain. The benefit
of choosing util_avg is that it is a metric of accumulated historic
activity, which seems to be smoother than instantaneous metrics
(such as rq->nr_running). Besides, choosing the sum of util_avg
would help predict the load of the LLC domain more precisely, because
SIS_PROP uses one CPU's idle time to estimate the total LLC domain idle
time.
In summary, the lower the util_avg is, the more select_idle_cpu()
should scan for idle CPU, and vice versa. When the sum of util_avg
in this LLC domain hits 85% or above, the scan stops. The reason to
choose 85% as the threshold is that this is the imbalance_pct(117)
when a LLC sched group is overloaded.
Introduce the quadratic function:
y = SCHED_CAPACITY_SCALE - p * x^2
and y'= y / SCHED_CAPACITY_SCALE
x is the ratio of sum_util compared to the CPU capacity:
x = sum_util / (llc_weight * SCHED_CAPACITY_SCALE)
y' is the ratio of CPUs to be scanned in the LLC domain,
and the number of CPUs to scan is calculated by:
nr_scan = llc_weight * y'
Choosing quadratic function is because:
[1] Compared to the linear function, it scans more aggressively when the
sum_util is low.
[2] Compared to the exponential function, it is easier to calculate.
[3] It seems that there is no accurate mapping between the sum of util_avg
and the number of CPUs to be scanned. Use heuristic scan for now.
For a platform with 112 CPUs per LLC, the number of CPUs to scan is:
sum_util% 0 5 15 25 35 45 55 65 75 85 86 ...
scan_nr 112 111 108 102 93 81 65 47 25 1 0 ...
For a platform with 16 CPUs per LLC, the number of CPUs to scan is:
sum_util% 0 5 15 25 35 45 55 65 75 85 86 ...
scan_nr 16 15 15 14 13 11 9 6 3 0 0 ...
Furthermore, to minimize the overhead of calculating the metrics in
select_idle_cpu(), borrow the statistics from periodic load balance.
As mentioned by Abel, on a platform with 112 CPUs per LLC, the
sum_util calculated by periodic load balance after 112 ms would
decay to about 0.5 * 0.5 * 0.5 * 0.7 = 8.75%, thus bringing a delay
in reflecting the latest utilization. But it is a trade-off.
Checking the util_avg in newidle load balance would be more frequent,
but it brings overhead - multiple CPUs write/read the per-LLC shared
variable and introduces cache contention. Tim also mentioned that,
it is allowed to be non-optimal in terms of scheduling for the
short-term variations, but if there is a long-term trend in the load
behavior, the scheduler can adjust for that.
When SIS_UTIL is enabled, the select_idle_cpu() uses the nr_scan
calculated by SIS_UTIL instead of the one from SIS_PROP. As Peter and
Mel suggested, SIS_UTIL should be enabled by default.
This patch is based on the util_avg, which is very sensitive to the
CPU frequency invariance. There is an issue that, when the max frequency
has been clamp, the util_avg would decay insanely fast when
the CPU is idle. Commit addca28512 ("cpufreq: intel_pstate: Handle no_turbo
in frequency invariance") could be used to mitigate this symptom, by adjusting
the arch_max_freq_ratio when turbo is disabled. But this issue is still
not thoroughly fixed, because the current code is unaware of the user-specified
max CPU frequency.
[Test result]
netperf and tbench were launched with 25% 50% 75% 100% 125% 150%
175% 200% of CPU number respectively. Hackbench and schbench were launched
by 1, 2 ,4, 8 groups. Each test lasts for 100 seconds and repeats 3 times.
The following is the benchmark result comparison between
baseline:vanilla v5.19-rc1 and compare:patched kernel. Positive compare%
indicates better performance.
Each netperf test is a:
netperf -4 -H 127.0.1 -t TCP/UDP_RR -c -C -l 100
netperf.throughput
=======
case load baseline(std%) compare%( std%)
TCP_RR 28 threads 1.00 ( 0.34) -0.16 ( 0.40)
TCP_RR 56 threads 1.00 ( 0.19) -0.02 ( 0.20)
TCP_RR 84 threads 1.00 ( 0.39) -0.47 ( 0.40)
TCP_RR 112 threads 1.00 ( 0.21) -0.66 ( 0.22)
TCP_RR 140 threads 1.00 ( 0.19) -0.69 ( 0.19)
TCP_RR 168 threads 1.00 ( 0.18) -0.48 ( 0.18)
TCP_RR 196 threads 1.00 ( 0.16) +194.70 ( 16.43)
TCP_RR 224 threads 1.00 ( 0.16) +197.30 ( 7.85)
UDP_RR 28 threads 1.00 ( 0.37) +0.35 ( 0.33)
UDP_RR 56 threads 1.00 ( 11.18) -0.32 ( 0.21)
UDP_RR 84 threads 1.00 ( 1.46) -0.98 ( 0.32)
UDP_RR 112 threads 1.00 ( 28.85) -2.48 ( 19.61)
UDP_RR 140 threads 1.00 ( 0.70) -0.71 ( 14.04)
UDP_RR 168 threads 1.00 ( 14.33) -0.26 ( 11.16)
UDP_RR 196 threads 1.00 ( 12.92) +186.92 ( 20.93)
UDP_RR 224 threads 1.00 ( 11.74) +196.79 ( 18.62)
Take the 224 threads as an example, the SIS search metrics changes are
illustrated below:
vanilla patched
4544492 +237.5% 15338634 sched_debug.cpu.sis_domain_search.avg
38539 +39686.8% 15333634 sched_debug.cpu.sis_failed.avg
128300000 -87.9% 15551326 sched_debug.cpu.sis_scanned.avg
5842896 +162.7% 15347978 sched_debug.cpu.sis_search.avg
There is -87.9% less CPU scans after patched, which indicates lower overhead.
