bench: add GC stats to bench

[Alizter]

Problem

We want to have a convenient way of tracking memory gains in PRs. There is no way to do this currently apart from measuring Dune's perf by yourself.

Solution

We start collecting Gc.stats in each of our build runs in the bench. In order to do so we introduce a Metrics module which is sufficiently polymorphic to support an unzipping operation which is useful when serialising to json. The mli of the Metrics module has some detailed documentation about the metrics that we collect about a process.

During the build runs, we garbage collect and compacitify before observing the statistics.

This PR changes the names of the benchmark jobs to be more uniform (use title case). This has the side effect of discontinuing previous runs, although the old data is still available in the website. Due to this I have bumped the version number of the bench.

We also enable sandboxing in the bench.

• wait for feature: add --dump-gc-stats option to dune build #8072
• Update hacking.rst bench documentation
• Remove unwanted metrics?

[Alizter]

I'm now testing this with #8052 to see any memory change.

Edit: As expected, there is a decrease of 200 in the minor word count.

[Alizter]

Now I will repush the bench commit and then push #8035.

Edit: No observable change in this one.

[Alizter]

This time I pushed a dummy commit allocating a pointless hashtable in a hotpath to see if it would be picked up and it appears to have done.

[rgrinberg]

You're handling the GC in the wrong process here. It's the dune command above that should spit out its GC stats.

To do this, you could add some --dump-gc-stats-end=FILE argument that would add an at_exit hook to dump this info.

[Alizter]

I've added that option in #8072 (which is currently also included here). The data I'm getting now makes much more sense. I'll try to test out your PRs again.

[Alizter]

OK this time testing #8052 again.

cc @rgrinberg

So the only observable changes are:

Major collections

On clean build went from 30 to 37 and on null build decreased from 23 to 22.

Heap words

On null build went from 2.039e+7 to 1.792e+7 but nothing changed on clean.

Free words

On null build free words appear to have halved from 5.126e+6 to 2.687e+6.

Conclusion

Overall, I'm not really sure if I understand what this change does to the memory.

[Alizter]

This time trying #8035.

No changes observed really.

[rgrinberg]

Awesome. Great work.