Potential performance regression in Array.Sort for arrays of reference types

[adamsitnik]

git clone https://github.com/dotnet/performance.git
py .\scripts\benchmarks_ci.py -f net5.0 net6.0 --filter "System.Collections.Sort<IntClass>.Array" --bdn-arguments "--miniterationcount 100 --maxiterationcount 101 --outliers dontremove --memoryrandomization true"

Both Windows and Unix are affected:

https://pvscmdupload.blob.core.windows.net/reports/allTestHistory%2frefs%2fheads%2fmain_x64_Windows%2010.0.18362%2fSystem.Collections.Sort(IntClass).Array(Size%3a%20512).html

https://pvscmdupload.blob.core.windows.net/reports/allTestHistory%2frefs%2fheads%2fmain_x64_ubuntu%2018.04%2fSystem.Collections.Sort(IntClass).Array(Size%3a%20512).html

[ghost]

Tagging subscribers to this area: @dotnet/area-system-runtime
See info in area-owners.md if you want to be subscribed.

Issue Details

---

git clone https://github.com/dotnet/performance.git
py .\scripts\benchmarks_ci.py -f net5.0 net6.0 --filter "System.Collections.Sort<IntClass>.Array" --bdn-arguments "--miniterationcount 100 --maxiterationcount 101 --outliers dontremove --memoryrandomization true"

Both Windows and Unix are affected:

https://pvscmdupload.blob.core.windows.net/reports/allTestHistory%2frefs%2fheads%2fmain_x64_Windows%2010.0.18362%2fSystem.Collections.Sort(IntClass).Array(Size%3a%20512).html

https://pvscmdupload.blob.core.windows.net/reports/allTestHistory%2frefs%2fheads%2fmain_x64_ubuntu%2018.04%2fSystem.Collections.Sort(IntClass).Array(Size%3a%20512).html

Author:	adamsitnik
Assignees:	-
Labels:	area-System.Runtime, tenet-performance
Milestone:	-

[adamsitnik]

There are more array sorting-related benchmarks that have regressed across multiple configs, so the better config would be:

--filter 'System.Collections.Sort*Array*'

[danmoseley]

Is it possible to isolate a commit range? Seems like a sharp change. Might help see whether PGO is a suspect or not.

[jeffhandley]

@DrewScoggins Could you help narrow down the commit range for this regression?

[DrewScoggins]

7b7954a...0df29f7

[DrewScoggins]

There is a PGO update in there, as well as an improvement for division by unsigned constants. Those are the only two things that look like they could be making a difference.

[jeffhandley]

@AndyAyersMS Do you think PGO could be the cause of this based on Drew's findings. (Thanks, @DrewScoggins!)

Since this was all the way back in May, I'm going to go ahead and move this to 7.0.0, as I can't imagine we'd change anything at this point in 6.0 for this regression.

[dakersnar]

Seems to have improved over time. https://pvscmdupload.blob.core.windows.net/reports/allTestHistory%2frefs%2fheads%2fmain_x64_Windows%2010.0.18362%2fSystem.Collections.Sort(IntClass).Array(Size%3a%20512).html