get_best_solver is costly

[redianthus]

It takes 3% of the whole execution time on test-comp with a 30 seconds timeout.

[hra687261]

Using strings as keys for the map of features is probably part of the problem. I'll change that soon.

[hra687261]

Did not mean to close this ...

[hra687261]

@redianthus do you have a simple way to reproduce the measure? are there command line options to pass to owi or symbocalypse to get these percentages or a test-comp run?

[hra687261]

ping @redianthus

[redianthus]

Oh sorry for the delay. No, it's not easy to reproduce. Basically, you need to run Test-Comp with perf and look at the flamegraph. You would start by recording with (you need to use a compiler with frame pointers enabled, and preferably flambda1 too) :

$ perf record -F 99 --call-graph fp,2000 -- dune exec -- symbocalypse testcomp owi 30 --solver=smtzilla --exploration=rarity-depth-aging

Then, here, in a normal situation you would go:

$ perf script report flamegraph

But it does not merge the various owi processes that are spawned by symbocalypse and it is thus unusable.
My first try was to use Brendan's scripts here: https://github.com/brendangregg/FlameGraph
And instead do:

$ perf script > out.perf
$ perl FlameGraph/stackcollapse-perf.pl out.perf | perl FlameGraph/flamegraph.pl > out.svg

This is very slow but it ultimately works (out.perf is 77Go...). So instead, you should open the recorded trace in hotspot and use the flamegraph visualization mode. The loading is slow but then it works well enough for me.