Implement APIs for some threading metrics (CoreCLR)

[kouvel]

• API review: https://github.com/dotnet/corefx/issues/35500

[kouvel]

There was no measurable change to perf from the changes

[kouvel]

Some perf numbers:

• Left score = Before any changes, work items per ms
• Right score = After any changes, work items per ms

Initial change:

CoreCLR 4-core 4-thread 1-node          Left score      Right score     ∆ Score %
--------------------------------------  --------------  --------------  ---------
TaskBurstWorkThroughput 1PcT 000.5PcWi   670.88 ±0.20%   677.22 ±0.16%      0.95%
TaskBurstWorkThroughput 1PcT 001.0PcWi  1642.87 ±0.39%  1661.91 ±0.45%      1.16%
TaskBurstWorkThroughput 1PcT 004.0PcWi  1859.70 ±0.61%  1877.41 ±0.63%      0.95%
TaskBurstWorkThroughput 1PcT 016.0PcWi  1696.58 ±0.60%  1681.65 ±0.57%     -0.88%
TaskBurstWorkThroughput 1PcT 064.0PcWi  1843.14 ±0.70%  1872.47 ±0.50%      1.59%
TaskBurstWorkThroughput 1PcT 256.0PcWi  2025.88 ±0.62%  2015.75 ±0.54%     -0.50%
TaskSustainedWorkThroughput 1PcT        2906.67 ±0.25%  2913.72 ±0.76%      0.24%
--------------------------------------  --------------  --------------  ---------

CoreRT 4-core 4-thread 1-node           Left score      Right score     ∆ Score %
--------------------------------------  --------------  --------------  ---------
TaskBurstWorkThroughput 1PcT 000.5PcWi   403.73 ±0.21%   407.09 ±0.12%      0.83%
TaskBurstWorkThroughput 1PcT 001.0PcWi   976.50 ±0.11%   981.36 ±0.08%      0.50%
TaskBurstWorkThroughput 1PcT 004.0PcWi  2913.15 ±0.07%  2892.59 ±0.10%     -0.71%
TaskBurstWorkThroughput 1PcT 016.0PcWi  2610.12 ±0.06%  2589.73 ±0.08%     -0.78%
TaskBurstWorkThroughput 1PcT 064.0PcWi  3057.35 ±0.11%  3038.80 ±0.10%     -0.61%
TaskBurstWorkThroughput 1PcT 256.0PcWi  3235.46 ±0.11%  3207.21 ±0.13%     -0.87%
TaskSustainedWorkThroughput 1PcT        3943.24 ±0.17%  3924.89 ±0.08%     -0.47%
--------------------------------------  --------------  --------------  ---------

CoreCLR 48-core 48-thread 4-node        Left score      Right score     ∆ Score %
--------------------------------------  --------------  --------------  ---------
TaskBurstWorkThroughput 1PcT 000.5PcWi   878.83 ±0.40%   877.38 ±0.40%     -0.16%
TaskBurstWorkThroughput 1PcT 001.0PcWi  4541.56 ±0.74%  4647.15 ±0.90%      2.32%
TaskBurstWorkThroughput 1PcT 004.0PcWi  5295.10 ±0.30%  5431.67 ±0.33%      2.58%
TaskBurstWorkThroughput 1PcT 016.0PcWi  5301.83 ±0.26%  5356.74 ±0.44%      1.04%
TaskBurstWorkThroughput 1PcT 064.0PcWi  5284.73 ±0.22%  5321.00 ±0.39%      0.69%
TaskBurstWorkThroughput 1PcT 256.0PcWi  5323.88 ±0.10%  5264.35 ±0.39%     -1.12%
TaskSustainedWorkThroughput 1PcT        7874.07 ±0.88%  8029.77 ±0.66%      1.98%
--------------------------------------  --------------  --------------  ---------

