skip heartbeat thread from CPU profile

[d-netto]

PR Description

Intentionally skips the heartbeat thread from the CPU profiler, so that we don't include a large number of useless samples from semaphore_wait_trap.

X-ref: https://relationalai.atlassian.net/browse/RAI-31811.

Checklist

Requirements for merging:

• I have opened an issue or PR upstream on JuliaLang/julia: N/A.
• I have removed the port-to-* labels that don't apply.
• I have opened a PR on raicode to test these changes: https://github.com/RelationalAI/raicode/pull/22572.

[d-netto]

Reminder for self: test on Linux before merging.

[d-netto]

Ran using Profile, PProf; @profile peakflops(); pprof(webport=11111).

No semaphore_wait_trap on MacOS:

Profile collected on Linux also shows no abnormalities:

[vustef]

Not an expert, but excited about this.

[NHDaly]

So simple and clear! 👏 👏

Thanks for the screenshots! It's great to see this resolved. 😎
Can you also share a screenshot of running with more threads?

[d-netto]

Can you also share a screenshot of running with more threads?

Ran using Profile, PProf; @profile peakflops(); pprof(webport=11111). Got this with 8 threads on my M2:

Seems expected given most threads will be blocked on a condition variable with no tasks to run.

[NHDaly]

Perfect. Thanks! Ship it

[vustef]

Can you also share a screenshot of running with more threads?

Ran using Profile, PProf; @profile peakflops(); pprof(webport=11111). Got this with 8 threads on my M2:

Seems expected given most threads will be blocked on a condition variable with no tasks to run.

What if all threads are busy?

[d-netto]

What if all threads are busy?

Still a large number of samples on cvwait, though the fraction seems to depend on how many GC threads are used.

Workload

using Profile
using PProf
using Base.Threads
    

function main()
    nt = Threads.nthreads()
    Base.@sync begin
        for i in 1:nt
            Threads.@spawn begin
                peakflops()
            end
        end
    end
end

@profile main()

8 compute threads, 1 GC thread.

8 compute threads, 8 GC threads

[vustef]

What if all threads are busy?

Still a large number of samples on cvwait, though the fraction seems to depend on how many GC threads are used.

Workload

using Profile
using PProf
using Base.Threads
    

function main()
    nt = Threads.nthreads()
    Base.@sync begin
        for i in 1:nt
            Threads.@spawn begin
                peakflops()
            end
        end
    end
end

@profile main()

8 compute threads, 1 GC thread.

## 8 compute threads, 8 GC threads

Thanks! Do you know why is the fraction not proportionate to the num default threads vs GC thread ratio? There are 8x more default threads, so I'd expect non-useful profile to be only 1/9th of the total.