Promise performance drop between v8 7.4 and 7.5

[Aschen]

I was doing some benchmark about promises performances and I noticed a performance drop in traditional promises (then/catch) between Node.js 10 and Node.js 12.

With further investigation I'm now able to say that the performance drop occur between v8 7.4 (Node.js 12.0.0) and v8 7.5 (Node.js 12.5.0).

The performance drop is still present in recent version of v8 (tested with v8 8.1 on Node.js 14.2.0)

Promises are between 3x and 5x slower in v8 7.5 than in v8 7.4 and they are still 2x slower in v8 8.1

Do you have any idea why?

If we look at the bigger picture, we can see that promise performance were much more better at some point (Node.js 10) and then start to deteriorate

You can run the script present on this Gist to reproduce the benchmarks: https://gist.github.com/Aschen/1d35ccb6dc1dc3d6796b00cecfebb6d0

[mscdex]

Out of curiosity, did you try the master branch, which now has V8 8.3?

[targos]

The following download has V8 8.3 if you want to test it: https://nodejs.org/download/nightly/v15.0.0-nightly202005135bb4d01fbe/

[Aschen]

@mscdex I updated the benchmark and the script to use the version @targos linked and the results are worse with the latest version of Node.js / v8

[Aschen]

This issue is being addressed here https://bugs.chromium.org/p/v8/issues/detail?id=10550

[camillobruni]

After some investigation I found the root cause: V8 changed the default page size from 512 KiB (v7.4) to 256 KiB (v7.5)

Some background:

• performance.js measures with process.memoryUsage()
• Memory usage in V8 is calculated by iterating over all Pages
• The number of pages doubled in v12.5 since V8 v7.5 reduced the default page size

Take-away:

• bluebird should probably not time the memory measurements (presuming this is not something that is happening very often)
• Node should provide a fast API for just accessing RSS itself (rss is cheap, heap-stats are more expensive)

[camillobruni]

FYI, to underly my comment about measuring the right thing:
Commenting out the memMax = Math.max(memMax, process.memoryUsage().rss); line in performance.js makes the benchmark run 3x faster in node v12.5.

[Aschen]

Thanks you very much @camillobruni for all these explanations!

I submitted a PR in Bluebird to fix those benchmarks and another one in Node.js to add a direct access to the rss.

Also updated the original article at https://blog.kuzzle.io/bluebird-native-async_await-javascript-promises-performances-2020

[Leodau]

@ryzokuken

[benjamingr]

Hey, I totally missed this issue - cool investigation I'll take a look at the bluebird issue.

Note that:

• V8 has their own forked version of the bluebird benchmark (which I think @bmerurer made?), IMO it's slightly worse (because it doesn't measure the overhead of promisification) but it's probably more maintained.
  • These benchmarks haven't been updated on the bluebird side in ~5 years and were written for CrankShaft (the old JIT), it (and bluebird) make a lot of (now incorrect) assumptions regarding how the benchmarks are run and I would not want people to use them as an authoritative source. It's kind of shameful Doxbee is still the best we have for this in non-micro-benchmark land.
  • I'm also a Node member - and there are lots of promise performance and debugging tasks Node needs helps with :]
  • If you want to add a promise performance benchmark to Node core that we run to detect regressions that could be a good place to contribute - though not where I'd start.
• We've been trying to get people to use native promises (and improve those!) where possible.
• Promises are generally "fast enough" for most use cases afaict.

[mmarchini]

@benjamingr I believe you're referring to these benchmarks?