Besides, with this patch applied, there is -13% less rq lock contention
in perf-profile.calltrace.cycles-pp._raw_spin_lock.raw_spin_rq_lock_nested
.try_to_wake_up.default_wake_function.woken_wake_function.
This might help explain the performance improvement - Because this patch allows
the waking task to remain on the previous CPU, rather than grabbing other CPUs'
lock.
Each hackbench test is a:
hackbench -g $job --process/threads --pipe/sockets -l 1000000 -s 100
hackbench.throughput
=========
case load baseline(std%) compare%( std%)
process-pipe 1 group 1.00 ( 1.29) +0.57 ( 0.47)
process-pipe 2 groups 1.00 ( 0.27) +0.77 ( 0.81)
process-pipe 4 groups 1.00 ( 0.26) +1.17 ( 0.02)
process-pipe 8 groups 1.00 ( 0.15) -4.79 ( 0.02)
process-sockets 1 group 1.00 ( 0.63) -0.92 ( 0.13)
process-sockets 2 groups 1.00 ( 0.03) -0.83 ( 0.14)
process-sockets 4 groups 1.00 ( 0.40) +5.20 ( 0.26)
process-sockets 8 groups 1.00 ( 0.04) +3.52 ( 0.03)
threads-pipe 1 group 1.00 ( 1.28) +0.07 ( 0.14)
threads-pipe 2 groups 1.00 ( 0.22) -0.49 ( 0.74)
threads-pipe 4 groups 1.00 ( 0.05) +1.88 ( 0.13)
threads-pipe 8 groups 1.00 ( 0.09) -4.90 ( 0.06)
threads-sockets 1 group 1.00 ( 0.25) -0.70 ( 0.53)
threads-sockets 2 groups 1.00 ( 0.10) -0.63 ( 0.26)
threads-sockets 4 groups 1.00 ( 0.19) +11.92 ( 0.24)
threads-sockets 8 groups 1.00 ( 0.08) +4.31 ( 0.11)
Each tbench test is a:
tbench -t 100 $job 127.0.0.1
tbench.throughput
======
case load baseline(std%) compare%( std%)
loopback 28 threads 1.00 ( 0.06) -0.14 ( 0.09)
loopback 56 threads 1.00 ( 0.03) -0.04 ( 0.17)
loopback 84 threads 1.00 ( 0.05) +0.36 ( 0.13)
loopback 112 threads 1.00 ( 0.03) +0.51 ( 0.03)
loopback 140 threads 1.00 ( 0.02) -1.67 ( 0.19)
loopback 168 threads 1.00 ( 0.38) +1.27 ( 0.27)
loopback 196 threads 1.00 ( 0.11) +1.34 ( 0.17)
loopback 224 threads 1.00 ( 0.11) +1.67 ( 0.22)
Each schbench test is a:
schbench -m $job -t 28 -r 100 -s 30000 -c 30000
schbench.latency_90%_us
========
case load baseline(std%) compare%( std%)
normal 1 mthread 1.00 ( 31.22) -7.36 ( 20.25)*
normal 2 mthreads 1.00 ( 2.45) -0.48 ( 1.79)
normal 4 mthreads 1.00 ( 1.69) +0.45 ( 0.64)
normal 8 mthreads 1.00 ( 5.47) +9.81 ( 14.28)
*Consider the Standard Deviation, this -7.36% regression might not be valid.
Also, a OLTP workload with a commercial RDBMS has been tested, and there
is no significant change.
There were concerns that unbalanced tasks among CPUs would cause problems.
For example, suppose the LLC domain is composed of 8 CPUs, and 7 tasks are
bound to CPU0~CPU6, while CPU7 is idle:
CPU0 CPU1 CPU2 CPU3 CPU4 CPU5 CPU6 CPU7
util_avg 1024 1024 1024 1024 1024 1024 1024 0
Since the util_avg ratio is 87.5%( = 7/8 ), which is higher than 85%,
select_idle_cpu() will not scan, thus CPU7 is undetected during scan.
But according to Mel, it is unlikely the CPU7 will be idle all the time
because CPU7 could pull some tasks via CPU_NEWLY_IDLE.
lkp(kernel test robot) has reported a regression on stress-ng.sock on a
very busy system. According to the sched_debug statistics, it might be caused
by SIS_UTIL terminates the scan and chooses a previous CPU earlier, and this
might introduce more context switch, especially involuntary preemption, which
impacts a busy stress-ng. This regression has shown that, not all benchmarks
in every scenario benefit from idle CPU scan limit, and it needs further
investigation.
Besides, there is slight regression in hackbench's 16 groups case when the
LLC domain has 16 CPUs. Prateek mentioned that we should scan aggressively
in an LLC domain with 16 CPUs. Because the cost to search for an idle one
among 16 CPUs is negligible. The current patch aims to propose a generic
solution and only considers the util_avg. Something like the below could
be applied on top of the current patch to fulfill the requirement:
if (llc_weight <= 16)
nr_scan = nr_scan * 32 / llc_weight;
For LLC domain with 16 CPUs, the nr_scan will be expanded to 2 times large.
The smaller the CPU number this LLC domain has, the larger nr_scan will be
expanded. This needs further investigation.
There is also ongoing work[2] from Abel to filter out the busy CPUs during
wakeup, to further speed up the idle CPU scan. And it could be a following-up
optimization on top of this change.
Suggested-by: Tim Chen <tim.c.chen@intel.com>
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Chen Yu <yu.c.chen@intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Yicong Yang <yangyicong@hisilicon.com>
Tested-by: Mohini Narkhede <mohini.narkhede@intel.com>
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Link: https://lore.kernel.org/r/20220612163428.849378-1-yu.c.chen@intel.com
70fb5ccf2e