Offline rendering is much slower when ended handlers disconnect source graphs

[yoniaiz]

Summary

During OfflineAudioContext.startRendering(), an AudioBufferSourceNode.onended handler that disconnects the finished source graph can make rendering much slower than leaving the graph connected.

This may be related to #168, but the symptom here is render-time performance rather than retained memory.

Environment

• node-web-audio-api: 2.0.0
• Node.js: v24.13.1
• OS: macOS arm64 (Darwin 25.3.0)

Reproduction

import { AudioBufferSourceNode, GainNode, OfflineAudioContext } from 'node-web-audio-api';
import { performance } from 'node:perf_hooks';

const sampleRate = 44_100;
const durationSeconds = 60;
const eventCount = 2_000;

async function bench(mode) {
  const ctx = new OfflineAudioContext(1, sampleRate * durationSeconds, sampleRate);
  const buffer = ctx.createBuffer(1, Math.round(sampleRate * 0.01), sampleRate);
  buffer.getChannelData(0)[0] = 1;

  let endedCount = 0;
  const activeSources = new Set();

  for (let i = 0; i < eventCount; i += 1) {
    const source = new AudioBufferSourceNode(ctx, { buffer });
    const gain = new GainNode(ctx, { gain: 1 });
    source.connect(gain);
    gain.connect(ctx.destination);

    if (mode === 'onended-disconnect') {
      activeSources.add(source);
      source.onended = () => {
        endedCount += 1;
        activeSources.delete(source);
        source.disconnect();
        gain.disconnect();
      };
    } else if (mode === 'onended-noop') {
      source.onended = () => {
        endedCount += 1;
      };
    }

    source.start((i / eventCount) * durationSeconds);
  }

  const t0 = performance.now();
  await ctx.startRendering();
  const ms = performance.now() - t0;
  console.log(`${mode}: ${Math.round(ms)}ms ended=${endedCount} active=${activeSources.size}`);
}

await bench('none');
await bench('onended-noop');
await bench('onended-disconnect');

Result on my machine

none: 6439ms ended=0 active=0
onended-noop: 6630ms ended=2000 active=0
onended-disconnect: 54124ms ended=2000 active=0

A larger 120s / 4,000-source version rendered the no-cleanup and no-op-handler cases in about 38-39s, but the disconnecting onended case was still running after more than two minutes, so I aborted it.

Expected

I expected disconnecting finished source subgraphs from their ended handlers during an offline render to be roughly comparable to leaving them connected, or at least not ~8x slower for this case.

In an application render path with many short sample hits, avoiding offline onended cleanup changed the native offline render from roughly 90s to 32s for the same audio. The workaround is fine for one-shot offline render processes, but a fix here would allow spec-shaped cleanup/pruning strategies without that render-time cliff.

[orottier]

Thanks for the report, it is probably orottier/web-audio-api-rs#129
Disconnecting a node forces a new graph ordering

[b-ma]

Hey, thanks, this is quite instructive indeed.

@orottier, I wonder if we actually need to recompute the graph in case of disconnect, maybe removing the edge is enough, the order should not change?

@yoniaiz, apart the disconnect issue, I'm actually not sure you can really improve the perf here, as even if the node is disconnected I' not sure it will garbage collected in such a setup

[orottier]

I have just landed a performance improvements on graph reorderings via orottier/web-audio-api-rs#629
It has not been released yet but you could try to run the given example on that patched dependency.

I wonder if we actually need to recompute the graph in case of disconnect, maybe removing the edge is enough, the order should not change?

Definitely. I added this to orottier/web-audio-api-rs#129 just now