[new release] moonpool (0.4)

[c-cube]

Pools of threads supported by a pool of domains

• Project page: https://github.com/c-cube/moonpool

CHANGES:

• add Fut.{reify_error,bind_reify_error}

• full lifecycle for worker domains, where a domain
  will shutdown if no thread runs on it, after a
  short delay.

• fix: generalize type of create_arg

• perf: in Bb_queue, only signal condition on push if queue was empty

[c-cube]

So that's going to be interesting to debug. What machine does this run on?

[avsm]

How many domains are the tests trying to create? It'll be running on a many core x86 machine in that first test, but may be running into cgroups resource limits.

[c-cube]

Potentially n-1, where n is recommended domain count. But this time it will create many more domains than that over the life time of the test, it just has at most n-1 alive at a time.

[c-cube]

So this is quite puzzling. I tried on a c3.4xlarge instance and a friend tried on his M1, to no avail — the test always passes. Can it be the cgroup settings?

[avsm]

It's very likely to do with cgroups restrictions in opam-repo-ci. /cc @mtelvers to perhaps shed light on this.

[mtelvers]

This has been a real puzzle for me this afternoon, taking a long look a cgroup limits on the cluster workers, running the spec file on different workers, and getting some expert advice from @dra27. The test fails when nproc >= Max_domains. The machines where it fails include: cree, pima and marpe which have nproc of 128, 128 and 256, respectively. On smaller machines where nproc < 127, there is no issue. This has also been tested as a bare metal installation on pima (no Docker or OBuilder) and the results are the same.

[dra27]

I'm not sure where the bug lies, but the domains are never being joined. I was able to reproduce the failures very easily with a custom compiler with Max_domains at 16 on the same machine (that triggered even more test failures with the same exception). I hacked together some domain joining logic (the hack was to record the Domain.t with the domain state and always join on it before re-using the slot). With that version, I'm not seeing the crash.

[c-cube]

Thank you @mtelvers and @dra27 for the thorough investigation and the fix!!

[avsm]

Such a success disaster, that we have machines with so many cores that we're blowing past the maximum available domains in OCaml already ;-) Thank you for the thorough investigation @mtelvers and @dra27, and @c-cube for the revised package!

[c-cube]

Thank y'all again for the help! That was a tricky one :)