Re: [PATCH] sched/fair: no sync wakeup from interrupt context

From: K Prateek Nayak
Date: Fri Jul 29 2022 - 00:47:48 EST


Hello Libo and Peter,

tl;dr

- We observed a major regression with tbench when testing the latest tip
sched/core at:
commit 14b3f2d9ee8d "sched/fair: Disallow sync wakeup from interrupt context"
Reason for the regression are the fewer affine wakeups that leaves the
client farther away from the data it needs to consume next primed in the
waker's LLC.
Such regressions can be expected from tasks that use sockets to communicate
significant amount of data especially on system with multiple LLCs.

- Other benchmarks have a comparable behavior to the tip at previous commit
commit : 91caa5ae2424 "sched/core: Fix the bug that task won't enqueue
into core tree when update cookie"

I'll leave more details below.

On 7/12/2022 4:17 AM, Libo Chen wrote:
> Barry Song first pointed out that replacing sync wakeup with regular wakeup
> seems to reduce overeager wakeup pulling and shows noticeable performance
> improvement.[1]
>
> This patch argues that allowing sync for wakeups from interrupt context
> is a bug and fixing it can improve performance even when irq/softirq is
> evenly spread out.
>
> For wakeups from ISR, the waking CPU is just the CPU of ISR and the so-called
> waker can be any random task that happens to be running on that CPU when the
> interrupt comes in. This is completely different from other wakups where the
> task running on the waking CPU is the actual waker. For example, two tasks
> communicate through a pipe or mutiple tasks access the same critical section,
> etc. This difference is important because with sync we assume the waker will
> get off the runqueue and go to sleep immedately after the wakeup. The
> assumption is built into wake_affine() where it discounts the waker's presence
> from the runqueue when sync is true. The random waker from interrupts bears no
> relation to the wakee and don't usually go to sleep immediately afterwards
> unless wakeup granularity is reached. Plus the scheduler no longer enforces the
> preepmtion of waker for sync wakeup as it used to before
> patch f2e74eeac03ffb7 ("sched: Remove WAKEUP_SYNC feature"). Enforcing sync
> wakeup preemption for wakeups from interrupt contexts doesn't seem to be
> appropriate too but at least sync wakeup will do what it's supposed to do.
>
> Add a check to make sure that sync can only be set when in_task() is true. If
> a wakeup is from interrupt context, sync flag will be 0 because in_task()
> returns 0.
>
> Tested in two scenarios: wakeups from 1) task contexts, expected to see no
> performance changes; 2) interrupt contexts, expected to see better performance
> under low/medium load and no regression under heavy load.
>
> Use hackbench for scenario 1 and pgbench for scenarios 2 both from mmtests on
> a 2-socket Xeon E5-2699v3 box with 256G memory in total. Running 5.18 kernel
> with SELinux disabled. Scheduler/MM tunables are all default. Irqbalance
> daemon is active.
>
> Hackbench (config-scheduler-unbound)
> =========
> process-pipes:
> Baseline Patched
> Amean 1 0.4300 ( 0.00%) 0.4420 ( -2.79%)
> Amean 4 0.8757 ( 0.00%) 0.8774 ( -0.20%)
> Amean 7 1.3712 ( 0.00%) 1.3789 ( -0.56%)
