Fs benchmarks take 10+ hours

[anonrig]

Current node:fs benchmarks take 10+ hours on our CI, making it impossible to run all of them and have a healthy emotional state. It also blocks other benchmark jobs.

I recommend aiming for 2-3 hours and changing the benchmarks according to fit into this deadline.

cc @nodejs/performance

[joyeecheung]

I think the fs benchmarks are having the same kind of issue that happen to many other benchmarks - they have a bit too many combinations of the parameters, and some of them are unnecessary. For example there are usually too many size / len variations. For example in this.

node/benchmark/fs/readfile-promises.js

Lines 17 to 24 in cdcb01a

	len: [
	1024,
	512 * 1024,
	4 * 1024 ** 2,
	8 * 1024 ** 2,
	16 * 1024 ** 2,
	32 * 1024 ** 2,
	],

We probably just need 2-3 combinations by default. Or at least only pick 2-3 combinations when we are running the entire fs benchmark. There could be something OS-specific about them to make certain options not representative enough to be in that default set but I think we can always expand the variations when there's actually any doubt.

[mscdex]

I think the fs benchmarks are having the same kind of issue that happen to many other benchmarks - they have a bit too many combinations of the parameters, and some of them are unnecessary.

This would probably be solved by #39285

[RafaelGSS]

Also related nodejs/build#3483.

[tniessen]

We could also make benchmark iterations scalable. Running a few benchmarks for 1e5 iterations is usually not a problem, but running all benchmarks for 1e5 iterations might be.

[Uzlopak]

I usually start with 1e7 iterations, and if it takes more time than I can bear, I reduce it by a factor of 10 and check if the benched result is not significantly different. Rinse and repeat.

[H4ad]

Is possible to create some kind of repo/issue to track slow benchmarks? Like we have for flaky tests?