CoreRT 48-core 48-thread 4-node         Left score      Right score     ∆ Score %
--------------------------------------  --------------  --------------  ---------
TaskBurstWorkThroughput 1PcT 000.5PcWi  3546.97 ±0.12%  3264.71 ±0.17%     -7.96%
TaskBurstWorkThroughput 1PcT 001.0PcWi  2133.60 ±0.23%  1823.23 ±0.49%    -14.55%
TaskBurstWorkThroughput 1PcT 004.0PcWi  4514.30 ±0.30%  3846.87 ±0.12%    -14.78%
TaskBurstWorkThroughput 1PcT 016.0PcWi  4867.67 ±0.29%  4115.50 ±0.21%    -15.45%
TaskBurstWorkThroughput 1PcT 064.0PcWi  4927.05 ±0.32%  4144.38 ±0.35%    -15.89%
TaskBurstWorkThroughput 1PcT 256.0PcWi  4541.31 ±0.50%  3946.43 ±0.45%    -13.10%
TaskSustainedWorkThroughput 1PcT        7061.97 ±0.41%  6064.36 ±0.37%    -14.13%
--------------------------------------  --------------  --------------  ---------

After change to use thread-locals for more things (monitor lock contention counting in CoreCLR, all of the relevant counting in CoreRT):

CoreCLR 4-core 4-thread 1-node          Left score      Right score     ∆ Score %
--------------------------------------  --------------  --------------  ---------
TaskBurstWorkThroughput 1PcT 000.5PcWi   673.14 ±0.28%   672.85 ±0.25%     -0.04%
TaskBurstWorkThroughput 1PcT 001.0PcWi  1658.72 ±0.32%  1669.33 ±0.23%      0.64%
TaskBurstWorkThroughput 1PcT 004.0PcWi  1845.45 ±0.49%  1874.94 ±0.51%      1.60%
TaskBurstWorkThroughput 1PcT 016.0PcWi  1674.97 ±0.68%  1687.24 ±0.72%      0.73%
TaskBurstWorkThroughput 1PcT 064.0PcWi  1846.55 ±0.46%  1839.79 ±0.58%     -0.37%
TaskBurstWorkThroughput 1PcT 256.0PcWi  2024.60 ±0.35%  2034.10 ±0.31%      0.47%
TaskSustainedWorkThroughput 1PcT        2904.78 ±0.33%  3051.85 ±0.42%      5.06%
--------------------------------------  --------------  --------------  ---------

CoreRT 4-core 4-thread 1-node           Left score      Right score     ∆ Score %
--------------------------------------  --------------  --------------  ---------
TaskBurstWorkThroughput 1PcT 000.5PcWi   405.24 ±0.13%   406.74 ±0.17%      0.37%
TaskBurstWorkThroughput 1PcT 001.0PcWi   978.79 ±0.06%   985.89 ±0.06%      0.73%
TaskBurstWorkThroughput 1PcT 004.0PcWi  2934.53 ±0.09%  2922.13 ±0.11%     -0.42%
TaskBurstWorkThroughput 1PcT 016.0PcWi  2620.20 ±0.05%  2592.52 ±0.06%     -1.06%
TaskBurstWorkThroughput 1PcT 064.0PcWi  3055.01 ±0.16%  3041.20 ±0.07%     -0.45%
TaskBurstWorkThroughput 1PcT 256.0PcWi  3243.40 ±0.08%  3237.24 ±0.12%     -0.19%
TaskSustainedWorkThroughput 1PcT        3984.48 ±0.09%  3974.17 ±0.08%     -0.26%
--------------------------------------  --------------  --------------  ---------

CoreCLR 48-core 48-thread 4-node        Left score      Right score     ∆ Score %
--------------------------------------  --------------  --------------  ---------
TaskBurstWorkThroughput 1PcT 000.5PcWi   874.75 ±0.33%   706.56 ±0.46%    -19.23%
TaskBurstWorkThroughput 1PcT 001.0PcWi  4472.33 ±1.28%  2294.10 ±1.53%    -48.70%
TaskBurstWorkThroughput 1PcT 004.0PcWi  5324.16 ±0.33%  5411.74 ±0.44%      1.64%
TaskBurstWorkThroughput 1PcT 016.0PcWi  5321.38 ±0.27%  5379.79 ±0.13%      1.10%
TaskBurstWorkThroughput 1PcT 064.0PcWi  5267.20 ±0.39%  5419.75 ±0.20%      2.90%
TaskBurstWorkThroughput 1PcT 256.0PcWi  5241.08 ±0.46%  5389.11 ±0.26%      2.82%
TaskSustainedWorkThroughput 1PcT        7535.42 ±0.43%  7880.48 ±0.70%      4.58%
--------------------------------------  --------------  --------------  ---------

