Add GitHub Action workflow to run benchmarks on CI for pull requests

[giordano]

No description provided.

[giordano]

It works: https://github.com/Team-RADDISH/ParticleDA.jl/blob/ceec818fcd6ea3b07c6f9e5abfb63d68fd0ab582/2021/02/22/164051/result.md 🥳 As I suspected, benchmark/benchmarks.jl should be already committed to master for this to work (if I understand correctly that's because it tries to compare performance with master, so the file needs to be there already).

[giordano]

Of the benchmark results published yesterday, only the first and the last seems meaningful (no performance change between the baseline on master and this PR), while the other two report a 50% speedup (with no changes to the code...).

I'm not sure what's going on, I've run the BenchmarkCI code locally multiple times and always found consistent results between the target and the baseline. I suspect that running benchmark on GitHub-hosted machines may not be the best idea possible (perhaps machine are randomly hit by other load?). In the past I tried to run benchmarks on Travis and results were unreliable at best, but "totally meaningless" is a better depiction of my experience, performance was extremely poor overall. Azure/GitHub machines are better than Travis, but may still be not always reliable.

My suggestion would be to merge this, but take results with a grain of salt.

[giordano]

Not sure this explains the issue (especially, not the difference between target and baseline), but the two "good runs" were using Intel(R) Xeon(R) Platinum 8171M CPU @ 2.60GHz, while the other two were on Intel(R) Xeon(R) CPU E5-2673 v3 @ 2.40GHz.