> Amean 12 2.3541 ( 0.00%) 2.3714 ( -0.73%)
> Amean 21 4.2229 ( 0.00%) 4.2439 ( -0.50%)
> Amean 30 5.9113 ( 0.00%) 5.9451 ( -0.57%)
> Amean 48 9.3873 ( 0.00%) 9.4898 ( -1.09%)
> Amean 79 15.9281 ( 0.00%) 16.1385 ( -1.32%)
> Amean 110 22.0961 ( 0.00%) 22.3433 ( -1.12%)
> Amean 141 28.2973 ( 0.00%) 28.6209 ( -1.14%)
> Amean 172 34.4709 ( 0.00%) 34.9347 ( -1.35%)
> Amean 203 40.7621 ( 0.00%) 41.2497 ( -1.20%)
> Amean 234 47.0416 ( 0.00%) 47.6470 ( -1.29%)
> Amean 265 53.3048 ( 0.00%) 54.1625 ( -1.61%)
> Amean 288 58.0595 ( 0.00%) 58.8096 ( -1.29%)
>
> process-sockets:
> Baseline Patched
> Amean 1 0.6776 ( 0.00%) 0.6611 ( 2.43%)
> Amean 4 2.6183 ( 0.00%) 2.5769 ( 1.58%)
> Amean 7 4.5662 ( 0.00%) 4.4801 ( 1.89%)
> Amean 12 7.7638 ( 0.00%) 7.6201 ( 1.85%)
> Amean 21 13.5335 ( 0.00%) 13.2915 ( 1.79%)
> Amean 30 19.3369 ( 0.00%) 18.9811 ( 1.84%)
> Amean 48 31.0724 ( 0.00%) 30.6015 ( 1.52%)
> Amean 79 51.1881 ( 0.00%) 50.4251 ( 1.49%)
> Amean 110 71.3399 ( 0.00%) 70.4578 ( 1.24%)
> Amean 141 91.4675 ( 0.00%) 90.3769 ( 1.19%)
> Amean 172 111.7463 ( 0.00%) 110.3947 ( 1.21%)
> Amean 203 131.6927 ( 0.00%) 130.3270 ( 1.04%)
> Amean 234 151.7459 ( 0.00%) 150.1320 ( 1.06%)
> Amean 265 171.9101 ( 0.00%) 169.9751 ( 1.13%)
> Amean 288 186.9231 ( 0.00%) 184.7706 ( 1.15%)
>
> thread-pipes:
> Baseline Patched
> Amean 1 0.4523 ( 0.00%) 0.4535 ( -0.28%)
> Amean 4 0.9041 ( 0.00%) 0.9085 ( -0.48%)
> Amean 7 1.4111 ( 0.00%) 1.4146 ( -0.25%)
> Amean 12 2.3532 ( 0.00%) 2.3688 ( -0.66%)
> Amean 21 4.1550 ( 0.00%) 4.1701 ( -0.36%)
> Amean 30 6.1043 ( 0.00%) 6.2391 ( -2.21%)
> Amean 48 10.2077 ( 0.00%) 10.3511 ( -1.40%)
> Amean 79 16.7922 ( 0.00%) 17.0427 ( -1.49%)
> Amean 110 23.3350 ( 0.00%) 23.6522 ( -1.36%)
> Amean 141 29.6458 ( 0.00%) 30.2617 ( -2.08%)
> Amean 172 35.8649 ( 0.00%) 36.4225 ( -1.55%)
> Amean 203 42.4477 ( 0.00%) 42.8332 ( -0.91%)
> Amean 234 48.7117 ( 0.00%) 49.4042 ( -1.42%)
> Amean 265 54.9171 ( 0.00%) 55.6551 ( -1.34%)
> Amean 288 59.5282 ( 0.00%) 60.2903 ( -1.28%)
>
> thread-sockets:
> Baseline Patched
> Amean 1 0.6917 ( 0.00%) 0.6892 ( 0.37%)
> Amean 4 2.6651 ( 0.00%) 2.6017 ( 2.38%)
> Amean 7 4.6734 ( 0.00%) 4.5637 ( 2.35%)
> Amean 12 8.0156 ( 0.00%) 7.8079 ( 2.59%)
> Amean 21 14.0451 ( 0.00%) 13.6679 ( 2.69%)
> Amean 30 20.0963 ( 0.00%) 19.5657 ( 2.64%)
> Amean 48 32.4115 ( 0.00%) 31.6001 ( 2.50%)
> Amean 79 53.1104 ( 0.00%) 51.8395 ( 2.39%)
> Amean 110 74.0929 ( 0.00%) 72.4391 ( 2.23%)
> Amean 141 95.1506 ( 0.00%) 93.0992 ( 2.16%)
> Amean 172 116.1969 ( 0.00%) 113.8307 ( 2.04%)
> Amean 203 137.4413 ( 0.00%) 134.5247 ( 2.12%)
> Amean 234 158.5395 ( 0.00%) 155.2793 ( 2.06%)
> Amean 265 179.7729 ( 0.00%) 176.1099 ( 2.04%)
> Amean 288 195.5573 ( 0.00%) 191.3977 ( 2.13%)
>
> Pgbench (config-db-pgbench-timed-ro-small)
> =======
> Baseline Patched
> Hmean 1 68.54 ( 0.00%) 69.72 ( 1.71%)
> Hmean 6 27725.78 ( 0.00%) 34119.11 ( 23.06%)
> Hmean 12 55724.26 ( 0.00%) 63158.22 ( 13.34%)
> Hmean 22 72806.26 ( 0.00%) 73389.98 ( 0.80%)
> Hmean 30 79000.45 ( 0.00%) 75197.02 ( -4.81%)
> Hmean 48 78180.16 ( 0.00%) 75074.09 ( -3.97%)
> Hmean 80 75001.93 ( 0.00%) 70590.72 ( -5.88%)
> Hmean 110 74812.25 ( 0.00%) 74128.57 ( -0.91%)
> Hmean 142 74261.05 ( 0.00%) 72910.48 ( -1.82%)
> Hmean 144 75375.35 ( 0.00%) 71295.72 ( -5.41%)

