Reduce memory usage of allocation profiles

[szegedi]

When the profiler hands off the collected V8 profiles to the rest of the processing code written in TypeScript, it transforms the V8 profile tree into a tree of JS objects. They all have the same string-valued keys. For time (wall and CPU) profiles, we have a mechanism to only create the key string values once per serialization and use them in every node object. For allocation profiles, we didn’t have this logic before, this PR implements it.

That’s 63 characters/node object less for keys name, scriptName, scriptId, lineNumber, columnNumber, children, and allocations and 14 characters/allocation object less for keys sizeBytes and count, plus extra overhead for the string value itself.

This tree only exists for a short period of time (once every 65 seconds between V8 capturing the allocation profile to us transforming it into pprof format) but its footprint still contributes to the peak memory use of the profiler.

Customers are sometimes running Node.js in constrained container environments, so reducing the peak memory use of the profiler is a worthwhile goal.

This currently handles periodic and on-shutdown profiles, but does not handle the special case (with a separate path) of allocation profile capture at OOM using kInterruptCallback mode but it would be easy to extend it for that as well.

JIRA: PROF-11123

[github-actions]

Overall package size

Self size: 9.52 MB
Deduped: 9.9 MB
No deduping: 9.9 MB

Dependency sizes

| name | version | self size | total size | |------|---------|-----------|------------| | source-map | 0.7.4 | 226 kB | 226 kB | | pprof-format | 2.1.0 | 111.69 kB | 111.69 kB | | p-limit | 3.1.0 | 7.75 kB | 13.78 kB | | delay | 5.0.0 | 11.17 kB | 11.17 kB | | node-gyp-build | 3.9.0 | 8.81 kB | 8.81 kB |

_{🤖 This report was automatically generated by heaviest-objects-in-the-universe}

[pr-commenter]

Benchmarks

Benchmark execution time: 2025-01-15 14:41:46

Comparing candidate commit 6ce12d0 in PR branch szegedi/cheaper-heap-profile with baseline commit 784f6f6 in branch main.

Found 1 performance improvements and 1 performance regressions! Performance is the same for 94 metrics, 24 unstable metrics.

scenario:profiler-idle-with-wall-profiler-18

• 🟩 cpu_user_time [-7.795ms; -2.211ms] or [-10.739%; -3.046%]

scenario:profiler-light-load-with-wall-profiler-22

• 🟥 cpu_user_time [+4.384ms; +10.687ms] or [+4.170%; +10.164%]

[nsavoire]

Methods defined inside class are already implicitly inlined:

Suggested change

	inlinev8::Local<v8::Object> NewObject() { returnv8::Object::New(isolate); }
	v8::Local<v8::Object> NewObject() { returnv8::Object::New(isolate); }

[nsavoire]

Suggested change

	inlineexplicitProfileTranslator() {}
	ProfileTranslator() = default;