[WIP] Big overhaul

[carstenbauer]

This is a WIP PR to improve the user-facing interface (and documetation) of, primarily, LIKWID.PerfMon.

Code

• one-based indexing
• provide simplified interface
  • perfmon
    • multiple perf groups
  • @perfmon
    • keyword args for autopin etc.?
    • multiple perf groups
• CI
  • check / overhaul / fix examples
    • CPU
    • GPU (and CPU+GPU)
  • setup CI on N2

Docs

• clarify examples wrt indices (not cpu ids, one-based)
• multiple group example
• multithreading example

TODO before merge

• tests for @perfmon_marker / perfmon_marker
• docs
  • howto perfmon
• @nvmon / nvmon
  • tests
• @nvmon_marker / nvmon_marker
• Update example in README.md

[carstenbauer]

Preview of the new simplified API

Simple single-threaded example:

using LIKWID
using LinearAlgebra

const N = 10_000
const a = 3.141
const x = rand(N)
const y = rand(N)
const z = zeros(N)

metrics, events = @perfmon "FLOPS_DP" begin
    for i in eachindex(x, y)
        z[i] = a * x[i] * y[i]
    end
end

@show events["RETIRED_SSE_AVX_FLOPS_ALL"];
@show metrics["DP [MFLOP/s]"];

Output

events["RETIRED_SSE_AVX_FLOPS_ALL"] = 20000.0                                                                       
metrics["DP [MFLOP/s]"] = 1363.3126039898723

Simple multi-threaded example:

using LIKWID
using LinearAlgebra

const N = 10_000
const a = 3.141
const x = rand(N)
const y = rand(N)
const z = zeros(N)

metrics, events = @perfmon "FLOPS_DP" begin
    Threads.@threads for i in eachindex(x, y)
        z[i] = a * x[i] * y[i]
    end
end

@show getindex.(events, "RETIRED_SSE_AVX_FLOPS_ALL"); # vector of events corresponding to Julia threads
@show getindex.(metrics, "DP [MFLOP/s]");

Output:

getindex.(events, "RETIRED_SSE_AVX_FLOPS_ALL") = [5954.0, 5000.0, 5000.0, 5000.0]
getindex.(metrics, "DP [MFLOP/s]") = [0.3990529460355565, 0.3351133238457814, 0.3351133238457814, 0.3351133238457814]

Note that Julia threads automatically get pinned to (the cpu threads they are currently running on) to get reliable results. The function equivalent perform(f, groupname; cpuids, autopin) provides more options.

(cc @antoine-levitt, @TomTheBear, @vchuravy)

[antoine-levitt]

That's very nice, thanks for doing this, will test once merged!

[TomTheBear]

There is a driver mismatch on the one CI node:

xxxxx@medusa ~ $ nvidia-smi 
Failed to initialize NVML: Driver/library version mismatch

This needs to be fixed by our admins.

[carstenbauer]

Thanks for the info. I'll also set up CI on Noctua 2 soon so that we test on two different systems, one with the daemon and one with the perf_events backend.

[carstenbauer]

docs preview: https://juliaperf.github.io/LIKWID.jl/previews/PR31

[vchuravy]

This is looking great!

[carstenbauer]

Will try to finish things next week before I record my JuliaCon talk 😊

[carstenbauer]

Let's merge this (I'll fix CI / doc build on main because it's easier). Remaining things like @nvmon_marker will have to wait.

[carstenbauer]

@antoine-levitt Feel free to try out the main branch now. Would be great to have someone test it :)

[antoine-levitt]

Hm, I wanted to try this for a course I was giving but it's over now and I don't want to go back to it :-p but I'll definitely have a use for it at some point, it looks great!