Deep dive discussion on {Async,Promise}_hooks & performance

[ofrobots]

During the WG meeting today we discussed that It would be good to meet and unblock the topic of async hooks / promise hooks performance. (#188).

Context:

• Proposal by @hashseed
• async_hooks: add currentResource node#21313

Here's a doodle: https://doodle.com/poll/fqkzvw929ueys2u9 I want to be able to get at least @hashseed @mcollina @AndreasMadsen to be able to attend. Please mark your availability by Friday Oct 19 and I will get a deep dive discussion scheduled.

[AndreasMadsen]

I would like @mike-kaufman or another from Microsoft to also attend, as they appear to be the primary driver of "JavaScript Async Context".

[Qard]

I'd like to attend if I can, so I put my times up. I have some thoughts on the subject. It's okay if it doesn't line up for me though. Definitely @hashseed, @mcollina and @AndreasMadsen are the core people we need most in the conversation.

[ofrobots]

It would be great to have as many people attend as possible. Please add your preferences, and I will try my best to find a good timeslot 🤞.

[mcollina]

Done! Il giorno gio 18 ott 2018 alle 23:34 Ali Ijaz Sheikh < [email protected]> ha scritto:

…

It would be great to have as many people attend as possible. Please add your preferences, and I will try my best to find a good timeslot 🤞. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#248 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AADL49un_DM_3_Gfk3Mg8gdIbrriEQx2ks5umPPJgaJpZM4Xru28> .

[Flarna]

These evening slots are in general hard for me and also difficult to predict. I don't add anything in doodle to avoid to bias it but I will try to attend.

[davisjam]

Will this conversation be recorded?

[mhdawson]

I'll try to attend as well, but think it is more important that times are good for other participants so won't complete the doodle.

[ofrobots]

Based on the poll, the best meeting time slot seems to be Oct 30, 11am PDT (1900 Central Europe) (your local time).

Hangouts meeting link: https://meet.google.com/tov-fcdh-jfw

[AndreasMadsen]

Is that 18:00 UTC? I'm switching to daylight saving time on the 28th, so I'm a bit worried and would like a fixed timezone specifier.

[wentout]

I'm very sorry, but I'm very interested in the result of this discussion.
Can I join and listen also?
I closed Issue #249, cause seems is was bit duplicate and everyone knows the idea, described there.

[ofrobots]

@AndreasMadsen AFAIK that should be Oct 30th 1800 UTC. Timezones are hard 😂.

@wentout you're welcome to join the hangouts at the time of the meeting!

[mmarchini]

Google Calendar should handle timezones and daylight saving time correctly, at least it does here when I look at the Node.js Foundation Calendar.

[wentout]

So, it made all necessary translations to my local time. Very easy and usable.

[ofrobots]

Just a reminder that this meeting will start in about 1-hr's time. It would be useful to review the prior discussions before the meeting (see OP).

At the time of the meeting, participants can join the meeting at: https://meet.google.com/tov-fcdh-jfw.

[mmarchini]

I'm on the road and won't make it :(

[mcollina]

Here is the link to the domain issue: nodejs/node#23862

[mcollina]

Here is the link to the currentAsyncResource() latest PR: https://github.com/mcollina/node/tree/current-resource-bis.

[ofrobots]

The recording of the meeting is here: https://youtu.be/2SCNsozMkF8. Thanks for everyone for participating.

Here's a brief summary of the action plan:

• We discussed that it would be great if majority of the async-hooks uses-cases didn't need to use the destroy hook. The hook can stay, but the expectation is that the availability of the resource on hook callbacks will remove the need for APMs to install destroy hooks. @mcollina's PR async_hooks: add currentResource node#21313 needs to be resurrected. Looking for volunteers. If there are no volunteers, @mcollina will pick this up again once he has some free time in December. There needs to be some consideration of whether it would problematic to expose the promise as the resource object. @hashseed to check with the V8 compiler team about gotchas. @ofrobots and @AndreasMadsen to provide concrete examples where we observe promises taking too long to GC.
• @hashseed to work with Maya on the V8 compiler team to introduce promise hooks from JS. Once these are added, Node core can switch to these.
• General agreement that non-observable promises don't need to trigger hooks as long as context preserved. V8 team to implement this optimization.
• V8 team expressed a desire for some spec modification describing promise hooks behavior. General agreement that this would be a good thing to get the interface between V8 and Node solidified. @hashseed to point to specific section and then WG can work drafting some spec test.