The two tests kernels used are:

- tip at commit: 14b3f2d9ee8d "sched/fair: Disallow sync wakeup from interrupt context"
- tip at commit: 91caa5ae2424 "sched/core: Fix the bug that task won't enqueue into core tree when update cookie"

Below are the tbench result on a dual socket Zen3 machine
running in NPS1 mode. Following is the NUMA configuration
NPS1 mode:

- NPS1: Each socket is a NUMA node.
Total 2 NUMA nodes in the dual socket machine.

Node 0: 0-63, 128-191
Node 1: 64-127, 192-255

Clients: tip (91caa5ae2424) tip (14b3f2d9ee8d)
1 569.24 (0.00 pct) 283.63 (-50.17 pct) *
2 1104.76 (0.00 pct) 590.45 (-46.55 pct) *
4 2058.78 (0.00 pct) 1080.63 (-47.51 pct) *
8 3590.20 (0.00 pct) 2098.05 (-41.56 pct) *
16 6119.21 (0.00 pct) 4348.40 (-28.93 pct) *
32 11383.91 (0.00 pct) 8417.55 (-26.05 pct) *
64 21910.01 (0.00 pct) 19405.11 (-11.43 pct) *
128 33105.27 (0.00 pct) 29791.80 (-10.00 pct) *
256 45550.99 (0.00 pct) 45847.10 (0.65 pct)
512 57753.81 (0.00 pct) 49481.17 (-14.32 pct) *
1024 55684.33 (0.00 pct) 48748.38 (-12.45 pct) *

This regression is consistently reproducible.

Below are the statistics gathered from schedstat data:

Kernel : tip + remove 14b3f2d9ee8d tip
HEAD commit : 91caa5ae2424 14b3f2d9ee8d
sched_yield cnt : 11 17
Legacy counter can be ignored : 0 0
schedule called : 12621212 15104022
schedule left the processor idle : 6306653 ( 49.96% of times ) 7547375 ( 49.96% of times )
try_to_wake_up was called : 6310778 7552147
try_to_wake_up was called to wake up the local cpu : 12305 ( 0.19% of times ) 12816 ( 0.16% of times )
total time by tasks on this processor (in jiffies) : 78497712520 72461915902
total time waiting tasks on this processor (in jiffies) : 56398803 ( 0.07% of times ) 34631178 ( 0.04% of times )
total timeslices run on this cpu : 6314548 7556630

