Firefox 71 Linux Performance Isn't Looking All That Great

[Michael]

Michael, is it possible to hook up Phoronix Test Suite with mozregression to locate where Firefox performance has regressed (like you have for Linux)? I am certain that the Linux community would greatly appreciate your help in finding these issues.

If I magically had the time/sponsorship.

[royce]

Mozilla is spending a lot of engineering resource in rolling out webrender to more platform combinations (OS + GPU combos) and not so much on improving performance. Hopefully after it's 100% rolled out they can get back into optimising firefox again.

[treba]

Mozilla is spending a lot of engineering resource in rolling out webrender to more platform combinations (OS + GPU combos) and not so much on improving performance. Hopefully after it's 100% rolled out they can get back into optimising firefox again.

Rolling out Webrender on all platforms is *the* big performance optimization in the baking
This is good news for Linux users as the Webrender code is mostly shared between the platforms, as opposed to the previous situation where Win, MacOS and Linux all used different backends (Direct2D, OpenGL, Basic/Software)

[Thaodan]

Why didn't you build Firefox with PGO/LTO like Chromium?

[royce]

Rolling out Webrender on all platforms is *the* big performance optimization in the baking

It is the beginning of it, yes. Webrender's performance in general today is on par with what it is replacing, and in some cases somewhat better. It's always been Mozilla's intention to get it rolled out first, then actually start reaping the benefits of a better, easier to optimise codebase.

[Michael]

Why didn't you build Firefox with PGO/LTO like Chromium?

Using Firefox release binaries, just as release binaries were used for Chrome.

[Weasel]

Must be because they wrote more stuff in Rust.

[kozman]

Problem with Mozilla, there putting in unwanted Features into firefox which is BLOATING the Browser right up . , disable features you dont need/want

Adding features, yes. Bloating? Not quite true if you're looking at installer binary size. 72 beta seems to be decreasing in size in relation to 71 and 70. Could be related to switching from LLVM 8 to 9 though. Must be some perf regressions not caught in 71.0. In most real world use cases, the average user won't notice.

FF 69.0 Win x64 binary is 48,380KB
FF 70.0 Win x64 binary is 50,309KB
FF 71.0 Win x64 binary is 50,119KB
FF 72.0b2 Win x64 binary is 49,814KB

[kozman]

It is the beginning of it, yes. Webrender's performance in general today is on par with what it is replacing, and in some cases somewhat better. It's always been Mozilla's intention to get it rolled out first, then actually start reaping the benefits of a better, easier to optimise codebase.

Exactly. And each FF rev gets Webrendering working on less powerful GPUs whereas initially it was targeting fairly recent hardware. Some old cruft/code has been removed from 71 and 72 is likely where these lower benchmark numbers will bump up a bit. I'm on a pretty old Latitude E6510 and 72 beta is feeling a bit snappier than 71 betas and final did. YMMV.

[kozman]

Must be because they wrote more stuff in Rust.

Possibly. Rust is evolving with each new version but if I remember correctly, they weren't able to bump to using Rust 1.38 and take advantage of it's new features for 71.0 due to something with 68.x ESR not building with it. Could be the reason perf number are lower than 70 or maybe just a regression somewhere. Rust 1.38 only just got into 72 beta and we should Rust 1.39 be in 73.0.