136677 - page-load (tp) increased by 1.5%

Status: RESOLVED FIXED

(Core :: XPCOM, defect)

Description

[Cathleen]

tinderbox showing page load number increased from 1195 to 1210s

check-ins possibly caused the increase are:
http://bonsai.mozilla.org/cvsquery.cgi?module=MozillaTinderboxAll&date=explicit&mindate=1018405320&maxdate=1018414079
http://bonsai.mozilla.org/cvsquery.cgi?module=MozillaTinderboxAll&date=explicit&mindate=1018414080&maxdate=1018415519

assign to pavlov for now...
cc'ing everyone else who checked in during those two checkin time frame, for
other possible suspects.  :-)

Comment 1

[Stuart Parmenter]

That 1195 was a fluke I think.  so it really jumped from 1205 to 1210 or so.  My
patch shouldn't have effected Linux at all, but I'll look to see.

Comment 2

[Cathleen]

we now think the number is a fluke...

btek reported 1215, 1217 
backed out pav's change, btek reported 1199, 1198
put pav's change back into btek, now reports 1189

unexplainable reporting... so putting btek back online anc close the bug.

Comment 3

[Cathleen]

after putting btek back online, btek is again reporting 1216 or somewhere there...
pavlov is going to try backout his changes later today on the trunk.

Comment 4

[Stuart Parmenter]

ok, I don't know what the deal with this is.  btek seems to go up, but luna
stays the same.  I backed myself out earlier and re-landed afterwards to see
what would happen with the tinderbox numbers.  I'm unable to reproduce the
slowdown on my solaris box.  Since the patch for 129953 gives a *signifigant*
speedup for DHTML pages on windows, I've chosen to leave the patch in right now.
 I will be out of the office until April 22.  If you want to back it out again,
thats fine with me, but please post something to bug 129953 letting them know
what the status is.

Comment 5

[Chris McAfee]

luna and btek are building slightly different things.
btek disables crypto, tests, and xprint.  luna enables
crypto, and has all extensions turned on.  I suspect you're
twiddling something that luna is already paying the price for,
but btek is not, hence the jump on btek.

Comment 6

[Brendan Eich [:brendan]]

So jrgm may be able to help -- by running timers on the main thread on Windows
when idle, we are probably adding a bit, noisily, to timer latency.  The page
load tests may care about that average increase in latency.

/be

Comment 7

[David Baron :dbaron:]

Considering the comments above, should this really be a smoketest blocker?

Comment 8

[Cathleen]

when pav backed out shortly last night, btek tp showed 1207, 1207, 1206, 1206
then pav re-checked in btek tp back to 1218

Comment 9

[Kathleen :Brade]

I have backed out pavlov's checkins.
Resolving this bug as fixed

Comment 10

[Randell Jesup [:jesup] (needinfo me)]

Final numbers given seem to be 1%.

For a significant improvement in DHTML functionality (not just performance) this
seems worth it.  DHTML functionality was broken by the previous timer checkins
to reduce latency; this fixes that regression.

Comments to be mirrored on 129953

Comment 11

[Stig Nygaard]

Is this more important than bug 129953 which a bugfix for seems to be the reason
for the little time-increase ?

Comment 12

[Chris McAfee]

reopening, I don't think we're done with this discussion yet.

Comment 13

[hacker formerly known as [e-mail]]

Here's another datapoint: the Txul numbers on slag jumped from ~2150 to ~2650
when brade backed out pavlov's checkin.  Looking at the history from last night,
a similar jump occurred when pavlov backed himself out last night.  slag isn't
one of our officially sanctioned performance boxes but the numbers seem
relatively consistent.

Comment 14

[Chris McAfee]

off blocker list, per discussion @ 3pm in my cube

Comment 15

[Jaime Rodriguez, Jr.]

Marking as nsbeta1+/ADT1

Comment 16

[Cathleen]

i think we should close this bug, as pavlov's change was backed out on Friday.

we need one more pass on the patch for bug 129953, the fix is only for windows
platform, but linux page-load was affected by the patch.  

please see bug 129953 for more details.

Comment 17

[Asa Dotzler [:asa]]

removing from RC1 tracking bug.

Comment 18

[Paul Wyskoczka]

Removing adt1 since it was make so for MachV.  Now tracking for Buffy