From c86ae1c4ad0183cf0987ec2da70b309aea01617b Mon Sep 17 00:00:00 2001
From: Bernhard Urban-Forster <lewurm@gmail.com>
Date: Thu, 31 Oct 2019 22:48:07 +0100
Subject: [PATCH] [mono] limit DegreeOfParallelism to 16

We started to see the `System.Core-xunit` step on CI to hit the timeout of 15 minutes with Linux/ARM64. That was weird, because the step used to be completed in around two minutes. With my local device (jetson board) I wasn't able to reproduce it either; it took around 100s there. We then realized it's specific to the new `taishan` CI machines, which are equipped with 64 cores. Hardcoding `mono_cpu_count` to return 16 restored the performance, however that isn't a viable fix.

Limiting `DefaultDegreeOfParallelism` to 16 fixes it, which is less extreme than limiting `mono_cpu_count ()`, still not ideal though. It seems to boil down to the fact that our non-netcore threadpool implementation doesn't handle a large number of cores well.

`repro.cs`, extracted from here https://github.com/dotnet/corefx/blob/a9b91e205a8794327a028cb4b29953127f0f194c/src/System.Linq.Parallel/tests/QueryOperators/ConcatTests.cs#L145-L154
```csharp
using System;
using System.Linq;
using System.Collections.Generic;
using System.Threading;

public class Repro {
    public static void Main (string []args) {
        const int ElementCount = 2048;
        ParallelQuery<int> leftQuery = ParallelEnumerable.Range(0, ElementCount / 4).Union(ParallelEnumerable.Range(ElementCount / 4, ElementCount / 4));
        ParallelQuery<int> rightQuery = ParallelEnumerable.Range(2 * ElementCount / 4, ElementCount / 4).Union(ParallelEnumerable.Range(3 * ElementCount / 4, ElementCount / 4));

        var results = new HashSet<int>(leftQuery.Concat(rightQuery));
        Console.WriteLine ("results.Count=" + results.Count + ", ElementCount=" + ElementCount);
    }
}
```

Before fix:
```console
$ time ./mono/mini/mono-sgen repro.exe
results.Count=2048, ElementCount=2048

real    0m5.846s
user    0m0.344s
sys     0m1.929s
$ make -C mcs/class/System.Core run-xunit-test
[...]
=== TEST EXECUTION SUMMARY ===
   net_4_x_System.Core_xunit-test  Total: 48774, Errors: 0, Failed: 0, Skipped: 6, Time: 536.005s
```

With this fix:
```console
$ time ./mono/mini/mono-sgen repro.exe
results.Count=2048, ElementCount=2048

real    0m1.247s
user    0m0.206s
sys     0m0.225s
$ make -C mcs/class/System.Core run-xunit-test
[...]
=== TEST EXECUTION SUMMARY ===
   net_4_x_System.Core_xunit-test  Total: 48774, Errors: 0, Failed: 0, Skipped: 6, Time: 131.143s
```
---
 .../src/System/Linq/Parallel/Scheduling/Scheduling.cs        | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/src/System.Linq.Parallel/src/System/Linq/Parallel/Scheduling/Scheduling.cs b/src/System.Linq.Parallel/src/System/Linq/Parallel/Scheduling/Scheduling.cs
index 945994e06ebd..4c9a89c1da05 100644
--- a/src/System.Linq.Parallel/src/System/Linq/Parallel/Scheduling/Scheduling.cs
+++ b/src/System.Linq.Parallel/src/System/Linq/Parallel/Scheduling/Scheduling.cs
@@ -47,8 +47,13 @@ internal static class Scheduling
         // The number of milliseconds before we assume a producer has been zombied.
         internal const int ZOMBIED_PRODUCER_TIMEOUT = Timeout.Infinite;
 
+#if MONO
+        /* limit to degree of 16 to avoid too much contention */
+        internal const int MAX_SUPPORTED_DOP = 16;
+#else
         // The largest number of partitions that PLINQ supports.
         internal const int MAX_SUPPORTED_DOP = 512;
+#endif
 
 
         //-----------------------------------------------------------------------------------