1352486 - (emberperf) [meta] Improve performance on EmberPerf benchmark

Status: NEW

(Core :: JavaScript Engine, defect, P3)

Description

[Tom S [:evilpie]]

http://emberperf.eviltrout.com/

I just ran "Render Complex List" on Nightly and we are only about half as fast as Chrome.

Comment 1

[Guilherme Lima]

Dupe of bug 1133864? Or at least that bug should depend on this one.

Comment 2

[Jan de Mooij [:jandem]]

I'll take a look next week. I also want to figure out where that high error % comes from.

Comment 3

[Chris Peterson [:cpeterson]]

Jeff Atwood reports that Chrome Canary and Edge both recently doubled their "Render Complex List" performance:

https://twitter.com/codinghorror/status/841775015917887488

https://twitter.com/codinghorror/status/847693460572225536

Comment 4

[Jan de Mooij [:jandem]]

(In reply to Jan de Mooij [:jandem] from comment #2)
> I also want to figure out where that high error % comes from.

Part of this is EAGER_ALLOC_TRIGGER GCs, if I just disable that GC trigger my error % drops from > 100% to 75-80%. That's still nothing to be proud of, but it suggests GC definitely plays a role here. The next step is understanding what kind of GC things we allocate here.

Comment 5

[Jan de Mooij [:jandem]]

OK, we're allocating (among other things) a lot of FAT_INLINE_STRING, STRING, and EXTERNAL_STRING arenas. There might be some easy wins here like bug 1352323. It's unfortunate we don't nursery allocate strings yet because I bet many of these are short-lived.

Comment 6

[Jan de Mooij [:jandem]]

Ember (still) has some terrible logging. We're allocating many strings here:

  _glimmerUtil.LOGGER.debug('[VM] OP ' + opcode.type);

The debug-function is quite complicated so it's not trivial for the JIT to optimize this concatenation away.

We really want to use nursery allocation for these strings so we don't trigger full GCs.

Comment 7

[Chris Peterson [:cpeterson]]

nominating for QF triage

Comment 9

[Tom S [:evilpie]]

I think we should be mentioning the logging issue to the Ember team, seems like something they could be interested in fixing on their side.

Comment 10

[Desigan Chinniah [:cyberdees] [:dees] [London - GMT]]

(In reply to Tom Schuster [:evilpie] from comment #9)
> I think we should be mentioning the logging issue to the Ember team, seems
> like something they could be interested in fixing on their side.

:jandem and others have access to our mailing list with the Ember core team. I'd encourage them reaching out if we deem that Tom Dale and others would find this interesting...

Comment 11

[Jan de Mooij [:jandem]]

The code is transpiled to JS (Babel) and Object.setPrototypeOf is used on functions to represent class inheritance. This confuses TI and we don't inline these functions in Ion, and then we end up with megamorphic ICs. This should be fixable, I just need to figure out the best way to do it.

Comment 12

[Chris Peterson [:cpeterson]]

Robert Jackson says Ember's new debug stripping should remove the logging function calls and avoid the string allocations. The Ember team is considering back porting to Ember LTS (2.12).

https://github.com/glimmerjs/glimmer-build/pull/42

Comment 13

[Mike Conley (:mconley) (:⚙️)]

Here is a link to a profile from a user that was viewing Twitch (which uses the Ember framework, apparently), in the background:

https://perfht.ml/2qAJFHe

Quite a bit of main thread jank on Content Process 5. Not sure if this is useful, but thought I'd attach.

Comment 14

[Florian Bender]

How's the situation changed with Fx 55 and Bug 1352323 fixed?

Comment 15

[Chris Peterson [:cpeterson]]

I see no improvement in Firefox 55 or 56 on my Windows 10 laptop. Results for the "Render Complex List" test with Ember version 2.13.0:

* Firefox 54 = 12.56 ops/sec
* Beta 55    = 12.89 ops/sec
* Nightly 56 = 12.04 ops/sec
* Edge 15    = 16.20 ops/sec
* Chrome 60  = 39.24 ops/sec

Comment 16

[Guilherme Lima]

You can follow on https://arewefastyet.com/#machine=35&view=breakdown&suite=emberperf