Node 24 Buffer GC patch

Verified on Linux x64. Independent preview, not an official Node.js release.

A reproducible Node 24 source patch and build/test harness for the short-lived Buffer garbage-collection regression tracked in nodejs/node#61967 and nodejs/node#63863.

The candidate restores the earlier 8 MiB scavenger-capacity default on builds without pointer compression and V8 older than 14. Explicit command-line flags still override it. It is a scoped workaround, not a backport of V8's newer external-memory-accounting implementation.

Verified results

• Both unpatched and patched runtimes compiled from the same source revision.
• 4,751 parallel/sequential test cases, zero failures (includes upstream skips for unsupported platform/features).
• Compiled-default, explicit-override, retained-buffer and worker-transfer checks passed.
• The Readable/Buffer reproducer measured 2.23× median throughput in a 15-sample-per-binary rerun; the first runner measured 2.11× across seven samples.

See the green verification run, full validation record, and raw evidence. Performance gains are specific to the measured workloads and machines.

Why this helps

V8 13.6 raised the capacity default from 8 to 32 MiB. Without pointer compression, this pushes the young external-backing-store collection threshold above the external-memory soft limit. Short-lived Buffers can then trigger major collections before a young collection reclaims them. Restoring the older default addresses that ordering, but can also affect workloads that benefit from a larger young heap.

Reproduce

The workflow pins Node source commit 0d10bfe3cb961ef997490f5c060bd1466b5affea (v24.x). It builds an unpatched binary, applies the patch, then rebuilds the changed code. It runs a compiled-default/override regression test, the upstream parallel and sequential suites, and interleaved benchmark samples against the unpatched binary. Logs, raw benchmark samples, runtime metadata, and checksums are retained as Actions artifacts. Timings are observations, not a promise of speedup for other applications.

Linux prerequisites: Git, GCC/G++ 12.2 or newer, Python 3, Make, and the other Node build prerequisites.

bash scripts/build.sh
python3 scripts/verify.py node-source results
python3 scripts/compare.py node-source results --runs 7
results/node-patched scripts/stress.cjs
cd node-source
mkdir -p /tmp/n24-tests
NODE_TEST_DIR=/tmp/n24-tests python3 tools/test.py --mode=release -j 4 parallel sequential

To try the established workaround without building a custom runtime:

node --scavenger-max-new-space-capacity-mb=8 app.js

Benchmark your own workload first. Do not use this flag on runtimes that no longer expose it. This repository does not change your installed Node or your application.

The custom build changes the default for the entire process. Applications that benefit from larger young heaps may behave differently. Explicit --scavenger-max-new-space-capacity-mb=32 restores the baseline setting. Windows measurements here use the official runtime with a flag; they do not validate a custom Windows binary. macOS, ARM64 and Node's full multi-platform CI are outside the validation performed here.

Windows workload comparison, using your own Node 24 executable:

./scripts/windows-check.ps1 -NodePath C:\path\to\node.exe

To repeat tests on already compiled Linux artifacts, run the Recheck compiled artifacts workflow with a build run ID. It verifies binary checksums before executing tests. This avoids another full compilation while retaining the exact same executables.

Attribution and scope

Upstream issue reporters and contributors identified this regression and its mechanism before this repository existed. The stream benchmark is attributed to santigimeno's reproducer in #63863; eliau2005 documented the bisection and capacity workaround in the upstream discussion. Node-derived code retains the upstream license in LICENSE.node. Harness additions use LICENSE. Development assisted by OpenAI Codex.

This is an independent patch repository. It is not affiliated with or endorsed by Node.js, and no upstream acceptance or release is claimed. An upstream submission requires human review and the Node project's contribution/DCO process.