commit e7c005de60f2d329656d852b863ea0a1302a600d
author Nicholas S. Husin <nsh@golang.org> Thu Oct 09 02:10:40 2025 -0400
committer Nicholas Husin <nsh@golang.org> Thu Oct 09 10:55:28 2025 -0700
tree 00f946d74182952194b6159c46c689e1bd8e5b70
parent b93acc2126d5ad1a8836390a0fd801eb7dceb5b6

http2: implement a more efficient writeQueue that avoids unnecessary copies.

Our previous implementation of writeQueue relies on one
[]FrameWriteRequests, forcing us to copy the rest of the slice's content
whenever we remove an item from the front.

This change remedies this problem by implementing writeQueue using
two-stage queues, similar to Okasaki's purely functional queue.

With 25 frames per stream, we are observing the following performance
improvement:
goos: linux
goarch: amd64
pkg: golang.org/x/net/http2
cpu: AMD EPYC 7B13
              │  /tmp/old   │              /tmp/new               │
              │   sec/op    │   sec/op     vs base                │
WriteQueue-64   508.3n ± 3%   305.7n ± 3%  -39.86% (p=0.000 n=10)

              │  /tmp/old  │            /tmp/new            │
              │    B/op    │    B/op     vs base            │
WriteQueue-64   0.000 ± 0%   0.000 ± 0%  ~ (p=1.000 n=10) ¹
¹ all samples are equal

              │  /tmp/old  │            /tmp/new            │
              │ allocs/op  │ allocs/op   vs base            │
WriteQueue-64   0.000 ± 0%   0.000 ± 0%  ~ (p=1.000 n=10) ¹
¹ all samples are equal

As the number of frames increases, the performance difference becomes
more stark as the old implementation does a quadratic amount of copying
in total to be able to fully consume a queue.

Change-Id: Ide816ebdd89a41275b5829683c0f10d48321af50
Reviewed-on: https://go-review.googlesource.com/c/net/+/710635
Reviewed-by: Damien Neil <dneil@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Nicholas Husin <husin@google.com>