416068 - 10% Linux Tp Regression on 1-29-2008

Status: RESOLVED FIXED

(Core :: Layout: Text and Fonts, defect, P1)

Description

[Mike Schroepfer]

See url for graph range.  This is Linux only. Checkin range:

http://bonsai.mozilla.org/cvsquery.cgi?treeid=default&module=PhoenixTinderbox&branch=HEAD&branchtype=match&dir=&file=&filetype=match&who=&whotype=match&sortby=Date&hours=2&date=explicit&cvsroot=%2Fcvsroot&mindate=1201593049&maxdate=1201616123

I'm going to guess Bug 401988

Comment 2

[Karl Tomlinson (:karlt)]

Thanks reed but bug 414692 was filed for changes from attachment 298411 [details] [diff] [review] in bug 401988 checked in 2008-01-26 23:37.

The patch from bug 401988 checked in during this range was attachment 299736 [details] [diff] [review], which changed preferred fonts for Japanese and Chinese languages.  This is likely to have affected performance with Japanese and Chinese languages, but wouldn't affect other languages at all, so I'm surprised to see such a large regression.  I'll look further into the data.

Comment 3

[Karl Tomlinson (:karlt)]

Checkin range from qm-mini-ubuntu01

http://bonsai.mozilla.org/cvsquery.cgi?module=PhoenixTinderbox&date=explicit&mindate=1201599420&maxdate=1201604999

And it is the Chinese and Japanese pages affected, so Bug 401988 was a good guess, Mike.

www.sogou.com is particularly badly affected (444->1416)

Comment 4

[Karl Tomlinson (:karlt)]

I couldn't reproduce a change in page load time from attachment 299736 [details] [diff] [review] on my system until I disabled the good Japanese and Chinese fonts that were installed so that there was only a bitmap Japanese font (and Chinese pages where rendered using the Japanese font).

This makes me wonder what font support there is for Japanese and Chinese on the talos machines.  I wonder if there is a quick way to find out what fonts are installed?  Anyone with a reference platform able to attach the output of "fc-list : family foundry lang file"?

I was planning to back this change out anyway because Masayuki pointed out that it is a hidden pref that can interfere with people's fontconfig settings resulting in confusion.

But the fact that this made such a difference (even if only when only bitmap Japanese fonts are available) is not expected and may be worth investigating because something may not be behaving as expected.

Comment 5

[Mike Schroepfer]

Aravind can you run "fc-list : family foundry lang file"? on the linux talos machines?

Comment 6

[Aravind Gottipati [:aravind]]

Attachment: Output from "fc-list : family foundry lang file"

Comment 7

[Aravind Gottipati [:aravind]]

Its the same across the other 4 machines.

Comment 8

[Karl Tomlinson (:karlt)]

Thanks very much Aravind.  That should help me to work out what is happening (although I haven't been able to reproduce yet with those fonts).  The fonts should provide Chinese and Japanese coverage AFAICT with my lack of knowledge of the scripts involved.  I can see why CJK Linux users prefer Western fonts for Latin characters.

The patch would have caused fallback to secondary fonts most of the time.  A couple of possible explanations for the poor performance resulting from this are:
This would have meant that GetMetrics would have been called for more fonts and this can be slow (Bug 403618 comment 5).
Also the word cache doesn't hold words constructed from secondary fonts across pages, which is part of the reason for bug 416475, but, as Chinese and Japanese pages don't really use space-separated words, that is less likely to be significant here.

The patch that caused the regression has now been backed out and the talos pageload time looks to be returning to its previous value.