Javascript Performance Thread

[Grantius]

It's nice to keep up to date with competition. There hasn't much news recently about Spidermonkey

[Zlip792]

I think this topic is for Firefox's Javascript engine, not Chrome's..

Fair enough. I'm fine to not post about other Javascript engines.

So self reminder, no competitor's JS Engine posts.

[Timvde]

I also think it's interesting to keep up with the competition. Although in the builds forum, the (current) title of this thread is "Javascript Performance Thread", not "Spidermonkey Performance Thread". So while this should not convert into a Chrome or Edge performance thread, I do think it's interesting to not only focus on Firefox here. It's an interesting world out there, let's embrace it.

[Ver Greeneyes]

I have no problem with bringing up the techniques behind the other JS engines either. I don't think there's much to report on the SM front right now. Various projects are underway, but nothing with fast payoff - but I've been very busy recently, so I might be a bit out of the loop.

[KilliK]

It's nice to keep up to date with competition. There hasn't much news recently about Spidermonkey

I am all for updates regarding the current competition , but his last link was about a V8 presentation article written 2 years ago.

[Ver Greeneyes]

djvj has put his patches to make parsing of blocking scripts asynchronous up for review (after finally managing to fix some extremely obscure crashes), with some interesting news: they improve the V8_v7 score in the Talos performance testing framework by 30% on Windows (and as much or more on Linux)! Unfortunately the per-test breakdown doesn't seem to load, so we can't tell what's improving, but it'll be interesting to see if the improvement is reflected in AWFY as well. I seriously doubt the effect will be as significant, though it would be amazing if it were - I think it's more likely that there's something about the Talos testing framework that was slowing things down.

[Grantius]

djvj has put his patches to make parsing of blocking scripts asynchronous up for review (after finally managing to fix some extremely obscure crashes), with some interesting news: they improve the V8_v7 score in the Talos performance testing framework by 30% on Windows (and as much or more on Linux)! Unfortunately the per-test breakdown doesn't seem to load, so we can't tell what's improving, but it'll be interesting to see if the improvement is reflected in AWFY as well. I seriously doubt the effect will be as significant, though it would be amazing if it were - I think it's more likely that there's something about the Talos testing framework that was slowing things down.

Wonder how this affects add-ons which deal with content code like AdBlockers

[Timvde]

djvj has put his patches to make parsing of blocking scripts asynchronous up for review (after finally managing to fix some extremely obscure crashes), with some interesting news: they improve the V8_v7 score in the Talos performance testing framework by 30% on Windows (and as much or more on Linux)! Unfortunately the per-test breakdown doesn't seem to load, so we can't tell what's improving, but it'll be interesting to see if the improvement is reflected in AWFY as well. I seriously doubt the effect will be as significant, though it would be amazing if it were - I think it's more likely that there's something about the Talos testing framework that was slowing things down.

Not sure what's going on when you try to load it, but it loads just fine here. Results (click for full size):



From here: https://treeherder.mozilla.org/perf.htm ... 349bc1f1a3

Quite possibly a bug, as there even is no result at all on OS X.

[Ver Greeneyes]

Not sure what's going on when you try to load it, but it loads just fine here. Results (click for full size):

Thanks, it was probably broken for a few days. Kannan reported the same problem at the time. But those scores make no sense! For one thing, why are Earley and Boyer, and Decrypt and Encrypt separate? Even ignoring those, why does RegExp have the highest score when it should have the lowest? It looks like it's displaying the times rather than the scores!

[Timvde]

Displaying time instead of score sounds likely with these results. Which would mean this is a 30% *regression* instead of an improvement

[Ver Greeneyes]

I've needinfo'd the module owner of Talos to clarify the situation, but I think that's likely

[Josa]

The patches were on mozilla-inbound enough time to run on AWFY. There were no improvements nor regressions.

[Timvde]

The patches were on mozilla-inbound enough time to run on AWFY. There were no improvements nor regressions.

Not sure, they got backed out.

Also, for whoever is interested:

[15:15:02] h4writer: Timvde, fyi the talos results are running v8 benchmark, not the updated octane benchmark
[15:15:15] h4writer: Timvde, that explains the huge regexp score etc
[15:15:46] h4writer: Timvde, v8-regexp didn't looked at the result, therefore in Ion we don't create the result array
[15:17:04] Timvde: h4writer: So the patch would be a legit performance win on the old V8 benchmark?
[15:17:28] Timvde: (still think the scores are weird though, most of them are in the hundreds instead of ten thousands)
[15:18:19] h4writer: Timvde, not sure. For the benchmark probably not. But probably the way talos loads the benchmarks alters the score a bit
[15:19:00] h4writer: Timvde, IIUC the patch improves performance of <script> tags inside the html to be not blocking
[15:19:39] h4writer: Timvde, so it most likely is a real world improvement, but probably not for benchmarks

[Grantius]

You mean they are looking at real world speed? Awesome

[Ver Greeneyes]

As I told Hannes on IRC, I don't think that applies to V8-v7. I ran it locally and the regexp score was the lowest, whereas the raytrace score was the highest. In the results on perfherder, the regexp number is the highest and the raytrace number is the lowest. What Joel said in the bug also seems to confirm this - they're in the process of getting the processed numbers to perfherder, but right now it's just getting the raw results. So the subtest results are accurate, but it's misinterpreting them as scores.

I'm willing to believe that there's no perf difference on AWFY though - there are differences between Octane and V8-v7, and the Talos harness could easily be causing weird differences as well.