BUG: load_sample is not MPI-safe

[matthewturk]

We should decorate the appropriate function in load_sample so that it doesn't get run on all processors after enable_parallelism has been called.

[neutrinoceros]

In fact almost the whole function would require special treatment to be parallel-safe. The current higher order function (yt.func.only_on_root) doesn't feel as appropriate as I'd like (though the goal is achievable with it), because it doesn't make processes await the working process (root).

I think this strategy would be better expressed as a new context manager.

[chrishavlin]

I think this strategy would be better expressed as a new context manager.

does yt/utilities/parallel_tools/parallel_analysis_interface.parallel_root_only not work? that seems to dispatch a function call and block the other processes. and looks to be fine to use when enable_parallelism has not been called.

[chrishavlin]

answered my own question -- and the answer is sort of. That parallel_root_only won't broadcast a result so would need to modify it to optionally broadcast a result (instead of the all_clear integer) or just write a new function... but I do have something that seems to work. I'll push it up in a moment.