node: improve GetActiveRequests performance

[trevnorris]

v8 is faster at setting object properties in JS than C++. Even when it
requires calling into JS from native code. Make
process._getActiveRequests() faster by doing this when populating the
array containing request objects.

Simple benchmark:

for (let i = 0; i < 22; i++)
fs.open(__filename, 'r', function() { });
let t = process.hrtime();
for (let i = 0; i < 1e6; i++)
process._getActiveRequests();
t = process.hrtime(t);
console.log((t[0] * 1e9 + t[1]) / 1e6);

Results between the two:

Previous: 4406 ns/op
Patched: 829 ns/op 4.3x faster

R=@bnoordhuis

Have another addition of improving the same for active handles, but wanted to solicit feedback on the approach early.

[rvagg]

Is "index" the right word to use here? you're just appending and not actually using an index

[ronkorving]

While performance improvements are nice, this definitely adds more code than it removes. Is the use case here continuous monitoring? This is still an undocumented API... for what reason I don't know though, seems it would help to make it public (I've used it too, and find it very useful, but that's usually been limited to shutdown-time).

[trevnorris]

@rvagg That's an artifact of an early implementation. I'll find a better name.

@ronkorving The use case is to not have a crappy implementation. This is a technique that I plan to expand through core. Search for Object::Set() and you'll see how heavily we rely on this slow operation. Here was simply the easiest location to kick off.

LOC should have little to no affect on a PR. Added complexity I can understand, and can agree with based on circumstance. This though is fairly straightforward.

[ronkorving]

Is there any way we could help the V8 team (I say we.. as if I really could) to make Object::Set as fast as JIT compiled JavaScript?

[bnoordhuis]

Can you use ARRAY_SIZE(argv) instead of hard-coding it in several places? If it gets unwieldy, I suggest writing it as:

static const size_t argc = 5;
Local<Value> argv[argc];
// ...
argv[i++ % argc] = ...;

[bnoordhuis]

This could be extended to numerous other places but I assume that's your plan anyway. :-)

[trevnorris]

@bnoordhuis Comments addressed (I think). Yes, I am planning on extending this across core but figured this would be a good low impact place to start off. :)

[bnoordhuis]

Shouldn't this check that i > 0? It's going to make a superfluous JS call now if I read it right.

EDIT: Never mind, didn't read it right. It's never zero.

[trevnorris]

this is the most succinct way I found, but even looking back at this over the weekend I needed to remember what the logic was doing. if you have something more readable in mind I'm open to suggestions. :)

[trevnorris]

@bnoordhuis made most suggested changes and added a simple test.

CI: https://ci.nodejs.org/job/node-test-pull-request/536/

[bnoordhuis]

This is the repetition I mean. I'd do the i % argc only once and cache the result in a const size_t remainder.

[trevnorris]

ah, got it. was trying to make it work w/ the for loop above.

[trevnorris]

@bnoordhuis comment addressed.

[jasnell]

Nice change, although it makes me sad that Object::Set is so slow. LGTM so long as CI is green and setPropByIndex is refactored to get away from the val0, ... val7 pattern.

[trevnorris]

@bnoordhuis@jasnellarguments change made.

CI: https://ci.nodejs.org/job/node-test-pull-request/557/

[jasnell]

LGTM if CI is green

[bnoordhuis]

LGTM

[trevnorris]

None of the failures are related to the PR. Landed on 494227b. Thanks much!

[MylesBorins]

@trevnorris what are your thought regarding adding this commit to LTS?

[MylesBorins]

@trevnorris ping

[trevnorris]

@thealphanerd oop. sorry. It's a micro-performance optimization. Merging will doubtfully prevent future conflicts. So, can land but don't think it's necessary.