Enabled gcServer and disabled gcConcurrent for test runners

[ivonin]

Description

Changed the GC settings to match the settings for MS Test v1 runners (QTAgent.exe)

Related issue

Fixes #1967

[msftclas]

All CLA requirements met.

[ivonin]

The check is failed with some acceptance tests broken, but I'm not sure if my changes could affect any of it... The tests have bad history and are failing randomly it seems?

[sharwell]

Locally I observed gcServer for testhost.x86 reducing test run time significantly. Here's a message I sent on the subject internally:

The only difference between these two test runs is I turned on server GC for testhost.x86.exe:

[4/2/2019 7:18:43 AM Informational] ------ Run test started ------
[4/2/2019 7:18:46 AM Informational] [xUnit.net 00:00:01.8254068]   Discovering: StyleCop.Analyzers.Test.CSharp7
[4/2/2019 7:18:48 AM Informational] [xUnit.net 00:00:03.9438646]   Discovered:  StyleCop.Analyzers.Test.CSharp7
[4/2/2019 7:18:48 AM Informational] [xUnit.net 00:00:03.9871306]   Starting:    StyleCop.Analyzers.Test.CSharp7
[4/2/2019 7:21:12 AM Informational] [xUnit.net 00:02:27.5838838]   Finished:    StyleCop.Analyzers.Test.CSharp7
[4/2/2019 7:21:13 AM Informational] ========== Run test finished: 6446 run (0:02:29.701225) ==========
[4/2/2019 7:21:52 AM Informational] ------ Run test started ------
[4/2/2019 7:21:55 AM Informational] [xUnit.net 00:00:01.6501179]   Discovering: StyleCop.Analyzers.Test.CSharp7
[4/2/2019 7:21:57 AM Informational] [xUnit.net 00:00:03.8203892]   Discovered:  StyleCop.Analyzers.Test.CSharp7
[4/2/2019 7:21:57 AM Informational] [xUnit.net 00:00:03.8693271]   Starting:    StyleCop.Analyzers.Test.CSharp7
[4/2/2019 7:22:33 AM Informational] [xUnit.net 00:00:40.0539226]   Finished:    StyleCop.Analyzers.Test.CSharp7
[4/2/2019 7:22:33 AM Informational] ========== Run test finished: 6446 run (0:00:41.4295855) ==========

Sam

I have not reviewed the scenario with gcConcurrent set.

[sharwell]

⚠️ After locally setting this I started to see frequent OutOfMemoryException failures in x86 test runs, which appear to be caused by the substantial increase in memory required to use Server GC on many cores (2990WX). It will be important to address #1985 at the same time.

[ivonin]

GCconcurrent=true makes performance slightly worse on our local test stand: 4 minutes 34 seconds median duration vs 4 minutes 28 seconds with GCconcurrent=false (this is the median from 10 runs).

QTAgent.exe has GCconcurrent=false, so combined with my benchmark it seems to be the right call.

[ivonin]

Some tests are indeed failing with my changes... I'll look into it shortly, but help will be appreciated

[singhsarab]

Closing stale PRs please re-open if required.

[sharwell]

I have not seen OOM errors since #1985 was addressed, even with gcServer enabled.

[yar-shukan]

@sharwell , @ivonin , can this be reopened?

There should be the ability to configure that through .runsettings file (I guess this is more complicated) or it should be the default setting as gcServer mode in theory should perform better for tests executed in parallel.

Considering the build agent case where it's not trivial (or even impossible due to permissions) to modify the shared vsconsole configuration I think enabling or giving the ability to configure gcServer mode has to be reconsidered again.