Restore: improve thread pool starvation fix by disabling asynchronous I/O

[dtivel]

Resolve NuGet/Home#7314.

The fix is to disable asynchronous I/O in the FileSystem class.

Ignoring the previous change, this change improves restore performance under .NET Core 2.1+ by ~70% using .NET Core 2.2 Preview 3 (2.2.100-preview3-009430) and the repro for https://github.com/dotnet/corefx/issues/31914.

Trial	Baseline (seconds)	With useAsync == false (seconds)
1	293.6	66.99
2	280.06	65.87
3	276.75	73.42
4	273.04	81.44
5	287.19	102.84
6	269.03	72.28
7	288.72	100.19
8	298.51	109.87
9	275.34	78.8
10	268.6	84.41
Average	281.084	83.611

All baseline trials ultimately failed with:

The SSL connection could not be established, see inner exception.
Authentication failed because the remote party has closed the transport stream.

None of the trials with the fix failed.

This fix was intended for .NET Core 2.1+ as its HTTP rewrite was the catalyst for this performance regression. However, I tested the same fix with NuGet.exe and found a ~44% improvement in restore performance.

Trial	Baseline (seconds)	With useAsync == false (seconds)
1	85.8	47.2
2	78.0	42.2
3	84.0	45.5
4	88.2	45.5
5	96.0	49.6
6	89.4	66.0
7	96.0	45.4
8	81.6	47.1
9	84.0	45.6
10	82.2	47.7
Average	86.5	48.2

[jainaashish]

💯

[loic-sharma]

However, I tested the same fix with NuGet.exe and found a ~44% improvement in restore performance.

Do we know why removing async improves performance here? Is this also due to the HTTP rewrite you mentioned?

[dtivel]

@loic-sharma, yes, we do.

See https://github.com/dotnet/corefx/issues/31914 and https://github.com/dotnet/corefx/issues/32837.