Investigate ~10% ts_paint improvements with XUL persist
Categories
(Core :: XUL, enhancement, P5)
Tracking
()
People
(Reporter: bdahl, Unassigned)
Details
(Keywords: perf:startup)
Attachments
(1 file)
|
4.07 KB,
text/plain
|
Details |
When investigating performance in htmlify browser.xhtml bug we ran into a 10% ts_paint improvement. However, the improvement was a from a bug where XULPersist was handling restoring the width/height attribute of the root element instead of AppWindow.
We could potentially try to mimic the bug and move more of the width/height handling into XULPersist, but I'm not sure if that's feasible with how the width/height is applied before the first layout.
|
||
Comment 1•6 years ago
|
||
As Comment 0 says this may not be feasible. But we wanted to get something on file in case perf folks want to investigate since the improvement we were seeing on windows ts_paint was pretty huge.
|
||
Comment 2•6 years ago
|
||
So I'm a little confused from comment #0 - can you elaborate a bit? Did the improvement happen and then it went away again? Did bug 1590044 regress ts_paint by about 10% again? Do we know what caused the improvement in the first place? Is it that we weren't applying the width/height at all? Is it the difference in xulpersist IO times? Or something else?
|
Reporter | |
Comment 3•6 years ago
|
||
The improvement showed up in tests we ran before landing <html> in browser.xhtml. However, that improvement went away when bug 1590044 was fixed. Things never regressed.
I did some talos pushes to see if the improvement is still there if I re-introduce bug 1590044 and it looks like it still shows the improvement.
I thought the improvement came from setting the width/height earlier, so I added some logging to see the ordering of applying attributes to the root element and various DOM events. I then tried to mimic that log by setting the root element attributes earlier. However, I saw no performance improvement doing that.
I'll attach a file with the logging with various patches applied.
|
|
Reporter | |
Comment 4•6 years ago
|
||
|
|
Reporter | |
Updated•6 years ago
|
|
|
||
Updated•6 years ago
|
|
||
Comment 5•6 years ago
|
||
It would be nice to compare profiles with and without the patch that caused the 10% improvement. Marking fxperf:p2 for now as it seems Gijs would like to investigate.
|
|
||
Comment 6•6 years ago
|
||
I've started gecko profiler jobs for ts_paint (o) and ts_paint_webext (g5) on both trypushes from comment #3 . I may have time to investigate tomorrow, otherwise next week. Keeping ni.
|
|
||
Comment 7•6 years ago
|
||
Well, that didn't work. I filed bug 1600279.
Will see if I can do my own trypushes with mach and get profiles that way...
|
||
Updated•5 years ago
|
|
|
||
Updated•3 years ago
|
|
|
||
Comment 8•3 years ago
|
||
Realistically I think we should close this incomplete. Florian, does that sound OK?
|
|
||
Comment 9•3 years ago
|
||
(In reply to :Gijs (he/him) from comment #8)
Realistically I think we should close this incomplete.
Agreed.
Description
•