201082 - DOM/Style/Perf regression

Status: RESOLVED WORKSFORME

(Core :: DOM: CSS Object Model, defect)

Description

[Marco Perez [:mpk]]

User-Agent:       Mozilla/5.0 (Windows; U; Win98; en-US; rv:1.4a) Gecko/20030401
Build Identifier: Mozilla/5.0 (Windows; U; Win98; en-US; rv:1.4a) Gecko/20030401

Between the releases of Moz1.3 and Moz1.4a something caused a regression
affecting at least the setting of element.style.backgroundColor . It now takes
more than twice the time it did before.

Reproducible: Always

Steps to Reproduce:




I verified it with a few older Phoenix-Win32 builds since the problem is visible
on Phoenix, too. Last known good build was phoenix-win32-2003030408. First known
bad build was phoenix-win32-2003030814. May this have been caused by the fix of
bug #171830 ?

Comment 1

[Marco Perez [:mpk]]

Attachment: changes background on 4096 elements and displays time used

The elements that get their background color changed are table cells. The
effects are also visible when avoiding tables (e.g. in text lines). On my
machine (Athlon XP1800+) the test with IE takes 110ms after the first run, with
Mozilla 1.4a it takes 990ms and with Mozilla1.3 it takes 440ms. If you replace
style.backgroundColor with style.background then IE slows down to Mozilla level
whereas Mozilla's time don't change.

Comment 2

[Boris Zbarsky [:bzbarsky]]

Yes, this is almost certainly a result of bug bug #171830.  Again, you're
testing builds that are in the middle of having changes happen to them.  We're
still working on the basic arch; the changes are yet to come.  So testing is not
very meaningful...

Comment 3

[Zibi Braniecki [:zbraniecki][:gandalf]]

Boris, but isn't it material for perf bug? My results:

Mozilla 1.4a:
Recorded times:
3.313s
1.578s
1.125s
1.281s
1.156s

IE6:
Recorded times:
0.313s
0.110s
0.125s
0.125s
0.109s

Comment 4

[Marco Perez [:mpk]]

Attachment: does basically the same as the first testcase

Since Mozilla is a bit slower at reading an element's title than IE, the
comparison wasn't very fair with the first testcase. In this testcase the color
values are read from an array rather than from each cell's title. On my
machine, this boosts Mozilla's performance by about 15-20%, whereas IE's times
remained the same.

Comment 5

[Marco Perez [:mpk]]

That's interesting. After surfing for a while (Mozilla 1.4a) when the memory
cache is almost filled, the recorded times drop by about 20-25%. They used to be
between 820ms and 880ms. Now they remain stable at 660ms. This seems to be
reproducible. about:cache displays the following stats:

Memory cache device
Number of entries: 	295
Maximum storage size: 	4194304 Bytes
Storage in use: 	4189116 Bytes
...
Disk cache device
Number of entries: 	1556
Maximum storage size: 	51200000 Bytes
Storage in use: 	21607424 Bytes
...

After restarting Mozilla the recorded times are at 880ms again.

Comment 6

[Boris Zbarsky [:bzbarsky]]

OK, I finally looked at the testcase closely.  The only number really worth
paying attention to is the very first one after a reload -- the others indicate
how long it takes to process a change when nothing changed.  The performance of
that specific case (an inline style change that changed absolutely nothing) did
indeed regress quite a bit; the changes for bug 158713 should bring it back
down.  Once that lands, we should retest this and see what it looks like.

Comment 7

[Max Kanat-Alexander]

Bug 158713 has been checked in. I see:

1.723s
0.481s
0.481s
0.481s

On 2003071608 Win2K.

That seems like a significant performance improvement to me, although not on the
first load (that's the same number I got with 1.4final on the first load), which
was the problem (as I understand it).

-M

Comment 8

[Max Kanat-Alexander]

By the way, that's on an Athlon 800 & Geforce2 GTS w/ 512M RAM, Win2K-SP4.

-M

Comment 9

[Marco Perez [:mpk]]

Attachment: improved testcase (same layout but better options)

This testcase allows a few settings (defaulting to more intelligent values than
my former testcases, e.g. random colors). Hiding the elements when repainting
improves performance (about 25%). Note that Mozilla sets "background-color"
about as fast as "background" while IE sets "background-color" faster.

Comment 10

[Marco Perez [:mpk]]

Wow! The performance improvements in todays builds are impressive to say the
least (thanks to the fix for bug #230170?). The results for the testcase are
better than ever. On Win2K, we're almost on par with IE6 and way faster than
Opera. :)
Does depending on the still open bug #213943 prevent this bug from being closed
or can/should it be marked as fixed anyway?

Comment 11

[Markus Hübner]

Fixing bug 213943 will speed things up even more!

Comment 12

[Boris Zbarsky [:bzbarsky]]

So is the regression fixed?  ;)  Can we resolve this bug?

Comment 13

[Marco Perez [:mpk]]

Yes, I think this can be safely marked as fixed,
especially considering that over the last few months
the test results remained stable or even improved a bit.
With the latest firefox builds the results from the
testcase are a bit flaky. They tend to switch between 2
to 4 "normal" results and then one with about 30% lower
times (and so forth). I can only speak for FreeBSD, since
right now I don't have any other platforms on
which to test. If nobody from the Win, Linux or Mac
camp objects I'll close this bug in a few days.

Comment 14

[Jason Bassford]

Last bug mentioned that might have resolved this was mentioned 2 years ago.

-> WORKSFORME