Multiple Request Completion 2

[jtronge]

Hello, I've been working with @hppritcha to attempt adding multiple request completion support. This is partly based on @AndrewGaspar's original PR #36 but is slightly different. Based on some of the discussion in #29, I experimented with attaching references to each request as a way to allow the user to immediately perform some computation on the data as soon as the request has been waited on.

Using the modified request object as a basis I added a RequestCollection object, similar to the former PR, but including support for attached references. Requests can be added to the collection and then waited on using the wait_() or test_() methods. To ensure that requests have completed, I included a new multiple_scope() function which passes in both a reference to a scope object and a request collection.

[hppritcha]

could someone kick off the CI for this PR @bsteinb @jedbrown thanks.

[jedbrown]

Thanks for working on this. I see some cargo fmt noise and have some questions about how to expose the completions.

[jtronge]

Thanks for taking a look at this. I'll fix the lints and take a closer look at the wait_some/test_some code at the start of next week.

[jedbrown]

I'd like to release version 0.6 soon so other crates can publish. Do you think this should be part of 0.6 and if so, how long do you anticipate needing?

[jedbrown]

Heads up that there are some new clippy warnings. This is clear on main.

$ cargo clippy -- --allow clippy::missing_safety_doc

[jtronge]

Ok, thanks. I think I managed to fix all of the warnings. I've updated the code and examples to use &mut Vec.

[jedbrown]

Looks great, just a couple comments to consider. And I'll run a pipeline.

[jedbrown]

I don't think this is from your branch (it passed on re-running), but do you understand the intermittent failure here? It looks like perhaps std::panic::catch_unwind not working correctly.

https://github.com/rsmpi/rsmpi/runs/7658512360?check_suite_focus=true#step:8:120

[jedbrown]

Many thanks for this important contribution. I'll merge now.

[jtronge]

I don't think this is from your branch (it passed on re-running), but do you understand the intermittent failure here? It looks like perhaps std::panic::catch_unwind not working correctly.

https://github.com/rsmpi/rsmpi/runs/7658512360?check_suite_focus=true#step:8:120

Hmmn, I'm not sure. I've run into a couple strange errors like this when Rust panics, or the c code that rust is calling aborts. It looks like openmpi is failing here and I'm not sure where the Rust code is handling this, or if it even is.

[jedbrown]

There is output from where Rust panics (that's supposed to be caught), but the unlink error is quite strange. It'll be something to watch, especially if we find a way to reproduce it.

[jtronge]

Talking with @hppritcha, he pointed me to this issue: open-mpi/ompi#7393. Seems like we might want to set TMPDIR for macOS at least. He also suggested adding a sleep between tests.

[jedbrown]

Thanks, I haven't seen this failure since #125.

I've been seeing non-deterministic MSMPI failures recently, but I don't think they're related to recent changes to rsmpi.