Avoid shared memory bug in openmpi>=3.1.2

[leofang]

Currently we build v3.1.2 in nsls2-tag:

lightsource2-recipes/recipes-tag/openmpi/meta.yaml

Line 2 in 87723db

{% set version = "3.1.2" %}

However, this version seems to be buggy. If one spawns a few MPI processes, let them do some work, but terminate them abnormally (Ctrl-C and whatnot), it can be seen that in /dev/shm/ there will be shared memory segments related to openmpi's vader component (don't ask me what this is...) that are not unlinked by openmpi during the cleanup phase:

leofang@xf03id-srv5:~$ ls -lt /dev/shm | more
total 28464
-rw------- 1 leofang leofang 4194312 Jan 25 23:06 vader_segment.xf03id-srv5.73dc0001.2
-rw------- 1 leofang leofang 4194312 Jan 25 23:06 vader_segment.xf03id-srv5.73dc0001.3
-rw------- 1 leofang leofang 4194312 Jan 25 23:06 vader_segment.xf03id-srv5.73dc0001.1
-rw------- 1 leofang leofang 4194312 Jan 25 23:06 vader_segment.xf03id-srv5.73dc0001.0
-rw------- 1 xjhuang xjhuang 4194312 Jan 15 11:08 vader_segment.xf03id-srv5.32270001.2
-rw------- 1 xjhuang xjhuang 4194312 Jan 15 11:08 vader_segment.xf03id-srv5.32270001.0
-rw------- 1 xjhuang xjhuang 4194312 Jan 15 11:08 vader_segment.xf03id-srv5.32270001.3
-rw------- 1 xjhuang xjhuang 4194312 Jan 15 11:08 vader_segment.xf03id-srv5.32270001.1
-rw------- 1 xjhuang xjhuang 4194312 Jan 15 11:07 vader_segment.xf03id-srv5.358e0001.3
-rw------- 1 xjhuang xjhuang 4194312 Jan 15 11:07 vader_segment.xf03id-srv5.358e0001.2
......

and they will remain there until the system is reboot, eating up slowly the system's memory!

Based on openmpi's changelog (see here), it seems that vader was reworked in v3.1.2, presumably this bug sneaked in by then. There's a bug fix in v3.1.3 hopefully would address this; if not, we can downgrade to v3.1.1, which I tested and worked without this issue.

(UPDATE: v3.1.3 also has this problem, has to use 3.1.1...)

I have other questions related to building conda packages for mpi4py, openmpi, and mpich, but perhaps I should ask them somewhere else...

[tacaswell]

@leofang can you sync our recipe with the conda-forge one (which sadly for now is a copy-paste job 😞 )

[leofang]

Sure, happy to do. But let me first check with Open MPI people and see if this is a known bug or caused by other problems.

[leofang]

@tacaswell since you brought up, I'd also like to copy and paste the conda-forge recipes for mpi4py and mpich. This would solve the MPI nuisance. Is it OK?

[stuartcampbell]

@leofang Definately - go ahead

[leofang]

ref: open-mpi/ompi/issues/6322