463592 - changes in Tp data from bug 449356

Status: RESOLVED WORKSFORME

(Core :: Graphics, defect)

Description

[Karl Tomlinson (:karlt)]

Summary of changes in Tp data from bug 449356:

                         pre-449356 post-449356   change (ms)  change (%)

new (large) page set:

  mean 1st page load       583         439        -144          -25
  mean median page load    427         425          -2.5         -0.6

old (small/fast) page set:

  mean 1st page load       480         345        -135          -28
  mean median page load    237         257         +19.2         +8


Note that talos loads each page in the page set, then cycles through the same
set another 9 times.

The "mean first page load" data comes only from the first run through the
page set.  These numbers are representative of viewing a page from a new site,
and closer to representative of zooming a page.

Talos normally reports the medians.  The "mean median page load" is the result
of taking the median over all 10 cycles for each single page, then taking the
mean over all pages.  These numbers represent a cached page load time and are
representative of reloading the existing page or viewing a page with very
similar fonts (probably on the same site).

Most of the numbers here are improvements, and when also considering the Ts
and Tp_RSS gains, we're heading in the right direction.

The one number is worse is the cached page load time for the old page set.
This is an indicator that there is room for improvement here.  These pages are
such that the fonts used when reloading the page were almost always found in
5000 entry gfxPangoFontCache (see bug 424622 and bug 428425), which we no
longer have.

The gfxPangoFontCache used several megabytes of memory, but I imagine we could
achieve similar (or better) reload performance with a much smaller
gfxFontGroupCache with font groups expired based on time.  I've filed bug
463576 on this.