Closed Bug 742086 Opened 8 years ago Closed 7 years ago

Need a more comprehensive "suite" of tests for testing checkerboarding performance with Eideticker


(Testing Graveyard :: Eideticker, defect)

Not set


(Not tracked)



(Reporter: wlach, Assigned: wlach)



(Whiteboard: [eideticker:p1])

Right now the eideticker dashboard uses to measure the amount of checkerboarding during a pan, and for testing a zoom. isn't necessarily a good benchmark, or at least not a good benchmark to use on its own. The reason why it's slow (use of gradients fills in the background) are not the same as the reasons why other sites as slow.  I don't really know about nightly, but chances are it's not totally representative either.

We should add some panning/zooming tests that represent real-world browsing experiences. Right now I am thinking:

* An interesting web page on (the full version)

Unfortunately the former two are copyrighted, which makes importing them into the eideticker repository annoying. :( I would really prefer to keep everything there for simplicity's sake. One option here would be to synthesize a test which mimics their essential properties without using copyrighted material. Other ideas welcome.
Assignee: nobody → wlachance
After talking to jmaher, I think we should also probably create this suite in a generic fashion that would also be usable by Talos and friends. This isn't really an eideticker-specific bug (though the work will prolly be used by Eideticker first)
I vote we specifically avoid even if we used it for internal testing. We should avoid testing on any page that is a known slow path of cairo such as a very hard gradients (like you said). We don't want to benchmark cairo slow path performance for our main Eideticker test but rather something that represent better 'typical webpages'. <-- Complex, lots of text, some images. IANAL Copyright should be okay to cache?
a reddit comment page <-- long, lots of text <-- Apparently a popular page according to a random source, public domain
Duplicate of this bug: 744724
I think we need to use timecube. Specifically for checkerboarding tests since our competition doesn't appear to checkerboard when using this page.
timecube is a reflection of skia vs. cairo and not a reflection of how good we are in general. I don't want to have a benchmark that isn't representative of typical web pages and where the work to improve means working towards a skia port.

(In reply to Bob Moss :bmoss from comment #4)
> Specifically for checkerboarding tests
> since our competition doesn't appear to checkerboard when using this page.

That argument can be made for many web pages. Esp. if we're testing against borwser version that don't have async panning.
I don't understand the objection. Does this mean it doesn't matter how we perform on this page relative to our competition? I don't think users care why it checkerboards, if checkerboarding is bad then we need to find a way to make it go away.
That page doesn't not represent typical web performance ours users will see. So no, we should not base our general benchmark on that page. We should still track issues with that page but not benchmark against it.
I think I see the problem. What do you mean by "general benchmark"? My purpose for this test is to track issues with it.
Should definitely do immediately.
Depends on: 745817
Let's do at a relatively high priority as well.
(In reply to JP Rosevear [:jpr] from comment #10)
> Let's do at a relatively high priority as well.

Mirroring planet mozilla is actually a bit tricky because of all the external images/assets in it. I'll try to make it happen though.
We've done this before with our tp pageset. Let's find out how they did it. Maybe joe knows?
Alice pulls the Alexa top500 pages, culls it somewhat, and makes them work offline. She's off right now, though, so I'll CC Lukas, who at one point knew about Talos things and will likely know who to ask if she doesn't know.
I'm adding on Armen, since he helped implement tp5 and is probably more up to date on how we can deploy new talos suites than my crusty knowledge from the past.
The knowledge of creating a pageset is afaik in Alice's head unless it is documented somewhere in the A-team's documentation pages.

I will be gone until Wednesday 25th and don't have knowledge at all on how the pageset can be created.
Easiest thing is probably just to modify the existing tp5 pages with the extra metadata eideticker needs. Going to try that route next.
So I had a look at the tp5 pageset. I really should have just done that a long time ago, as it would have saved me some effort in cleaning up/modifying the nytimes and Anyway,  there's lots of good stuff in there. No planet mozilla, though I imagine there's something that has similar behaviour in there. Should be very straightforward to modify most/all of these to work with Eideticker. Here's the manifest (should be fairly easy to figure out which pages are which from it)

What should we prioritize? My thought is that wikipedia, along with maybe a few more news sites, would be the most useful.
My vote is for: (more interesting testcase then (would be nice to get a real page instead of the log-in but not worth spending much time on this)
Added tests for reddit, imgur, and wikipedia:

Having some issues with imgur, there seems to be issues downloading all the images. Solution is probably to move mozhttpd into its own process. For now, you can see reddit and wikipedia here:
Filed dependent bug 747937 to deal with the image latency issue with imgur.
Depends on: 747937
No longer depends on: 745817
Depends on: 745817
Whiteboard: [eideticker:p1]
This bug seems a bit ambiguous to me. If there's a particular site we want to test with eideticker on a regular basis, IMO we should file a bug to handle that. If no one objects, I'm going to close this bug soon. (setting needinfo to myself to remind myself to do this)
Flags: needinfo?(wlachance)
Closing. Please feel free to file specific bugs if there's something you want to see!
Closed: 7 years ago
Flags: needinfo?(wlachance)
Resolution: --- → FIXED
Product: Testing → Testing Graveyard
You need to log in before you can comment on or make changes to this bug.