Wakeups on same SMT : 39 ( 0.00062 ) 263 ( 0.00348 )
Wakeups on same MC : 6297015 ( 99.78% of time ) <--- 1079 ( 0.01429 )
Wakeups on same DIE : 559 ( 0.00886 ) 7537909 ( 99.81147 ) <--- With the patch, the task will prefer
Wakeups on same NUMA : 860 ( 0.01363 ) 80 ( 0.00106 ) to wake on the same LLC where it previously
Affine wakeups on same SMT : 25 ( 0.00040 ) 255 ( 0.00338 ) ran as compared to the LLC of waker.
Affine wakeups on same MC : 6282684 ( 99.55% of time ) <--- 961 ( 0.01272 ) This results in performance degradation as
Affine wakeups on same DIE : 523 ( 0.00829 ) 7537220 ( 99.80235 ) <--- the task is farther away from data it will
Affine wakeups on same NUMA : 839 ( 0.01329 ) 46 ( 0.00061 ) consume next.

All the statistics are comparable except for the reduced number of affine
wakeup on the waker's LLC that resulting in task being placed on the previous
LLC farther away from the data that resides in the waker's LLC that the wakee
will consume next.

All wakeups in the tbench, happens in_serving_softirq() making in_taks() false
for all the cases where sync would have been true otherwise.

We wanted to highlight there are workloads which would still benefit from
affine wakeups even when it happens in an interrupt context. It would be
great if we could spot such cases and bias wakeups towards waker's LLC.

Other benchmarks results are comparable to the tip in most cases.
All benchmarks were run on machine configured in NPS1 mode.
Following are the results:

~~~~~~~~~
hackbench
~~~~~~~~~

Test: tip (91caa5ae2424) tip (14b3f2d9ee8d)
1-groups: 4.22 (0.00 pct) 4.48 (-6.16 pct) *
1-groups: 4.22 (0.00 pct) 4.30 (-1.89 pct) [Verification run]
2-groups: 5.01 (0.00 pct) 4.87 (2.79 pct)
4-groups: 5.49 (0.00 pct) 5.34 (2.73 pct)
8-groups: 5.64 (0.00 pct) 5.50 (2.48 pct)
16-groups: 7.54 (0.00 pct) 7.34 (2.65 pct)

~~~~~~~~
schbench
~~~~~~~~

#workers: tip (91caa5ae2424) tip (14b3f2d9ee8d)
1: 22.00 (0.00 pct) 22.00 (0.00 pct)
2: 22.00 (0.00 pct) 27.00 (-22.72 pct) * Known to have run to run
4: 33.00 (0.00 pct) 38.00 (-15.15 pct) * variations.
8: 48.00 (0.00 pct) 51.00 (-6.25 pct) *
16: 70.00 (0.00 pct) 70.00 (0.00 pct)
32: 118.00 (0.00 pct) 120.00 (-1.69 pct)
64: 217.00 (0.00 pct) 223.00 (-2.76 pct)
128: 485.00 (0.00 pct) 488.00 (-0.61 pct)
256: 1066.00 (0.00 pct) 1054.00 (1.12 pct)
512: 47296.00 (0.00 pct) 47168.00 (0.27 pct)

Note: schbench results at lower worker count have a large
run-to-run variance and depends on certain characteristics
of new-idle balance.

~~~~~~
stream
~~~~~~

- 10 runs

Test: tip (91caa5ae2424) tip (14b3f2d9ee8d)
Copy: 336140.45 (0.00 pct) 334362.29 (-0.52 pct)
Scale: 214679.13 (0.00 pct) 218016.44 (1.55 pct)
Add: 251691.67 (0.00 pct) 249734.04 (-0.77 pct)
Triad: 262174.93 (0.00 pct) 260063.57 (-0.80 pct)

