MemoryMappedFile.CreateNew is 12-4 times slower on Linux compared to Windows

[adamsitnik]

MemoryMappedFile.CreateNew is 12-4 times slower on Linux compared to Windows. MemoryMappedFile.CreateFromFile is 3x faster so there is a chance that we are doing something wrong.

Slower	diff/base	Windows Median (ns)	Linux Median (ns)	Modality
System.IO.MemoryMappedFiles.Tests.Perf_MemoryMappedFile.CreateNew(capacity: 10000)	12.77	2321.27	29641.41
System.IO.MemoryMappedFiles.Tests.Perf_MemoryMappedFile.CreateNew(capacity: 100000)	12.12	2462.10	29849.96
System.IO.MemoryMappedFiles.Tests.Perf_MemoryMappedFile.CreateNew(capacity: 1000000)	8.53	3967.32	33824.88	bimodal
System.IO.MemoryMappedFiles.Tests.Perf_MemoryMappedFile.CreateNew(capacity: 10000000)	4.83	6632.03	32033.40	bimodal

Faster	Win/Lin	Windows Median (ns)	Linux Median (ns)	Modality
System.IO.MemoryMappedFiles.Tests.Perf_MemoryMappedFile.CreateFromFile(capacity: 10000000)	3.68	150431.05	40823.03
System.IO.MemoryMappedFiles.Tests.Perf_MemoryMappedFile.CreateFromFile(capacity: 1000000)	2.93	119037.20	40617.88
System.IO.MemoryMappedFiles.Tests.Perf_MemoryMappedFile.CreateFromFile(capacity: 10000)	2.75	112842.91	41088.96
System.IO.MemoryMappedFiles.Tests.Perf_MemoryMappedFile.CreateFromFile(capacity: 100000)	2.67	110914.50	41580.16	several?
The contributor who wants to work on this issue should:

1. Run this simple benchmark from dotnet/performance repository and confirm the problem

git clone https://github.com/dotnet/performance.git
python3 ./performance/scripts/benchmarks_ci.py -f netcoreapp5.0 --filter System.IO.MemoryMappedFiles.Tests.Perf_MemoryMappedFile.CreateNew

2. Build CoreFX locally: https://github.com/dotnet/performance/blob/master/docs/profiling-workflow-corefx.md#Build
3. Create a small repro app: https://github.com/dotnet/performance/blob/master/docs/profiling-workflow-corefx.md#repro
4. Use PerfCollect to identify issue https://github.com/dotnet/performance/blob/master/docs/profiling-workflow-corefx.md#PerfCollect
5. Solve the issue

[JeremyKuhne]

Triage: We need to investigate, but this might be unavoidable due to platform implementation differences.

[adamsitnik]

Based on the data available in dotnet/perf-autofiling-issues#40288 and dotnet/perf-autofiling-issues#40339 I can say that #105178 has addressed this issue. There is still a gap, but we use best available sys-calls now so there is very little we can do about it. Thank you again @am11 !