Pool the dictionary buffer when training a Zstandard dictionary

[alinpahontu2912]

ZstandardDictionary.Train allocates new byte[maxDictionarySize] on every call to hold the trained dictionary returned by the native ZDICT_trainFromBuffer. Upstream zstd guidance puts a reasonable dictionary size at ~100 KB and the CLI defaults to 112,640 bytes — both above the .NET 85,000-byte LOH threshold. So for the recommended-and-up sizes, every training call costs a fresh Large Object Heap allocation that survives until the next gen2 collection. This change rents the buffer from ArrayPool<byte>.Shared and returns it (with clearArray: true) once the trained dictionary has been copied into its own array.

Why dictionary sizes are LOH-heavy

Upstream zstd (zdict.h):

"A reasonable dictionary size, the dictBufferCapacity, is about 100KB. The zstd CLI defaults to a 110KB dictionary."

CLI manpage (zstd.1.md): --maxdict=# default is 112640 bytes.

The managed wrapper repeats this guidance (ZstandardDictionary.cs:85-87) and only guards the lower bound (ThrowIfLessThan(maxDictionarySize, 256, ...) at line 128) — there is no upper-bound check, so MB-sized buffers are also legal.

.NET LOH threshold is 85,000 bytes, so the recommended ~100 KB dictionary and the 112,640-byte CLI default both land on the LOH on every call.

Empirical evidence

I modelled the exact allocation pattern (new byte[N] vs Rent/Return) and measured per-call allocations with GC.GetAllocatedBytesForCurrentThread (Release build, 1,000 iterations, workstation GC, .NET 11 preview):

Buffer size	new byte[] per call	Rent per call (steady state)	Reduction	Hits LOH?
4 KB	4,120 B	0 B	100 %	no
50 KB	50,024 B	0 B	100 %	no
100 KB (zstd recommended)	102,424 B	0 B	100 %	YES
110 KB (zstd CLI default)	112,664 B	0 B	100 %	YES
1 MB	1,048,600 B	0 B	100 %	YES

For the recommended/default dictionary sizes, every Train call previously produced a ~100 KB LOH allocation that only a gen2 collection could reclaim. After this change, steady-state Train allocations for the dictionary buffer drop to zero.

Reproducer

Program.cs (run from a scratch net11.0 console project, dotnet run -c Release):

// Licensed to the .NET Foundation under one or more agreements.
// The .NET Foundation licenses this file to you under the MIT license.
using System;
using System.Buffers;
using System.Runtime.CompilerServices;

int[] sizes = { 4_096, 50_000, 102_400, 112_640, 1_048_576 };
const int Iterations = 1_000;

foreach (int size in sizes)
{
    BeforePath(size);
    AfterPath(size);
    GC.Collect(); GC.WaitForPendingFinalizers(); GC.Collect();

    long b0 = GC.GetAllocatedBytesForCurrentThread();
    for (int i = 0; i < Iterations; i++) BeforePath(size);
    long b1 = GC.GetAllocatedBytesForCurrentThread();

    GC.Collect(); GC.WaitForPendingFinalizers(); GC.Collect();

    long a0 = GC.GetAllocatedBytesForCurrentThread();
    for (int i = 0; i < Iterations; i++) AfterPath(size);
    long a1 = GC.GetAllocatedBytesForCurrentThread();

    Console.WriteLine($"size={size,9:N0}  new/iter={(b1 - b0) / Iterations,9:N0} B  rent/iter={(a1 - a0) / Iterations,9:N0} B");
}

[MethodImpl(MethodImplOptions.NoInlining)]
static void BeforePath(int n)
{
    byte[] b = new byte[n];
    b[0] = 1;
    GC.KeepAlive(b);
}

[MethodImpl(MethodImplOptions.NoInlining)]
static void AfterPath(int n)
{
    byte[] b = ArrayPool<byte>.Shared.Rent(n);
    try
    {
        b[0] = 1;
        GC.KeepAlive(b);
    }
    finally
    {
        ArrayPool<byte>.Shared.Return(b, clearArray: true);
    }
}

Output (workstation GC, single thread):

size=    4,096  new/iter=    4,120 B  rent/iter=        0 B
size=   50,000  new/iter=   50,024 B  rent/iter=        0 B
size=  102,400  new/iter=  102,424 B  rent/iter=        0 B
size=  112,640  new/iter=  112,664 B  rent/iter=        0 B
size=1,048,576  new/iter=1,048,600 B  rent/iter=        0 B

The 24-byte overhead per allocation is the 64-bit byte[] object header (sync block + method table pointer + length). Rent reports 0 bytes/iter because the pool returns the same array each iteration after warm-up. The pre-PR path stays on the SOH for the first two rows and lands on the LOH for the last three.

[dotnet-policy-service]

Tagging subscribers to this area: @dotnet/area-system-collections
See info in area-owners.md if you want to be subscribed.

[Copilot]

Pull request overview

This PR reduces repeated large allocations in ZstandardDictionary.Train by renting the native training output buffer from ArrayPool<byte>.Shared instead of allocating a new byte[] every call, then returning the rented buffer after copying the trained dictionary into its own array.

Changes:

• Replace per-call new byte[maxDictionarySize] with ArrayPool<byte>.Shared.Rent(maxDictionarySize) for the native training output buffer.
• Ensure the rented dictionary buffer is returned in a finally block, currently using clearArray: true to avoid retaining caller-derived data in the pool.

[MihaZupan]

Suggested change

	// Clear before returning: the trained dictionary is derived from caller-supplied samples.
	ArrayPool<byte>.Shared.Return(dictionaryBuffer, clearArray: true);
	ArrayPool<byte>.Shared.Return(dictionaryBuffer);

We practically never clear buffers outside of crypto, I don't see a reason to do it here

[MihaZupan]

We already have a try/finally for the lengthsArray buffer, can we avoid even more nesting by reusing the existing blocks?