Performance regression in Node.js v14

[mcollina]

See #46 (comment)

[mcollina]

@puzpuzpuz I think it's a regression in core.

v12:

benchCore*1000: 5758.926ms

v14:

benchCore*1000: 6.924s

[mcollina]

@ronag do you know what could have slown down https://github.com/mcollina/sonic-boom/blob/63bc3a4024c17302c7b0d4d3d0ca2ee0f2e772a9/bench.js#L55-L60 so much in v14?

[puzpuzpuz]

Yes, it looks a lot like a core-side regression. I'm going to profile the benchmark tomorrow and try to find the bottleneck in v14.

[ronag]

How does it look if you try with node 15?

[mcollina]

same

[ronag]

Yea, we would need to do a bisect.

[puzpuzpuz]

@mcollina @ronag
A quick bisect over versions showed that the regression was introduced in v14.5.0. So, the regression must be somewhere here: nodejs/node#34093

[mcollina]

There are so many suspects there, it'll take a while :/.

[puzpuzpuz]

I've tried profiling the benchmark on 14.4.0 and 14.5.0 and noticed one strange thing. In 14.5.0 flatstr takes 32.1% (1656 of 5157 samples), while in 14.4.0 it's only 4.9% (179 of 3665 samples). So, maybe it's a regression on V8 side?

[ronag]

I can't really see anything suspicious in the change log for 14.5. At least not related to streams or fs.

[mcollina]

I've tried profiling the benchmark on 14.4.0 and 14.5.0 and noticed one strange thing. In 14.5.0 flatstr takes 32.1% (1656 of 5157 samples), while in 14.4.0 it's only 4.9% (179 of 3665 samples). So, maybe it's a regression on V8 side?

removing flatstr does not change the bench

[puzpuzpuz]

removing flatstr does not change the bench

Hmm, then it's something else.

[mcollina]

I'm currently bisect, let's see where I land.