Gradual regression of System.Buffers.Tests.ReadOnlySequenceTests<Char>.IterateGetPositionArray from net6.0 to net7.0-rc2

[jozkee]

This benchmark has been jumping up and down throughout the whole net7.0 cycle but looking at the 6.0 data point seems like it gradually regressed.

Link to the chart:
https://pvscmdupload.blob.core.windows.net/reports/allTestHistory/refs/heads/main_x64_ubuntu%2018.04/System.Buffers.Tests.ReadOnlySequenceTests(Char).IterateGetPositionSingleSegment.html

Screenshot:

Benchmark results from 7.0-RC2 vs 6.0:

System.Buffers.Tests.ReadOnlySequenceTests.IterateGetPositionArray

Result	Ratio	Alloc Delta	Operating System	Bit	Processor Name	Modality
Slower	0.77	+0	ubuntu 18.04	Arm64	Unknown processor
Noise	-	+0	Windows 11	Arm64	Unknown processor
Slower	0.86	+0	Windows 11	Arm64	Microsoft SQ1 3.0 GHz
Slower	0.90	+0	Windows 11	Arm64	Microsoft SQ1 3.0 GHz
Same	0.92	+0	macOS Monterey 12.6	Arm64	Apple M1
Same	0.94	+0	macOS Monterey 12.6	Arm64	Apple M1 Max
Slower	0.67	+0	Windows 10	X64	Intel Core i7-8650U CPU 1.90GHz (Kaby Lake R)	bimodal
Noise	-	+0	Windows 11	X64	AMD Ryzen Threadripper PRO 3945WX 12-Cores
Same	0.98	+0	Windows 11	X64	AMD Ryzen 9 5900X
Same	0.95	+0	Windows 11	X64	AMD Ryzen 9 7950X
Slower	0.84	+0	Windows 11	X64	Intel Core i7-8700 CPU 3.20GHz (Coffee Lake)
Slower	0.73	+0	debian 11	X64	Intel Core i7-8700 CPU 3.20GHz (Coffee Lake)
Same	1.07	+0	ubuntu 18.04	X64	AMD Ryzen 9 5900X
Slower	0.84	+0	ubuntu 18.04	X64	Intel Xeon CPU E5-1650 v4 3.60GHz
Same	0.88	+0	ubuntu 20.04	X64	AMD Ryzen 9 5900X	several?
Same	0.94	+0	ubuntu 20.04	X64	Intel Core i7-8650U CPU 1.90GHz (Kaby Lake R)
Same	0.93	+0	ubuntu 20.04	X64	Intel Core i7-8700 CPU 3.20GHz (Coffee Lake)
Slower	0.87	+0	macOS Big Sur 11.7	X64	Intel Core i5-4278U CPU 2.60GHz (Haswell)
Slower	0.85	+0	macOS Monterey 12.6	X64	Intel Core i7-4870HQ CPU 2.50GHz (Haswell)

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

Issue Details

---

This benchmark has been jumping up and down throughout the whole net7.0 cycle but looking at the 6.0 data point seems like it gradually regressed.

Link to the chart:
https://pvscmdupload.blob.core.windows.net/reports/allTestHistory/refs/heads/main_x64_ubuntu%2018.04/System.Buffers.Tests.ReadOnlySequenceTests(Char).IterateGetPositionSingleSegment.html

Screenshot:

Benchmark results from 7.0-RC2 vs 6.0:

