CompareExchange_long benchmark sometimes reports very long execution time on x86

[adamsitnik]

By looking at the data I got for x86 the CompareExchange_long benchmark typically reports time around 10ns, but sometimes it's even x100 more.

It's either a very weird hardware issue or a BenchmarkDotNet bug, possibly: dotnet/BenchmarkDotNet#837

The problem is that I can't reproduce it locally... so if anyone ever faces this problem please copy-paste the content of BDN .log file here.

git clone https://github.com/dotnet/performance.git
py .\performance\scripts\benchmarks_ci.py -f netcoreapp5.0 --filter System.Threading.Tests.Perf_Interlocked.CompareExchange_long --architecture x86

[AndyAyersMS]

Here's a repro... issue is lock cmpxchg8 splitting a cache line, so it requires a particular data alignment.

The code below has a 50/50 chance that one of those long fields does this.

public class X
{
     int x;
     long l1;
     long l2;
     long l3;
     long l4;
     long l5;
     long l6;
     long l7;
     long l8;
     
     [Benchmark]
     public long CompareExchange_long1() => Interlocked.CompareExchange(ref l1, 1, 0);
     
     [Benchmark]
     public long CompareExchange_long2() => Interlocked.CompareExchange(ref l2, 1, 0);
     
     [Benchmark]
     public long CompareExchange_long3() => Interlocked.CompareExchange(ref l3, 1, 0);
     
     [Benchmark]
     public long CompareExchange_long4() => Interlocked.CompareExchange(ref l4, 1, 0);
     
     [Benchmark]
     public long CompareExchange_long5() => Interlocked.CompareExchange(ref l5, 1, 0);
     
     [Benchmark]
     public long CompareExchange_long6() => Interlocked.CompareExchange(ref l6, 1, 0);
     
     [Benchmark]
     public long CompareExchange_long7() => Interlocked.CompareExchange(ref l7, 1, 0);
     
     [Benchmark]
     public long CompareExchange_long8() => Interlocked.CompareExchange(ref l8, 1, 0);
}

Sample data

BenchmarkDotNet=v0.12.1, OS=Windows 10.0.19041.572 (2004/?/20H1)
Intel Core i7-8650U CPU 1.90GHz (Kaby Lake R), 1 CPU, 8 logical and 4 physical cores
.NET Core SDK=5.0.100-rc.2.20479.15
  [Host]     : .NET Core 5.0.0 (CoreCLR 5.0.20.47505, CoreFX 5.0.20.47505), X86 RyuJIT
  DefaultJob : .NET Core 5.0.0 (CoreCLR 5.0.20.47505, CoreFX 5.0.20.47505), X86 RyuJIT

Method	Mean	Error	StdDev	Median
CompareExchange_long1	9.083 ns	0.5248 ns	0.5615 ns	9.009 ns
CompareExchange_long2	11.070 ns	0.4912 ns	0.4102 ns	11.115 ns
CompareExchange_long3	11.995 ns	0.5210 ns	0.9785 ns	11.782 ns
CompareExchange_long4	11.636 ns	0.5162 ns	0.6891 ns	11.585 ns
CompareExchange_long5	1,273.176 ns	25.7278 ns	71.7188 ns	1,254.676 ns
CompareExchange_long6	10.818 ns	0.7369 ns	2.1380 ns	9.710 ns
CompareExchange_long7	9.389 ns	0.2530 ns	0.2243 ns	9.346 ns
CompareExchange_long8	9.397 ns	0.3290 ns	0.3078 ns	9.449 ns

[AndyAyersMS]

Also explains why we don't see this on x64, because classes there are 8 byte aligned. On x86, classes are only 4 byte aligned.

[danmoseley]

cc @kunalspathak