Let me know if I missed or misunderstood anything. Full details in the recording.

[mmarchini]

@mcollina's PR nodejs/node#21313 needs to be resurrected. Looking for volunteers. If there are no volunteers

I can do that. Should I just resubmit the same PR or are there any changes required before resubmitting it?

[mcollina]

@mmarchini there are some comments in the PR that needs to be fixed, mainly there is an ask that we maintain a stack of the current async resources.

[mmarchini]

Ok, thanks. I'll take a look into it next week.

[Flarna]

@mcollina Could you please point to the fix regarding http_parser which allows to use the async resource in before/after?

[mcollina]

@Flarna here it is nodejs/node#23201.

[AndreasMadsen]

@mcollina That PR looks like it is just a small fix. As far as I can see, the same resource object is still being reused. I would expect a fix that allows WeakMap to be used to involve something with the freelist and the HTTP_PARSER allocator.

[Flarna]

I agree with @AndreasMadsen. If a user stores some metadata into the HTTPParser object it will stay alive even after the parser is freed.
Maybe some dummy/wrapper object (similar as for Promises) should be created on every alloc and passed as resource to the callbacks. Doc states nothing really concrete users can rely on about the resource so it should be don't care.

[AndreasMadsen]

@Flarna to be clear the docs specifically says:

In some cases the resource object is reused for performance reasons, it is thus not safe to use it as a key in a WeakMap or add properties to it.

Once the issue with HTTPParser is fixed, I'm fine with changing that.

[mcollina]

@AndreasMadsen I cannot reproduce the problem anymore. Can you share a snippet that reproduces it?

'use strict'

const sleep = require('util').promisify(setTimeout)
const { createHook, executionAsyncResource } = require('async_hooks')
const asyncHook = createHook({init})
const sym = Symbol('cls')

asyncHook.enable()

function getCLS () {
  return executionAsyncResource()[sym]
}

function setCLS (data) {
  return executionAsyncResource()[sym] = data
}

function init (asyncId, type, triggerAsyncId, resource) {
  var cr = executionAsyncResource()
  if (cr) {
    if (type === 'HTTP_PARSER') {
      if (cr.noReuse) {
        throw new Error('no reuse')
      }

      cr.noReuse = true
    }

    resource[sym] = cr[sym]
  }
}


var counter = 0

const server = require('http').createServer(function (req, res) {
  setCLS(counter++)
  sleep(10).then(function () {
    res.setHeader('content-type', 'application/json')
    res.end(JSON.stringify({ cls: getCLS() }))
  })
})
server.listen(3000)

[AndreasMadsen]

@mcollina I'm not in a position to compile anything right now, but

'use strict'

const http = require('http')
const { createHook } = require('async_hooks')
const reused = Symbol('reused')

const asyncHook = createHook({
  init (asyncId, type, triggerAsyncId, resource) {
    if (resource[reused]) {
      throw new Error('reuse detected')
    }
    resource[reused] = true;
  }
})
asyncHook.enable()

const server = http.createServer(function (req, res) {
  res.end();
})
server.listen(3000, function () {
  http.get('http://127.0.0.1:3000', function () {
    http.get('http://127.0.0.1:3000', function () {
      console.log('passed');
      server.close();
    });
  });
});

is the test you need.

[AndreasMadsen]

@mcollina tested it now, the issue is still there.

