Port Channels benchmarks

[adamsitnik]

Fixes #72

@ViktorHofer this is a good example of benchmarks that require to setup them once before running and produce very stable results.

Results reported by xunit-performance:

System.Threading.Channels.Performance.Tests.dll	Metric	Unit	Iterations	Average	STDEV.S	Min	Max
System.Threading.Channels.Tests.BoundedChannelPerfTests.PingPong	Duration	msec	4	2504.966	14.556	2486.916	2519.338
System.Threading.Channels.Tests.BoundedChannelPerfTests.PingPong	Allocation Size on Benchmark Execution Thread	bytes	4	-6874.000	34188.135	-39888.000	29808.000
System.Threading.Channels.Tests.BoundedChannelPerfTests.ReadAsyncThenWriteAsync	Duration	msec	100	92.164	3.370	89.415	116.241
System.Threading.Channels.Tests.BoundedChannelPerfTests.ReadAsyncThenWriteAsync	Allocation Size on Benchmark Execution Thread	bytes	100	0.000	0.000	0.000	0.000
System.Threading.Channels.Tests.BoundedChannelPerfTests.TryWriteThenTryRead	Duration	msec	100	49.657	0.656	48.887	53.348
System.Threading.Channels.Tests.BoundedChannelPerfTests.TryWriteThenTryRead	Allocation Size on Benchmark Execution Thread	bytes	100	0.000	0.000	0.000	0.000
System.Threading.Channels.Tests.BoundedChannelPerfTests.WriteAsyncThenReadAsync	Duration	msec	100	77.709	6.011	72.436	103.316
System.Threading.Channels.Tests.BoundedChannelPerfTests.WriteAsyncThenReadAsync	Allocation Size on Benchmark Execution Thread	bytes	100	0.000	0.000	0.000	0.000
System.Threading.Channels.Tests.SpscUnboundedChannelPerfTests.PingPong	Duration	msec	5	2451.046	35.630	2414.985	2505.484
System.Threading.Channels.Tests.SpscUnboundedChannelPerfTests.PingPong	Allocation Size on Benchmark Execution Thread	bytes	5	-440.000	29448.079	-40208.000	42952.000
System.Threading.Channels.Tests.SpscUnboundedChannelPerfTests.ReadAsyncThenWriteAsync	Duration	msec	100	91.562	0.901	90.826	97.535
System.Threading.Channels.Tests.SpscUnboundedChannelPerfTests.ReadAsyncThenWriteAsync	Allocation Size on Benchmark Execution Thread	bytes	100	0.000	0.000	0.000	0.000
System.Threading.Channels.Tests.SpscUnboundedChannelPerfTests.TryWriteThenTryRead	Duration	msec	100	30.574	0.191	30.373	31.380
System.Threading.Channels.Tests.SpscUnboundedChannelPerfTests.TryWriteThenTryRead	Allocation Size on Benchmark Execution Thread	bytes	100	0.000	0.000	0.000	0.000
System.Threading.Channels.Tests.SpscUnboundedChannelPerfTests.WriteAsyncThenReadAsync	Duration	msec	100	72.449	0.744	71.742	75.072
System.Threading.Channels.Tests.SpscUnboundedChannelPerfTests.WriteAsyncThenReadAsync	Allocation Size on Benchmark Execution Thread	bytes	100	0.000	0.000	0.000	0.000
System.Threading.Channels.Tests.UnboundedChannelPerfTests.PingPong	Duration	msec	4	2565.998	32.485	2525.836	2602.982
System.Threading.Channels.Tests.UnboundedChannelPerfTests.PingPong	Allocation Size on Benchmark Execution Thread	bytes	4	-7770.000	184523.815	-256784.000	189120.000
System.Threading.Channels.Tests.UnboundedChannelPerfTests.ReadAsyncThenWriteAsync	Duration	msec	94	107.246	1.138	106.359	113.047
System.Threading.Channels.Tests.UnboundedChannelPerfTests.ReadAsyncThenWriteAsync	Allocation Size on Benchmark Execution Thread	bytes	94	0.000	0.000	0.000	0.000
System.Threading.Channels.Tests.UnboundedChannelPerfTests.TryWriteThenTryRead	Duration	msec	100	44.448	2.991	42.320	64.041
System.Threading.Channels.Tests.UnboundedChannelPerfTests.TryWriteThenTryRead	Allocation Size on Benchmark Execution Thread	bytes	100	0.000	0.000	0.000	0.000
System.Threading.Channels.Tests.UnboundedChannelPerfTests.WriteAsyncThenReadAsync	Duration	msec	100	72.118	4.670	66.332	84.937
System.Threading.Channels.Tests.UnboundedChannelPerfTests.WriteAsyncThenReadAsync	Allocation Size on Benchmark Execution Thread	bytes	100	0.000	0.000	0.000	0.000