System.Buffers.Tests.ReadOnlySequenceTests.IterateGetPositionArray

Result	Ratio	Alloc Delta	Operating System	Bit	Processor Name	Modality
Slower	0.77	+0	ubuntu 18.04	Arm64	Unknown processor
Noise	-	+0	Windows 11	Arm64	Unknown processor
Slower	0.86	+0	Windows 11	Arm64	Microsoft SQ1 3.0 GHz
Slower	0.90	+0	Windows 11	Arm64	Microsoft SQ1 3.0 GHz
Same	0.92	+0	macOS Monterey 12.6	Arm64	Apple M1
Same	0.94	+0	macOS Monterey 12.6	Arm64	Apple M1 Max
Slower	0.67	+0	Windows 10	X64	Intel Core i7-8650U CPU 1.90GHz (Kaby Lake R)	bimodal
Noise	-	+0	Windows 11	X64	AMD Ryzen Threadripper PRO 3945WX 12-Cores
Same	0.98	+0	Windows 11	X64	AMD Ryzen 9 5900X
Same	0.95	+0	Windows 11	X64	AMD Ryzen 9 7950X
Slower	0.84	+0	Windows 11	X64	Intel Core i7-8700 CPU 3.20GHz (Coffee Lake)
Slower	0.73	+0	debian 11	X64	Intel Core i7-8700 CPU 3.20GHz (Coffee Lake)
Same	1.07	+0	ubuntu 18.04	X64	AMD Ryzen 9 5900X
Slower	0.84	+0	ubuntu 18.04	X64	Intel Xeon CPU E5-1650 v4 3.60GHz
Same	0.88	+0	ubuntu 20.04	X64	AMD Ryzen 9 5900X	several?
Same	0.94	+0	ubuntu 20.04	X64	Intel Core i7-8650U CPU 1.90GHz (Kaby Lake R)
Same	0.93	+0	ubuntu 20.04	X64	Intel Core i7-8700 CPU 3.20GHz (Coffee Lake)
Slower	0.87	+0	macOS Big Sur 11.7	X64	Intel Core i5-4278U CPU 2.60GHz (Haswell)
Slower	0.85	+0	macOS Monterey 12.6	X64	Intel Core i7-4870HQ CPU 2.50GHz (Haswell)

Author:	Jozkee
Assignees:	-
Labels:	tenet-performance, tenet-performance-benchmarks, area-CodeGen-coreclr
Milestone:	-

[AndyAyersMS]

I'll take a look.

[AndyAyersMS]

Locally I see only a small regression:

BenchmarkDotNet=v0.13.1.1847-nightly, OS=Windows 11 (10.0.22621.674)
Intel Core i7-8700 CPU 3.20GHz (Coffee Lake), 1 CPU, 12 logical and 6 physical cores
.NET SDK=7.0.100-rc.2.22477.23
[Host] : .NET 6.0.10 (6.0.1022.47605), X64 RyuJIT AVX2
Job-GJZRLG : .NET 6.0.10 (6.0.1022.47605), X64 RyuJIT AVX2
Job-FDXEMW : .NET 7.0.0 (7.0.22.47203), X64 RyuJIT AVX2

PowerPlanMode=00000000-0000-0000-0000-000000000000 IterationTime=250.0000 ms MaxIterationCount=20
MinIterationCount=15 WarmupCount=1

Method	Job	Runtime	Toolchain	Mean	Error	StdDev	Median	Min	Max	Ratio	RatioSD	Allocated	Alloc Ratio
IterateGetPositionSingleSegment	Job-GJZRLG	.NET 6.0	net6.0	25.38 ns	0.890 ns	1.025 ns	25.23 ns	23.90 ns	27.47 ns	1.00	0.00	-	NA
IterateGetPositionSingleSegment	Job-FDXEMW	.NET 7.0	net7.0	26.25 ns	0.295 ns	0.262 ns	26.32 ns	25.54 ns	26.56 ns	1.03	0.05	-	NA

Profiling (7.0) shows:

[AndyAyersMS]

x64 windows codegen for IterateGetPosition is very similar in 6 and 7. Some minor differences in allocation and in the tail part of the method, despite the fact that in 6 this method bypassed tiering.

Codegen for IterateGetPositionSingleSegment is even closer, the only difference is that in 7 we call IterateGetPosition indirectly.

However, just realized this is an x64 ubuntu issue, not a windows one. FWIW windows perf does not show this slow creep:

So let me redo some of the above looking at ubuntu x64 codegen.

[AndyAyersMS]

Perf has since improved and then regressed back a bit.

Jan 2023 improvement looks like it was #81095
April 2023 regression possibly a pgo update #83624

[AndyAyersMS]

Subsequently has improved again, and now faster than it's been; likely explanation was the pgo update in #85275