149224 - Slow scrolling and much disk activity; page with 2px x 3000px bg image tiled horizontally

Status: RESOLVED DUPLICATE of bug 148598

(Core :: Layout, defect, P3)

Description

[John Morrison]

I was looking at a bug (can't remember which one) and found a site where I had
really poor scrolling behaviour. Not surprisingly, the page had a large tiled
background image: 2px wide x 3000px high.

One particularly curious aspect of the scrolling behaviour was that there was
_audible_ disk drive activity each time I scrolled on the page. It is just
grinding my disk drive. This continues even when I set disk cache size to
zero.

I then tried this in a current branch build, and found that I had no problem
in that build. So I searched back through builds and found that this was
introduced on the trunk (for my system) between 8am/Apr23 and 8am/Apr24.
Which would seem to make this a consequence of the checkin for bug 102321
'ESPN's new page layout is dog slow in mozilla'.

But instead of reopening that bug, I thought I'd file this as a separate
issue, especially considering the disk activity (wtf?). I can reproduce this
in a current trunk build, so if you (dcone) want to try some experiments, send
me patches.

System info:
  HP Kayak XAs / 500MHz / 128MB / win2k
  Elsa Gloria Synergy (Permedia 2 R17) 8MB
  ViewSonic Professional P810
  1280x1024, 32bit, 75Hz
  Drivers are the latest available for that card.

I'll attach a simple testcase in a bit and others can try and see how this
works with their system configuration. My times for running that test, with
a single maximized window, at 1280x1024, 32bit, 75Hz are:

  current trunk mozilla  : 23 seconds, with lots of disk activity
  current branch mozilla : 11 seconds, with negligible disk activity

Comment 1

[John Morrison]

Attachment: background image; 2px wide by 3000px high (gif)

Comment 2

[John Morrison]

Attachment: testcase; automated, timed scrolling of simple document with background image

Comment 3

[John Morrison]

... and while the disk activity might suggest a cache issue, if I completely 
cover the browser while that testcase runs, the times reported are ~0.5 
seconds to scroll a page (i.e., painting/tiling is still a factor).

Comment 4

[dcone (gone)]

duplicate

*** This bug has been marked as a duplicate of 141786 ***

Comment 5

[John Morrison]

I tried the patch in bug 141786, but that does not cure the problem here.

I put a 'printf("numTiles: %d\n", numTiles);' just before that 'if'
statement, and I get this repeated sequence of 4 lines.
  numTiles: 1
  numTiles: 1
  numTiles: 630
  numTiles: 3432
where sometimes the 630 is instead a 1260. So, I cut out 2 of 4 calls to the
following routine, which is good. But my time the testcase is unchanged (might
be a fraction of a second better), and I still grind my disk drive.

Still, I am better off with the call to 'ProgressiveDoubleBlit' since, if I
'if (0) {}' that section and drop through to the 'slow blit backstop', my time
in the attached testcase shoots up to ~60 seconds, from 23 seconds.

Let me know what other information you need. I'll try the testcase on a few 
other systems and see how they react.

Comment 6

[Ken Harris]

Hmm.  Bug 141786 is for big tables, and John's testcase doesn't have any tables.

WFM on Mac OS X 10.1.5, G4 / 500MHz, Mozilla 2002061014.

If it's a Windows-specific small-background-image-tiling problem, that sounds
more like bug 133261.

Comment 7

[dcone (gone)]

The results you have had on this bug show that progressive doubling does work, 
but maybee the way we store 8 bit gifs as 24 bit images may slow us up.  I have 
a bug open for these issues.. so I will dup this against that bug.

*** This bug has been marked as a duplicate of 148598 ***