Possible external-memory / GC regression from v22 to... | nodejs/node

还没有人认领这个 Issue。

I’m seeing a pretty consistent regression from Node.js v22.22.3 to v24.16.0 on a small stream benchmark that creates lots of short-lived Readables and drains them immediately. The code in the benchmark file benchmark/streams/readable-push-buffer-burst-minimal.js is not optimal, but that's not the point as it demonstrates the issue.

'use strict';

const common = require('../common.js');
const assert = require('assert');
const { Readable } = require('stream');

const bench = common.createBenchmark(main, {
  n: [2e4],
  bodySize: [120 * 1024],
});

function repeat(size, ch) {
  return ch.repeat(size);
}

function makeBurstChunks(bodySize) {
  return [
    '<!doctype html><html><head><title>burst</title><style>',
    repeat(1024, 'a'),
    '</style></head><body><header>',
    repeat(2048, 'b'),
    '</header><main>',
    repeat(bodySize, 'c'),
    '</main><script>',
    repeat(2048, 'd'),
    '</script></body></html>',
  ];
}

function toBuffers(chunks) {
  return chunks.map((chunk) => Buffer.from(chunk, 'utf8'));
}

function buildAndDrain(buffersIn) {
  const buffersOut = [];
  for (let i = 0; i < buffersIn.length; i++) {
    const readable = new Readable({ read() {} });
    readable.push(buffersIn[i]);
    readable.push(null);
    let chunk;
    while ((chunk = readable.read()) !== null)
      buffersOut.push(chunk);
  }
  return Buffer.concat(buffersOut);
}

function main({ n, bodySize }) {
  const inputBuffers = toBuffers(makeBurstChunks(bodySize));
  const expectedLength = inputBuffers.reduce(
    (sum, chunk) => sum + chunk.length,
    0,
  );

  const warmup = buildAndDrain(inputBuffers);
  assert.strictEqual(warmup.length, expectedLength);

  bench.start();
  for (let i = 0; i < n; ++i) {
    const buffer = buildAndDrain(inputBuffers);
    if (buffer.length !== expectedLength)
      throw new Error(`Unexpected buffer length: ${buffer.length}`);
  }
  bench.end(n);
}

So running the benchmark comparing v22.22.3 vs v24.16.0, I see the following results:

$ node benchmark/compare.js --old ~/.nvm/versions/node/v22.22.3/bin/node --new ~/.nvm/versions/node/v24.16.0/bin/node --filter readable-push-buffer-burst-minimal streams > readable-push-buffer-burst-minimal.csv
[00:00:58|% 100| 1/1 files | 60/60 runs | 1/1 configs]: Done
$ npx node-benchmark-compare readable-push-buffer-burst-minimal.csv
                                                                      confidence improvement accuracy (*)   (**)  (***)
streams/readable-push-buffer-burst-minimal.js bodySize=122880 n=20000        ***    -37.18 %       ±4.12% ±5.55% ±7.36%

I also profiled both cases with perf and generated flamegraphs. The main difference seems to be GC behavior: v22 is mostly dominated by scavenges, while v24 is much more major-GC-heavy. See attached flamegraph-22.svg and flamegraph-24.svg.

One more datapoint that seems relevant: if I run v24 with --external-memory-accounted-in-global-limit, the regression drops a lot. Here are the results using a patched benchmark/compare.js that supports per-binary flags:

$ node benchmark/compare.js --old ~/.nvm/versions/node/v22.22.3/bin/node --new ~/.nvm/versions/node/v24.16.0/bin/node --new-flags "--external-memory-accounted-in-global-limit" --filter readable-push-buffer-burst-minimal streams > readable-push-buffer-burst-minimal_with_flag.csv
[00:00:50|% 100| 1/1 files | 60/60 runs | 1/1 configs]: Done
$ npx node-benchmark-compare readable-push-buffer-burst-minimal_with_flag.csv
                                                                      confidence improvement accuracy (*)   (**)  (***)
streams/readable-push-buffer-burst-minimal.js bodySize=122880 n=20000        ***    -11.18 %       ±4.28% ±5.76% ±7.62%

Be aware that when doing many comparisons the risk of a false-positive result increases.
In this case, there are 1 comparisons, you can thus expect the following amount of false-positive results:
  0.05 false positives, when considering a   5% risk acceptance (*, **, ***),
  0.01 false positives, when considering a   1% risk acceptance (**, ***),
  0.00 false positives, when considering a 0.1% risk acceptance (***)

So in summary, when running the benchmark, I see a 37% regression which drops to only 11% when using v24 with the --external-memory-accounted-in-global-limit flag.

Does this ring a bell as a known regression or is this to be expected in v24?

Flamegraphs

flamegraph-22.svg (v22.22.3)
flamegraph-24.svg (v24.16.0)
flamegraph-24-flag.svg (v24.16.0 --external-memory-accounted-in-global-limit)

/cc @RafaelGSS

首先，使用 benchmark/compare.js 针对两个 Node.js 版本运行 benchmark/streams/readable-push-buffer-burst-minimal.js，然后比较附带的火焰图。调查 GC 行为的变化以及 --external-memory-accounted-in-global-limit 的影响。完成的标准是确定 v22 到 v24 的性能差异是预期的变化还是回归，并提供来自可复现 benchmark 结果的证据。

由索引模型根据 Issue 内容生成。

Possible external-memory / GC regression from v22 to v24 in short-lived Readable chunk workloads

描述

Flamegraphs

贡献指南

从这里开始

调研方向

评估

Possible external-memory / GC regression from v22 to v24 in short-lived Readable chunk workloads

描述

Flamegraphs

贡献指南

从这里开始

调研方向

评估

把新 issue 发到你的邮箱