[mmarchini]

Implementing a wrapper on HTTPParser will be tricky 🤔. Do we really need to reuse HTTPParsers though? I changed the max amount of parsers in the FreeList at to 0, which forces Node.js to always create a new parser, and our benchmarks don't show significant regressions:

                                      confidence improvement accuracy (*)   (**)  (***)
 http/bench-parser.js n=100000 len=16                -2.47 %       ±3.05% ±4.07% ±5.34%
 http/bench-parser.js n=100000 len=32                -1.58 %       ±3.74% ±4.98% ±6.48%
 http/bench-parser.js n=100000 len=4                 -0.02 %       ±4.74% ±6.31% ±8.22%
 http/bench-parser.js n=100000 len=8                 -0.94 %       ±3.44% ±4.58% ±5.96%

Be aware that when doing many comparisons the risk of a false-positive
result increases. In this case there are 4 comparisons, you can thus
expect the following amount of false-positive results:
  0.20 false positives, when considering a   5% risk acceptance (*, **, ***),
  0.04 false positives, when considering a   1% risk acceptance (**, ***),
  0.00 false positives, when considering a 0.1% risk acceptance (***)

(I might be running the wrong benchmarks though. Are there any other recommended benchmarks I should run?)

[mcollina]

Why don’t you send a PR that removes that reuse? Then we can run our benchmark ci. I think it might not be needed anymore because: a) we have a new GC b) referencing from “old” objects to “new” objects is costly, and this reuse was causing that. In an ideal world, reusing that should not be needed. Il giorno sab 10 nov 2018 alle 00:02 Matheus Marchini < [email protected]> ha scritto:

…

Implementing a wrapper on HTTPParser will be tricky 🤔. Do we really need to reuse HTTPParsers though? I changed the max amount of parsers in the FreeList at to 0, which forces Node.js to always create a new parser, and our benchmarks don't show significant regressions: confidence improvement accuracy (*) (**) (***) http/bench-parser.js n=100000 len=16 -2.47 % ±3.05% ±4.07% ±5.34% http/bench-parser.js n=100000 len=32 -1.58 % ±3.74% ±4.98% ±6.48% http/bench-parser.js n=100000 len=4 -0.02 % ±4.74% ±6.31% ±8.22% http/bench-parser.js n=100000 len=8 -0.94 % ±3.44% ±4.58% ±5.96% Be aware that when doing many comparisons the risk of a false-positive result increases. In this case there are 4 comparisons, you can thus expect the following amount of false-positive results: 0.20 false positives, when considering a 5% risk acceptance (*, **, ***), 0.04 false positives, when considering a 1% risk acceptance (**, ***), 0.00 false positives, when considering a 0.1% risk acceptance (***) [image: compare-plot] <https://user-images.githubusercontent.com/4048656/48292582-c08f1b00-e42f-11e8-858d-a38f762a5be5.png> (I might be running the wrong benchmarks though. Are there any other recommended benchmarks I should run?) — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#248 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AADL4x8gQWOahwr1PvostXXUaOMNCBuPks5utgmYgaJpZM4Xru28> .

[mmarchini]

That's what I was thinking. Also, I don't think there would be any difference between removing the reuse and wrapping HTTPParse (we would need to allocate the wrapper every time anyway).

[Flarna]

I think the main difference is that a wrapper object would be a simple javascript object linked to HTTPParser and this link is cut once the parser is moved into the freelist whereas a HTTPParser object has a native part attached so it's more heavy.
But I fully agree that it's even better to avoid the complexity of introducing a wrapper just because of this.

[Flarna]

Besides the reuse of HTTPParser I found two more places for reuse in node core:

• HttpAgent reuses it's handle
• FileHandle::ReadStart() reuses FileHandleReadWrap (happens at least in Http2)

For the first one I created nodejs/node#27581 to avoid the reuse. Depending on the result of this one I can create one more to avoid the reuse of FileHandleReadWrap.

[github-actions]

This issue is stale because it has been open many days with no activity. It will be closed soon unless the stale label is removed or a comment is made.

[mmarchini]

@Flarna is the FileHandleReadWrap still reused? If so I suggest creating a separate issue for it on nodejs/core and closing this one.

[Flarna]

Reuse of FileHandleReadWrap should be fixed by nodejs/node#31972