892768 - [Meta] Improve breakpad unwinder performance

Status: RESOLVED WONTFIX

(Core :: Gecko Profiler, defect)

Description

[Benoit Girard (:BenWa)]

This bug should track possibles improvements to breakpad performance

jld collected a profile with breakpad execution:
https://people.mozilla.com/~bgirard/cleopatra/#report=2af1780356d123e3e2c91102441e2e48c0d95611&search=breakpad

Comment 1

[Jed Davis [:jld] ⟨⏰|UTC-7⟩ ⟦he/him⟧]

I've applied the proposed fixes for bug 894264 and bug 892774 and done some testing.

Here's a perf_event profile of me typing nonsense letters into the Messages app, with breakpad unwinding set to 10ms, on a keon:

https://people.mozilla.com/~bgirard/cleopatra/#report=a914daa692ec402403f401e3ed96f4febc1378ca

The main process's unwinder thread takes 25.2% of the CPU, and generates this:

https://people.mozilla.com/~bgirard/cleopatra/#report=084c24d3fb6f2e366233c1c8813217a18b6dfd22

So, with about 14.3ms average interval in the region that roughly corresponds to the meta-profile, that's around 3.6ms per sample.  This workload involves relatively large C++ stacks and not much script.

Comment 2

[Jed Davis [:jld] ⟨⏰|UTC-7⟩ ⟦he/him⟧]

And here's more sunspidering (which, in constrast, should spend a lot of time with jitcode on the stack).  The perf_event profile:

https://people.mozilla.com/~bgirard/cleopatra/#report=dd56120921a0037176932da98e9f8714401f414d

The SPS profile of the browser app:

https://people.mozilla.com/~bgirard/cleopatra/#report=2855f38d65a3faec0fe6e5e52d7110d20131cfd4

26.4% of 11.8ms = about 3.1ms

Comment 3

[(not currently active) Ted Mielczarek]

The Breakpad unwinder is deprecated in favor of LUL, so I say we close this.