Perf Testing Meeting Notes

classic Classic list List threaded Threaded
3 messages Options
Reply | Threaded
Open this post in threaded view
|

Perf Testing Meeting Notes

Vladimir Vukicevic-3

Here are my notes from the perf testing meeting that happened last Wednesday;
I hope I didn't miss anything, as I was trying to both participate and take
notes.

- Tp does a full network-to-gfx test
-- hard to isolate different components

- Tgfx/Tr test just gfx performance
- Tp2 is using local http server instead of file URLs

- perf characteristics for loading over http and loading file are very different
-- testing http over a throttled connection so that we can control the exact
packet latency/bandwidth

- timeline
-- revive timeline stuff?

- Tp
-- onload only fires after final paint
-- we want to be able to measure the first paint (but first paint might not
be useful at all!)
-- but we really want "how does it feel to the user"

- alternative tests
-- adding cpu usage - load machine with 50% another program doing cpu usage
-- reading os scheduling info or instruction counter

- new Tp pagesets
-- graydon's memory profile pageset
-- darin's extension for saving and santizing a page
-- google team working on Tp set

- baseline generation
-- 1.0, 1.0.8, 1.5, 1.5.0.*
-- rhelmer has some tools for generating baseline numbers, needs more automation
-- would be useful to be able to give it a time range and intervals and have
it test (e.g. test one trunk build each day between A and B)

- running simultaneous test runs (on different but identical machines)
-- with:
--- profiling: jprof/oprofile/sysprof
--- valgrind
--- timeline
--- other tools
-- contentsink makes decisions based on wall clock time, this is bad

- more specific tests
-- gfx (improve Tr/Tgfx some, pull out svg)
-- layout: trigger gfx repaint, reflow, full frame reconstruct for the same
cotent so that you can compare the differences
-- html parser, xml parser
-- darin: 30% of Tp2 is everything but layout

- function-level profiling
-- take advantage of profiling tools e.g. oprofile, and create a comprehensive
list of gecko functions for profiling purposes
-- assign each function's time to caller or to a category bucket, e.g. "layout"
"network" "parser" "gfx" etc.
-- lets us get very useful performance data without having to manually instrument
code
-- create one master list, write tools to transform that list into something
that's usable by jprof/oprofile/vtune/shark/etc.

- to do
-- vlad: improve Tr/Tgfx tests
-- vlad: new graph server bring-up
-- darin: get firefox extension for page saving out
-- rhelmer: add new tinderbox that can run tests under jprof/oprofile/etc.
-- bz: revisit long-running DHTML tests
-- look into doing tests with extra cpu load
-- create standard include/exclude list of functions that assigns costs of
specific functions to specific modules
--- dbaron: has a version of this for jprof, should look into typing it in
to oprofile as well
--- crowder: to look into whether we can do this with vtune/shark
-- look into copyright issues with pageload pages
-- rhelmer: i18n text pages, need fonts and the like installed on (some?)
tinderbox machines


_______________________________________________
dev-performance mailing list
[hidden email]
https://lists.mozilla.org/listinfo/dev-performance
Reply | Threaded
Open this post in threaded view
|

Re: Perf Testing Meeting Notes

Robert O'Callahan-3
Vladimir Vukicevic wrote:
> -- darin: 30% of Tp2 is everything but layout

Does that mean that 70% of Tp2 is layout?

Rob
_______________________________________________
dev-performance mailing list
[hidden email]
https://lists.mozilla.org/listinfo/dev-performance
Reply | Threaded
Open this post in threaded view
|

Re: Perf Testing Meeting Notes

Boris Zbarsky
Robert O'Callahan wrote:
> Does that mean that 70% of Tp2 is layout?

Note that the "overall" number Tp2 is reporting is just bogus.  See bug
350631.

Not sure whether that's the number that was used for the 30%/70% thing; I seem
to recall Darin looking at individual page numbers sometimes...

-Boris
_______________________________________________
dev-performance mailing list
[hidden email]
https://lists.mozilla.org/listinfo/dev-performance