- 100 runs

Test: tip (91caa5ae2424) tip (14b3f2d9ee8d)
Copy: 336576.38 (0.00 pct) 334646.27 (-0.57 pct)
Scale: 219124.86 (0.00 pct) 223480.29 (1.98 pct)
Add: 251796.93 (0.00 pct) 250845.95 (-0.37 pct)
Triad: 262286.47 (0.00 pct) 258020.57 (-1.62 pct)

~~~~~~~~~~~~
ycsb-mongodb
~~~~~~~~~~~~

tip (91caa5ae2424): 290479.00 (var: 1.53)
tip (14b3f2d9ee8d): 287361.67 (var: 0.80) (-1.07 pct)

>
> [..snip..]
>

We also ran tbench on a dual socket Icelake Xeon system (2 x 32C/64T)
(Intel Xeon Platinum 8362) and following are the results:

Clients: tip (91caa5ae2424) tip (14b3f2d9ee8d)
1 131.56 (0.00 pct) 145.05 (10.25 pct)
2 279.46 (0.00 pct) 245.94 (-11.99 pct) *
4 552.19 (0.00 pct) 577.94 (4.66 pct)
8 1176.12 (0.00 pct) 1309.40 (11.33 pct)
16 2685.86 (0.00 pct) 2724.96 (1.45 pct)
32 11582.50 (0.00 pct) 10817.42 (-6.60 pct) *
64 18309.47 (0.00 pct) 18287.92 (-0.11 pct)
128 15689.97 (0.00 pct) 15322.08 (-2.34 pct)
256 37130.34 (0.00 pct) 37332.91 (0.54 pct)
512 36448.68 (0.00 pct) 36162.17 (-0.78 pct)

Both kernel versions show lot of run to run variance but the two highlighted
cases are relatively more stable and the case with 32 clients is outside of
the NUMA imbalance threshold.

Below are detailed stats of each individual runs on Icelacke machine:

Kernel: tip tip
(91caa5ae2424) (14b3f2d9ee8d)

o 1 Clients

Min : 126.23 128.01
Max : 137.75 168.99
Median : 131.20 143.83
AMean : 131.73 146.94
GMean : 131.64 145.99
HMean : 131.56 145.05
AMean Stddev : 5.78 20.66
AMean CoefVar : 4.39 pct 14.06 pct

o 2 clients *

Min : 261.71 234.26
Max : 293.38 253.03
Median : 285.27 251.43
AMean : 280.12 246.24
GMean : 279.79 246.09
HMean : 279.46 245.94
AMean Stddev : 16.45 10.40
AMean CoefVar : 5.87 pct 4.22 pct

o 4 clients

Min : 515.98 497.81
Max : 611.56 744.49
Median : 537.71 543.82
AMean : 555.08 595.37
GMean : 553.61 586.31
HMean : 552.19 577.94
AMean Stddev : 50.10 131.17
AMean CoefVar : 9.03 pct 22.03 pct

o 8 clients

Min : 1101.30 1178.59
Max : 1293.18 1442.30
Median : 1150.15 1334.56
AMean : 1181.54 1318.48
GMean : 1178.80 1313.97
HMean : 1176.12 1309.40
AMean Stddev : 99.72 132.59
AMean CoefVar : 8.44 pct 10.06 pct

o 16 clients

Min : 2478.03 2522.28
Max : 2829.35 3043.49
Median : 2777.96 2660.31
AMean : 2695.11 2742.03
GMean : 2690.54 2733.37
HMean : 2685.86 2724.96
AMean Stddev : 189.75 270.04
AMean CoefVar : 7.04 pct 9.85 pct

o 32 clients *

Min : 11479.70 10569.00
Max : 11670.50 11154.90
Median : 11598.90 10744.90
AMean : 11583.03 10822.93
GMean : 11582.77 10820.17
HMean : 11582.50 10817.42
AMean Stddev : 96.38 300.64
AMean CoefVar : 0.83 pct 2.78 pct

Please let me know if you would like me to gather any more data
on the test system.
--
Thanks and Regards,
Prateek