I observed some long GC times while loading the Facebook.com newsfeed: ~300ms of GC for the facebook.com page which loads in 2 seconds on my i7 MacBook Pro. Similar behaviour on Windows. See attached logs.
SPS profile (click on Content Process panel): https://cleopatra.io/#report=ba3a919e09652a157707aee442c0d735e8ee70de&search=%3A%3Agc%3A%3AGCruntime&selection=0,3213,3214,3215,3216,2777,3217,8,9,152,153,154,155,156,157,158,159,160,161,162,163,164,165,166,3787,3788,3505,3506,3507,876,877,3789,3790,3791,3792
<vladan> did you see anything interesting in that log? <sfink> anyway, there wasn't anything huge, but one slice missed its 40ms target (took 67ms), which we should probably clean up <sfink> and given that you're timing initial page load, I'm guessing that the incremental slices aren't doing you any favors to begin with <sfink> I'm not sure how much overhead they induce, but if it's significant, it might be worth knowing that yielding to the mutator isn't going to do any good yet <sfink> and then there were some suspiciously long sweep phases <sfink> they may be the most actionable <sfink> it looked like GC/CC interaction (gray marking)
2 years ago
plat-rel changed to - as we're already tracking the meta as priority
platform-rel: + → -
Is there anything we can do here, Steve or Olli? Will any of the recent GC changes have a positive effect?
Is this still happening? Especially when e10s is enabled? Nightly should trigger way more zone GCs which should be faster than full GCs.
Flags: needinfo?(bugs) → needinfo?(vladan.bugzilla)
What is the STR here? Load Facebook when logged in? Or does newsfeed mean something else than the frontpage there?
Whiteboard: [platform-rel-Facebook] → [qf: investigate: p1][platform-rel-Facebook]
(In reply to Olli Pettay [:smaug] (work week, slower than usual reviews) from comment #8) > What is the STR here? Load Facebook when logged in? Or does newsfeed mean > something else than the frontpage there? I can no longer reproduce the long GC times in Nightly, perhaps it has something to do with e10s or longer-running sessions. currently I see about 50-100ms GC in the parent process and 20-50ms in the child process during newsfeed loading, which I think is ok. For future reference, I made a simple static benchmark representing a logged-in Facebook newsfeed, you can find it & setup instructions here: https://www.dropbox.com/sh/fw5odcyb2mpn7ag/AADGuImIda4B1Rz03eoMYN33a?dl=0
We do compartment GCs these days way more often, perhaps that has helped.
Closing as per comment 9, please re-open if this is still a problem.
Status: NEW → RESOLVED
Last Resolved: 10 months ago
Resolution: --- → WORKSFORME
You need to log in before you can comment on or make changes to this bug.