Port ConditionalWeakTable from .NET Core 3.x to .NET Standard 2.0 about lucenenet

[NightOwl888]

Since WeakDictionary is a port of WeakHashSet in Java , has anyone taken a long hard look at the WeakDictionary code to see why it's not more performant and why some of it's tests fail in debug mode?

Yes, Vincent Van Den Burghe looked at this a while back and this was a "best effort" that could be done using pure managed code. It uses WeakReference to allow references to be GC'd, however, the dictionary entries still need to be cleaned up after the WeakReference dies and therein lies the bottleneck. Most read/write operations do a CleanIfNeeded() operation which isn't very efficient.

Using native resources is really the only way we can be sure that dictionary entries are "removed" as soon as their contained reference goes out of scope. ConditionalWeakTable does exactly what we need, but in .NET Standard 2.0 it is missing 2 of the APIs we need, which is a deal breaker in those 5 cases mentioned above.

Since no version of .NET Framework yet supports .NET Standard 2.1

Microsoft considers .NET Framework to be "finished" and there will be no further development on it, including upgrading to .NET Standard 2.1. This is a problem that won't go away until everyone moves from .NET Framework to .NET 5+, which is now the best upgrade path.

This problem dies with .NET Framework (meaning it will become less of an issue over time) and we have a solution that works for now, but is less than ideal. Additionally, WeakDictionary has been marked internal. Therefore, this issue is fairly low priority.

That being said, there are other components of J2N that could benefit performance-wise by using native code as well. It would be great if we set up a precedent for building and deploying native code with J2N.

[clambertus]

Is related to: GH-610

[shadowcaptain]

The source code for the dotnetcore implementation is here https://github.com/dotnet/runtime/blob/master/src/libraries/System.Private.CoreLib/src/System/Runtime/CompilerServices/ConditionalWeakTable.cs

[NightOwl888]

@shadowcaptain

Thanks. An attempt has already been made on this. The blocker is the fact that it depends on DependentHandle which has native resources.

DependentHandle calls native methods using [MethodImpl(MethodImplOptions.InternalCall)].

        [MethodImpl(MethodImplOptions.InternalCall)]
        private static extern IntPtr nInitialize(object primary, object? secondary);

        [MethodImpl(MethodImplOptions.InternalCall)]
        private static extern object? nGetPrimary(IntPtr dependentHandle);

        [MethodImpl(MethodImplOptions.InternalCall)]
        private static extern object? nGetPrimaryAndSecondary(IntPtr dependentHandle, out object? secondary);

        [MethodImpl(MethodImplOptions.InternalCall)]
        private static extern void nSetPrimary(IntPtr dependentHandle, object? primary);

        [MethodImpl(MethodImplOptions.InternalCall)]
        private static extern void nSetSecondary(IntPtr dependentHandle, object? secondary);


        [MethodImpl(MethodImplOptions.InternalCall)]
        private static extern void nFree(IntPtr dependentHandle);

It is not currently known how to translate that into a call that can be done from a 3rd party library such as J2N. Some important questions to answer:

• Can we assume the native method exists for the current platform, since .NET Standard should "just work"?
• Do we need to bundle the native resources in J2N to call them?
• If we need to bundle native resources, do they work on all platforms or do we need separate builds for specific platforms?

Note that [MethodImpl(MethodImplOptions.InternalCall)] doesn't exist on .NET Standard 1.x, but it is probably time to drop support for that target, anyway.

[rclabo]

Regarding DependentHandle :

Can we assume the native method exists for the current platform

This is a head scratcher. On one hand the static extern is likely referencing some native binary implementation. This would suggest the binary implementation for each platform would have to be bundled with J2N so that they could be available and that doesn't feel like the right direction. On the other hand this call is being made from .Net Core 3.0 code which would suggest that the appropriate binaries must already be in place on all platforms that .Net Core 3.0 supports. So in the end, my money is on those extern methods being available everywhere .Net Core 3.0 supports. That of course does not ensure .NET Standard 2.0 support by any means. Which brings us back to having to embed OS specific binaries which sounds like Yuk!

So I did some poking around and it appears there is an identical mono class that does not make extern calls. Incorporating this as the J2N implementation may be the way to go. https://github.com/dotnet/runtime/blob/master/src/mono/netcore/System.Private.CoreLib/src/System/Runtime/CompilerServices/DependentHandle.cs

[NightOwl888]

So I did some poking around and it appears there is an identical mono class that does not make extern calls. Incorporating this as the J2N implementation may be the way to go. https://github.com/dotnet/runtime/blob/master/src/mono/netcore/System.Private.CoreLib/src/System/Runtime/CompilerServices/DependentHandle.cs

Good find. However, upon closer inspection, GC.register_ephemeron_array(data); is also a native call and so is the source of GC.EPHEMERON_TOMBSTONE.

So, either way it appears we need to go native to solve this. However, it looks like using the approach of Mono is simpler.

[rclabo]

Yep it looks like you are correct.

Another options, and maybe one I like even better is to use compiler directives to use the existing WeakDictionary for NET Standard 2.0 and to use ConditionalWeakTable for NET Standard 2.1 or later.

This avoids the whole "port ConditionalWeakTable to NET Standard 2.0" all together. This way the code works on NET Standard 2.0 but is even more performant in highly concurrent environments when running in NET Standard 2.1 or later.

[NightOwl888]

We already are using conditional compilation to minimize usage of WeakDictionary, this issue is mainly about

• Improving performance on .NET Framework and other platforms that don't support .NET Standard 2.1
• Factoring WeakDictionary (and its tests) out of Lucene.NET altogether

Do note that WeakDictionary is basically a straight port from WeakHashSet in Java and some of its tests fail when running in Debug mode in addition to having performance issues with cleanup. AFAIK this does not cause issues with Lucene.NET, but it is not very encouraging.

We could potentially add an option in .NET Standard 2.0 for CachedOrdinalsReader to use ConditionalWeakTable by disabling the functionality of RamBytesUsed() when the user doesn't require it and eliminate one usage of WeakDictionary.

However, that would still leave WeakDictionary in

• FieldCacheImpl
• AssertingScorer
• CachingWrapperFilter
• ThreadedIndexingAndSearchingTestCase

Since the main usage (and main bottleneck) of WeakDictionary on .NET Standard 2.0 is in FieldCacheImpl, that wouldn't gain us very much.

[rclabo]

Ok, that's all great information. Thank you!

Since no version of .NET Framework yet supports .NET Standard 2.1 and since .NET Framework 4.6.1 does supports .NET Standard 2.0, if the performance hit of using WeakDictionary is large then I can totally see where you are coming from.

Since WeakDictionary is a port of WeakHashSet in Java , has anyone taken a long hard look at the WeakDictionary code to see why it's not more performant and why some of it's tests fail in debug mode?

[NightOwl888]

Moving this to Future since this is not blocking the 4.8.0 release.