1613473 - Fenix is slower to load nytimes.com then other browsers

Status: RESOLVED WORKSFORME

(GeckoView :: General, defect, P2)

Description

[Jeff Muizelaar [:jrmuizel]]

See here:
https://www.reddit.com/r/firefox/comments/ez2vty/fenix_loads_pages_incredibly_slowly_34_slower/

Comment 1

[Andrew Creskey [:acreskey]]

I did a visual metrics run with browsertime and Fenix compared favorably to Fennec 68 (Pixel 3, 5 cold page loads, live site)

Fennec68
[2020-02-05 13:41:43] INFO: https://www.nytimes.com 0, backEndTime: 126ms (±3.54ms), firstPaint: 827ms (±3.48ms), firstVisualChange: 966ms (±5.47ms), DOMContentLoaded: 979ms (±28.68ms), Load: 7.70s (±471.58ms), speedIndex: 4421 (±13.32), perceptualSpeedIndex: 5985 (±37.58), contentfulSpeedIndex: 966 (±5.47), visualComplete85: 14.13s (±640.23ms), lastVisualChange: 14.13s (±640.23ms), rumSpeedIndex: 279 (±13.75) (5 runs)

Fenix beta
[2020-02-05 13:45:06] INFO: https://www.nytimes.com 0, backEndTime: 157ms (±10.76ms), firstPaint: 631ms (±134.02ms), firstVisualChange: 741ms (±136.02ms), DOMContentLoaded: 1.06s (±59.64ms), Load: 3.40s (±70.79ms), speedIndex: 4095 (±258.38), perceptualSpeedIndex: 5618 (±154.92), contentfulSpeedIndex: 741 (±136.02), visualComplete85: 11.75s (±183.97ms), lastVisualChange: 11.75s (±183.97ms), rumSpeedIndex: 346 (±73.14) (5 runs)

However when pasting in the URL, Fenix's performance may be worse than Chrome or Fennec 68's.
(Note, n=1)
https://streamable.com/3j0w8

Comment 2

[Jeff Muizelaar [:jrmuizel]]

Emily, what info do you want from Kris?

Comment 3

[Emily Toop (:fluffyemily)]

Kris was in our triage meeting and said that he would take a look at this issue, so I tagged him. I should have written that as a summary, sorry.

Comment 4

[Kris Taeleman (:ktaeleman)]

I was just going to make some profiles to see if there is anything obvious that we can do.

Non WR profile: https://perfht.ml/2UuNfBn
WR profile: https://perfht.ml/2OzlLa6

Jeff: does anything jump out to you?

Comment 5

[Jeff Muizelaar [:jrmuizel]]

Most of the time is being spent in Javascript. We do spend some time doing main-thread paint in the non-wr profile that is eliminated with WebRender, but it's not that much of the total time.

It is worth noting that the original reporter claims that enabling WebRender made a big difference
" As it turns out, the issue on my end seemed to be WebRender being disabled. Since there's no way to enable it on the regular Firefox Preview build I didn't even know that could be a factor, but after I turned it on in the nightly build on about:config the page load speed increased dramatically and the browser stopped being unresponsive on loads. I'd say now the speed is more or less the same as Fennec, which is a huge improvement from what I was dealing with before. Thank you to all those who tried to help."

Comment 6

[Michael Comella (:mcomella) [NI reported issues only: ex-Mozilla]]

However when pasting in the URL, Fenix's performance may be worse than Chrome or Fennec 68's.

acreskey, what does this mean? Can this be a front-end/application-level issue?

---

Note that Vesta has similarly long delays for page load when clicking google search results in https://github.com/mozilla-mobile/fenix/issues/7903. Without much other evidence than these, I'm concerned that a small minority of our users could be experiencing very poor performance that we may not be reproducing locally.

Comment 7

[Andrew Creskey [:acreskey]]

(In reply to Michael Comella (:mcomella) [needinfo or I won't see it] from comment #6)

However when pasting in the URL, Fenix's performance may be worse than Chrome or Fennec 68's.

acreskey, what does this mean? Can this be a front-end/application-level issue?

---

Note that Vesta has similarly long delays for page load when clicking google search results in https://github.com/mozilla-mobile/fenix/issues/7903. Without much other evidence than these, I'm concerned that a small minority of our users could be experiencing very poor performance that we may not be reproducing locally.

Michael, yes, I share this concern.
Browsertime navigates by setting window.location to the given URL. So it does measure most of the codepaths involved pageload, but it also misses some.
I'll leave my needInfo on this as I'll be trying to make it navigate in a more 'realistic' manner.
The hope is that it will either find additional problems, or else provide evidence that this isn't a problem.

Comment 8

[Michael Comella (:mcomella) [NI reported issues only: ex-Mozilla]]

I'll leave my needInfo on this as I'll be trying to make it navigate in a more 'realistic' manner.

Marc has a WIP where he triggers UI navigation via code built into the APK (link – Marc might pull this code out of the PR though): maybe you can sync up with Marc on that approach and if it'd work for y'all.

Comment 9

[Andrew Creskey [:acreskey]]

Jeff, can you reproduce this on a recent build of Fenix? e.g.
https://firefox-ci-tc.services.mozilla.com/tasks/index/project.mobile.fenix.v2.fennec-production.2020.02.20/latest

Comment 10

[Jeff Muizelaar [:jrmuizel]]

I've never tried reproducing this. I was reporting on behalf of the reddit user.

Comment 11

[Andrew Creskey [:acreskey]]

This may have been caused by this (now fixed) issue:
https://github.com/mozilla-mobile/android-components/issues/5880

Comment 12

[Tom S]

My issue was implicitly linked to this bug and I can't reproduce the big slowdown when loading HN anymore. +1