CoreRT 48-core 48-thread 4-node         Left score      Right score     ∆ Score %
--------------------------------------  --------------  --------------  ---------
TaskBurstWorkThroughput 1PcT 000.5PcWi  3531.92 ±0.11%  3302.87 ±0.11%     -6.49%
TaskBurstWorkThroughput 1PcT 001.0PcWi  2153.34 ±0.23%  1971.45 ±0.36%     -8.45%
TaskBurstWorkThroughput 1PcT 004.0PcWi  4518.53 ±0.21%  3803.60 ±0.14%    -15.82%
TaskBurstWorkThroughput 1PcT 016.0PcWi  4818.32 ±0.24%  4002.37 ±0.28%    -16.93%
TaskBurstWorkThroughput 1PcT 064.0PcWi  4917.06 ±0.43%  4065.70 ±0.29%    -17.31%
TaskBurstWorkThroughput 1PcT 256.0PcWi  4507.61 ±0.50%  3863.19 ±0.48%    -14.30%
TaskSustainedWorkThroughput 1PcT        6984.64 ±0.40%  6150.49 ±0.24%    -11.94%
--------------------------------------  --------------  --------------  ---------

Summary:

• Updated task burst tests to scale a bit better, code is here and I had modified it to use a medium-length delay (RandomMediumDelay) for the numbers above
• None of these regressions or improvements are caused by either change (counting with interlocked or thread-local)
• According to profiles, perf changes appear to be in code that was not changed. Perhaps slight changes in timing are affecting where and how badly contention occurs.
• For monitor lock contention counting, I tried a test that has multiple threads wait on an already acquired lock, then when all threads reach a waiting state the lock is released, and this repeats. I did not see any noticeable contention from using interlocked counting for Monitor lock contentions before waiting.
• Decided to go with the thread-local approach for all of the relevant counting, as in the profiles it doesn't appear to be any worse than counting with interlocked operations, and it shouldn't be a cause of contention in the future. At the moment the difference doesn't seem to be noticeable. For thread pool, probably there are other bottlenecks that would all need to be improved to see a difference.

[kouvel]

The RT implementation is broken, ignore those perf numbers, I'll fix and retest

[kouvel]

Perf numbers after fixes were similar to before, updated inline above

[kouvel]

@dotnet-bot test this please

[kouvel]

Ping for review please

[kouvel]

@dotnet-bot test Ubuntu arm Cross Checked crossgen_comparison Build and Test

[jkotas]

I have been waiting for API review to be done before looking at this.

[kouvel]

Oh ok, I'll ping that one as well

[kouvel]

The API review was approved. Couple of pending items is:

• Whether to expose PendingLocalWorkItemCount and PendingGlobalWorkItemCount. I have asked for more info on whether they would be useful, by default I'll go ahead and remove them
• Whether we want to expose CompletedWorkItemCount at all.
  • I think that's the most expensive one to track of the lot, as it involves a bit of extra work per work item
  • It is possible to mostly eliminate the extra cost for managed work items in CoreCLR and CoreRT. Work items that may have some extra cost are timer callbacks in CoreRT Windows and IO completion work items in CoreCLR Windows.
  • The info would enable easily determining work item throughput and whether the thread pool is stalled

[jkotas]

LGTM

[jkotas]

Preserve flush for hill climbing

Why do you think we need the flush for hill climbing? Any reasonable processor will flush the caches to the main memory at reasonable rate. I would think that it is fine for the hill climbing to miss last few microsecond worth of updates.

[kouvel]

Why do you think we need the flush for hill climbing?

Hill climbing wants to know what the current counts are so that it can respond and having the flush may help it respond sooner. Some processors may take longer than others to sync those updates. I doubt that it would take long enough to sync those updates to be comparable to the hill climbing interval, I'm not sure about that. I don't think it will make a noticeable difference on recent processors, not sure about older ones.

[jkotas]

having the flush may help it respond sooner.

At the cost of slowing down every processor in the system...

[kouvel]

The hill climbing interval is between 10 and 200 ms, higher when the thread count is already at the min number of threads. I'm not sure if the perf hit would be significant on those time scales. I also doubt that the flush would improve anything, I just kept it because I don't have all of the information on why it was there to begin with.

[kouvel]

Changed my mind again, I'll just remove it. If not having a flush would be an issue it likely would have shown up in a bunch of other places before.

[kouvel]

@dotnet-bot test Windows_NT x64 Checked CoreFX Tests

[kouvel]

Merged in #24113