Results reported by BenchmarkDotNet:

Type	Method	Mean	Error	StdDev	Median	Min	Max	Gen 0	Allocated
BoundedChannelPerfTests	TryWriteThenTryRead	51.29 ms	0.1146 ms	0.1016 ms	51.29 ms	51.12 ms	51.46 ms	-	0 B
SpscUnboundedChannelPerfTests	TryWriteThenTryRead	29.64 ms	0.0197 ms	0.0184 ms	29.65 ms	29.61 ms	29.67 ms	-	0 B
UnboundedChannelPerfTests	TryWriteThenTryRead	42.51 ms	0.1769 ms	0.1477 ms	42.47 ms	42.34 ms	42.87 ms	-	0 B
BoundedChannelPerfTests	WriteAsyncThenReadAsync	73.55 ms	2.0606 ms	1.9275 ms	72.83 ms	72.38 ms	79.19 ms	-	0 B
SpscUnboundedChannelPerfTests	WriteAsyncThenReadAsync	57.50 ms	0.1168 ms	0.0975 ms	57.47 ms	57.34 ms	57.67 ms	-	0 B
UnboundedChannelPerfTests	WriteAsyncThenReadAsync	69.54 ms	2.2016 ms	2.4470 ms	69.72 ms	66.63 ms	74.75 ms	-	0 B
BoundedChannelPerfTests	ReadAsyncThenWriteAsync	93.21 ms	2.7695 ms	3.0783 ms	91.41 ms	90.31 ms	100.63 ms	-	0 B
SpscUnboundedChannelPerfTests	ReadAsyncThenWriteAsync	92.76 ms	2.2358 ms	2.5748 ms	91.40 ms	90.85 ms	99.89 ms	-	0 B
UnboundedChannelPerfTests	ReadAsyncThenWriteAsync	109.82 ms	2.7086 ms	3.1192 ms	108.83 ms	106.55 ms	118.01 ms	-	0 B
BoundedChannelPerfTests	PingPong	2,544.90 ms	25.1021 ms	23.4805 ms	2,547.47 ms	2,512.98 ms	2,577.61 ms	20000.0000	792 B
SpscUnboundedChannelPerfTests	PingPong	2,441.26 ms	20.7453 ms	19.4052 ms	2,440.15 ms	2,398.93 ms	2,486.24 ms	20000.0000	792 B
UnboundedChannelPerfTests	PingPong	2,606.69 ms	34.3391 ms	32.1208 ms	2,607.97 ms	2,556.90 ms	2,659.98 ms	20000.0000	792 B

Hint: To join the results of multiple types from the same namespace into a single summary I used --join combined with --namespace filter (dotnet run -c Release -f netcoreapp2.1 -- --namespace=System.Threading.Channels.Tests --join)

Comments:

1. BenchmarkDotNet needed 205.31 sec to run the benchmarks, xunit-performance only half of it. The reason is that xunit-performance does not run many iterations for time-consuming benchmarks. In this case for the PingPong benchmark which takes 2.5s to execute BDN run 20 iterations when xunit-performance run 5 iterations. I added the info about this to Ideas for reducing the time required to run all benchmarks #78
2. BenchmarkDotNet produced much better results (smaller standard deviation, the distribution was more narrow)
3. For some reason xunit-performance reported negative values for allocated memory for PingPong benchmarks.
4. PingPongBenchmarks suffer from MemoryDiagnoser should include memory allocated by all Threads that were live during benchmark execution BenchmarkDotNet#723 which could be solved by https://github.com/dotnet/corefx/issues/30644 (we show the number of allocated bytes per single thread, whereas the benchmarks include many)

[jorive]

1_000_000 [](start = 32, length = 9)

This magic number is in 6 different places in the source file. Could we make it a const int?

[adamsitnik]

@jorive done!

[jorive]

LGTM modulo comment.