benchmark: add benchmarks for Buffer.from()

[addaleax]

Checklist

• make -j8 test (UNIX), or vcbuild test nosign (Windows) passes
• tests and/or benchmarks are included
• commit message follows commit guidelines

Affected core subsystem(s)

benchmarks

Description of change

Adds benchmarks for Buffer.from() and its various argument combinations.

Ref: #8733

It also looks like there’s a performance regression on the current master that isn’t on v6 and which exceeds the one described in #8733 notably:

$ nvm use 6.6
Now using node v6.6.0
$ node benchmark/buffers/buffer-from.js
buffers/buffer-from.js n=1024 len=10 source="array": 1352.177972206885
buffers/buffer-from.js n=1024 len=2048 source="array": 108.78659817646022
buffers/buffer-from.js n=1024 len=10 source="arraybuffer": 2849.9419428697133
buffers/buffer-from.js n=1024 len=2048 source="arraybuffer": 3303.4248644402187
buffers/buffer-from.js n=1024 len=10 source="arraybuffer-middle": 3115.8790362447385
buffers/buffer-from.js n=1024 len=2048 source="arraybuffer-middle": 3143.4771954881257
buffers/buffer-from.js n=1024 len=10 source="buffer": 1362.2701961428957
buffers/buffer-from.js n=1024 len=2048 source="buffer": 367.766503317392
buffers/buffer-from.js n=1024 len=10 source="uint8array": 1480.231221483884
buffers/buffer-from.js n=1024 len=2048 source="uint8array": 139.1829921904438
buffers/buffer-from.js n=1024 len=10 source="string": 685.3303955985581
buffers/buffer-from.js n=1024 len=2048 source="string": 163.58717416397928
buffers/buffer-from.js n=1024 len=10 source="string-base64": 504.0783343717544
buffers/buffer-from.js n=1024 len=2048 source="string-base64": 110.29233159776292
$ ./node benchmark/buffers/buffer-from.js 
buffers/buffer-from.js n=1024 len=10 source="array": 522.2764604853
buffers/buffer-from.js n=1024 len=2048 source="array": 97.72530491842134
buffers/buffer-from.js n=1024 len=10 source="arraybuffer": 776.7268797100915
buffers/buffer-from.js n=1024 len=2048 source="arraybuffer": 658.5854603318716
buffers/buffer-from.js n=1024 len=10 source="arraybuffer-middle": 690.3042490225982
buffers/buffer-from.js n=1024 len=2048 source="arraybuffer-middle": 742.2573140765411
buffers/buffer-from.js n=1024 len=10 source="buffer": 496.9498820166972
buffers/buffer-from.js n=1024 len=2048 source="buffer": 296.6712882896401
buffers/buffer-from.js n=1024 len=10 source="uint8array": 561.3881977090617
buffers/buffer-from.js n=1024 len=2048 source="uint8array": 124.81186442783465
buffers/buffer-from.js n=1024 len=10 source="string": 386.8836136075654
buffers/buffer-from.js n=1024 len=2048 source="string": 173.50450305089902
buffers/buffer-from.js n=1024 len=10 source="string-base64": 370.7046939007373
buffers/buffer-from.js n=1024 len=2048 source="string-base64": 120.02970526090957

:(

[Fishrock123]

Ouch! Even if these "micro benchmarks" are not 100% accurate, they could still be good for discovering large deltas.

[lpinca]

LGTM

[addaleax]

Ouch! Even if these "micro benchmarks" are not 100% accurate, they could still be good for discovering large deltas.

Yeah. The difference actually seems pretty reproducible to me, so I’m going to starting bisecting now…

[addaleax]

The first bad commit could be any of:
785506a1fc460323074dbeffb27b4518c404e9b4
fbfc15c51b7cb8de96d6719d3b2096c086530ee3
7292a1e954e0db348ff78c704df9e105ba5667ad
ec02b811a8a5c999bab4de312be2d732b7d9d50b
2187bd3b4a3430880ec1c2b99e4eff57f27afd89

Those are, unfortunately, the V8 5.4 update from #8317. /cc @nodejs/v8

(For completeness: benchmark results at all bisect points; bisected from the v6.0.0 release up to the current HEAD.)

[jasnell]

@addaleax ... ouch. That's a big drop.

For the bench mark itself, can you perhaps move each case into a separate function then use common.v8ForceOptimization() to force V8 to optimize before doing the actual benchmark. I'm curious to see if that would have any tangible impact. (I doubt it, but worth checking)

[addaleax]

@jasnell I can try that later, but since you’re obviously having something in specific in mind and I’ll be afk for a bit, feel free to use the Allow edits from maintainers. thingy to make any changes you’d like to see. :)

[jasnell]

heh... yeah, considered that but I need to be away from the keyboard for a bit also (my two boys have a cross country track meet at school and they need a cheering section). When I get back on later today I'll see what I can work up.

[targos]

I found the cause of the regression:

In the new FastBuffer(allocPool, poolOffset, size); call in allocate, V8 creates and uses an array iterator (for a yet unknown reason).

[addaleax]

I'm curious to see if that would have any tangible impact.

@jasnell it doesn’t seem to have much of an effect. Would you still prefer that change?

[jasnell]

No that's OK. Given what appears to be the cause of the regression, it's
not going to have much impact. Thank you for checking tho!

On Friday, September 23, 2016, Anna Henningsen notifications@github.com
wrote:

I'm curious to see if that would have any tangible impact.

@jasnell https://github.com/jasnell it doesn’t seem to have much of an
effect. Would you still prefer that change?

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
#8738 (comment), or mute
the thread
https://github.com/notifications/unsubscribe-auth/AAa2eXoEhGmffn7XJ5o8EY-AU6b9Q-ZTks5qtEC-gaJpZM4KFAv2
.

[imyller]

LGTM

[addaleax]

Landed in 289d862

[fhinkel]

We improved spread-super performance (my microbenchmark is as fast in 5.8 as it was in 5.1 when it wasn't spec compliant). Can we revisit this, once we have 5.8 in Node?

[addaleax]

Can we revisit this, once we have 5.8 in Node?

What exactly? The change in #8754? I would assume we could just revert that if we want and it doesn’t have an performance impact.

[fhinkel]

All I meant was to re-run the benchmarks. But we'll do that anyways. Sorry for the confusion. I got to excited by the performance improvements.