This performance test is failed - Akka.Streams.Tests.Performance.GraphBuilderBenchmark+Graph_with_1000_imported_flows
[FAIL] Expected Elapsed Time to must be less than or equal to 70.00 ms; actual value was 72.67 ms.

Also I've found a lot of similar exceptions in performance tests, like this

 Invoking setup for Akka.Streams.Tests.Performance.FlowSelectBenchmark+Flow_Select_100k_elements_with_buffer_8_and_1_select_and_without_GraphStage_Identifier
[00:56:45][Step 2/2] ERROR: Error occurred during $Akka.Streams.Tests.Performance.FlowSelectBenchmark+Flow_Select_100k_elements_with_buffer_8_and_1_select_and_without_GraphStage_Identifier SETUP.
[00:56:45][Step 2/2] System.ArgumentNullException: Value cannot be null.
[00:56:45][Step 2/2] Parameter name: stream
[00:56:45][Step 2/2]    at System.IO.StreamReader..ctor(Stream stream, Encoding encoding, Boolean detectEncodingFromByteOrderMarks, Int32 bufferSize, Boolean leaveOpen)
[00:56:45][Step 2/2]    at System.IO.StreamReader..ctor(Stream stream)
[00:56:45][Step 2/2]    at Akka.Configuration.ConfigurationFactory.FromResource(String resourceName, Assembly assembly) in D:\BuildAgent\work\d26c9d7f36545acd\src\core\Akka\Configuration\ConfigurationFactory.cs:line 135
[00:56:45][Step 2/2]    at Akka.Streams.Tests.Performance.FlowSelectBenchmark.Setup(BenchmarkContext context) in D:\BuildAgent\work\d26c9d7f36545acd\src\core\Akka.Streams.Tests.Performance\FlowSelectBenchmark.cs:line 56
[00:56:45][Step 2/2]    at NBench.Sdk.ReflectionBenchmarkInvoker.InvokePerfSetup(BenchmarkContext context)
[00:56:45][Step 2/2]    at NBench.Sdk.Benchmark.PreRun()

@Silv3rcircl3 you should check this test StreamLayout_should_not_fail_materialization_when_building_a_large_graph_with_simple_computation_when_using_Via

Xunit can't execute it and just fails. I've checked it 3 times

@alexvaluyskiy addressed your remarks

LGTM

just needs a rebase

@Aaronontheweb rebased but @Horusiath wanted to take a look too so he should merge it once he has reviewed it.

This is enormous amount of work, but guys - that performance drop seriously worries me (and no, setting perf threshold to be more forgiving is not a way to go). From what I know, performance in JVM never drops in higher versions, and I think we should not let this happen on our side.

This is strange, as i.e. Keep optimization and forgetting materialized values should make it faster.

The perf specs passed yesterday and today failed with this ~50% perf drop without any code changes. So I guess it has something to do with the build server.

For example the Graph_with_1000_junction failed in the last run
[FAIL] Expected Elapsed Time to must be less than or equal to 1,000.00 ms; actual value was 1,422.67 ms.

but passed the last 3 runs before
[PASS] Expected Elapsed Time to must be less than or equal to 1,000.00 ms; actual value was 808.33 ms.
[PASS] Expected Elapsed Time to must be less than or equal to 1,000.00 ms; actual value was 787.67 ms.
[PASS] Expected Elapsed Time to must be less than or equal to 1,000.00 ms; actual value was 771.33 ms.