Open Bug 1503559 Opened 6 years ago Updated 2 years ago

Experiment: train PGO build with full websites, evaluate performance with Visual Metrics

Categories

(Core :: Performance, enhancement, P3)

Desktop
Windows
enhancement

Tracking

()

Performance Impact none

People

(Reporter: acreskey, Unassigned)

References

(Blocks 1 open bug)

Details

Currently the PGO builds use the following set of tests and pages to train the optimizer:
https://searchfox.org/mozilla-central/source/build/pgo/index.html

These tests may or may not exercise sufficient codepaths to optimize for page load.
This bug is to run an experiment where the PGO is trained on live websites. For example, the sites in the tp6-p1 corpus.
Whether or not this produces a better optimization can be tested by running the a visual metrics comparison with Browsertime.
Blocks: 1505256
Assignee: nobody → brennie
Status: NEW → ASSIGNED
FWIW (sorry if I'm repeating something you already know) in Firefox 63 we switched from MSVC to clang-cl on Windows.  Before that for many years we didn't change the PGO training set all that much mostly out of the fear of triggering new PGO crashes caused by MSVC bugs.  Now we don't have that limitation any more \o/.

Bug 1356652 was the last major change to our PGO training set where we imported the Speedometer benchmark.  Besides that, the rest of what we run in the PGO training set hardly tests anything interesting in the browser.  There is sunspider and a couple of other pages that test a few DOM APIs and that's about it.

I'm very excited to see what improvements we can unlock by including better pagesets in our PGO training set.  Thanks a lot for taking on this challenge.
Priority: -- → P3

Hi Barret - we're thinking now that this would be better done with Visual Metrics.
Are you interested in trying this again?
I can run the visual metrics if you can supply the binaries...

Flags: needinfo?(brennie)

Updated: a simpler strategy of training against live sites and comparing performance with the more meaningful visual metrics

Summary: Experiment: train PGO build with raptor tp-6 tests → Experiment: train PGO build with full websites, evaluate performance with Visual Metrics

Yeah I know the incantations to make PGO point at different sites, but I am busy probably for the rest of the quarter. I can take a look at this then.

Flags: needinfo?(brennie)

I did a quick test where I saved a few sites (facebook/google/bing) to disk and ran the android PGO training run on them.

Maybe some modest ~3% improvements on facebook in this compare
and also here with an -O2 optimized binary.

https://treeherder.mozilla.org/perf.html#/compare?originalProject=try&originalRevision=7482c843963313d2cd4104545cdb0f466121ab14&newProject=try&newRevision=44752d230a0c19256fe0f7b136cef4b37a852f56&framework=10

Blocks: 1543849
See Also: 1543849
Assignee: brennie → nobody
Status: ASSIGNED → NEW
Performance Impact: --- → -
Severity: normal → S3
You need to log in before you can comment on or make changes to this bug.