Performance regression Lite vs no extension

[heubergen]

I think I found a reproducible performance regression and wanted to know if you see anything from your side that you can fix or if I would have to report it to Mozilla.

Always tested with a fresh profile, only setting changed was "Enhanced Tracking Protection" to Strict (except for the first line). No extensions installed unless specified, no hard refresh:

Test Case	Load Event	DOMContentLoaded	Finish
Firefox fresh profile, no settings changed at all, no extension, first reload of page	4.23s	1.47s	18.5s
Firefox enabling strict tracking test 1	3.38s	1.49s	8.83s
Firefox enabling strict tracking test 2	4.22s	1.69s	9.5s
Ublock Lite basic protection, no filter list changes test 1	10.45s	1.51s	14.37s
Ublock Lite basic protection, no filter list changes test 2	9.47s	1.11s	14.02s
Ublock Lite No protection, permission for all websites revoked test 1	10.27s	1.55s	16.87s
Ublock Lite No protection, permission for all websites revoked test 2	11.20s	1.67s	17.03s
Ublock Lite disabled	2.92s	1.20s	8.22s
Firefox enabling strict tracking test 3	3.89s	1.58s	8.78s
Ublock Origin, no filter list or setting changes test 1	4.91s	1.38s	9.41s
Ublock Origin, no filter list or setting changes test 2	8.54s	1.34s	15s
Ublock Origin, no filter list or setting changes test 3	5.12s	1.47s	9.67s

Summary: DOMContentLoaded is not affected, Finish has some irregularities so I think Load Event is the best to use for judgement. It seems that Lite has the worst regression, while Origin only has a minor one.

I took the numbers from the network tab of the developer tools.

Tested URL: https://www.tripadvisor.ch/Restaurant_Review-g186338-d10460592-Reviews-Bonoo-London_England.html

Impact: While I only did the extensive testing with that specific URL (because I could feel the longest loading time on tripadvisor), I think that Firefox in general feels slower. But I wanted to give you specifics so I made this tests.

Firefox: 115.2.1esr
OS: macOS 12.6.9

[gorhill]

Please use the Firefox profiler to find out whether the issue is in uBO's code or elsewhere.

[heubergen]

I did check them, but can only see that it spends most of the time with __psynch_cvwait, not sure how one can check what the trigger is of that. I made two recordings and uploaded them:

With Lite: https://share.firefox.dev/3riWPKy
Without Lite: https://share.firefox.dev/3Rpmj3F

[gwarser]

Look at parent process. It's in Firefox DNR code.

[heubergen]

Ah right, so a bug in Firefox or does uBlock use the code incorrectly?

[gwarser]

AFAIK if you did not enable anything uBOL only delivers filter list and that's all. All work is done in browser.

[heubergen]

I've opened https://bugzilla.mozilla.org/show_bug.cgi?id=1853569, hopefully they confirm the issue.

Feel free to edit it if you have a better description/summary etc.

[gorhill]

I see at least two obvious performance issues in Firefox's code:

For each network request, it iterates through all the rules to find best match:
https://searchfox.org/mozilla-central/rev/0afa5b067589195a45ed5e858b51a4ee30c8c0a7/toolkit/components/extensions/ExtensionDNR.sys.mjs#1676

For some of the rules containing long list of domains, it potentially (and likely) iterates through all the domains to find a match:
https://searchfox.org/mozilla-central/rev/0afa5b067589195a45ed5e858b51a4ee30c8c0a7/toolkit/components/extensions/ExtensionDNR.sys.mjs#1792

For instance, there is one rule in default ruleset which requestDomains condition contains over 55.3K entries. So just for that one rule (out of over 17K rules), Firefox's code potentially loop over 55.3K entries to find a match against hostname for every single network request. Even without state of the art optimization, using a Set() to lookup matches would be far better as this would require only to loop at most as many times as there are labels in the hostname to match, i.e. single low digit number of iterations versus over 55.3K iterations.

---

It looks like Firefox's DNR implementation is a "minimal viable product" state, and no work has been done on it since April 2023.

[gorhill]

I am not seeing issues with uBO itself, so changing title. If you think there are performance issues with uBO, please provide supporting profiling data.

[gorhill]

Follow progress at https://bugzilla.mozilla.org/show_bug.cgi?id=1853569.