Address Github issue #11532 by translating legacy parameters for direct launches

[qkoziol]

Borrow code from the OMPI schizo module in PRRTE that translates legacy MCA parameters when an application is direct launched (PRRTE will translate legacy parameters when natively launched).

Addresses issue #11532

[qkoziol]

Requires PRRTE PR: openpmix/prrte#1777 and OpenPMIX PR: openpmix/openpmix#3133

[gpaulsen]

bot:ibm:retest

[rhc54]

Hmmmm....why would an MPI process care about PRRTE params? It has no involvement with PRRTE, so the params will just be ignored.

You certainly would want the PMIx values as the MPI proc uses that library - but I cannot see any reason to handle PRRTE params.

[qkoziol]

Hmmmm....why would an MPI process care about PRRTE params? It has no involvement with PRRTE, so the params will just be ignored.

You certainly would want the PMIx values as the MPI proc uses that library - but I cannot see any reason to handle PRRTE params.

OK - I had a feeling that might be the case. I'll update this PR to just look for PMIX params and remove the PRRTE part. I'll also update the PR for PRRTE to just have the refactor of the schizo ompi file and remove the other work that creates the PRRTE framework header.

[qkoziol]

Revised to just check PMIX params, and to use the existing PMIX framework header.

[qkoziol]

bot:nvidia:retest

[lrbison]

bot:nvidia:retest

[lrbison]

CI cannot find the file:

base/pmix_base_fns.c:43:10: fatal error: pmix/src/include/pmix_frameworks.h: No such file or directory
   43 | #include "pmix/src/include/pmix_frameworks.h"
      |          ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
compilation terminated.

I saw in openpmix/openpmix#3133 we expected it to be in ${PREFIX}/include/pmix/src/include. I'm not sure what the disconnect is here, do we search PREFIX?

[qkoziol]

CI cannot find the file:

base/pmix_base_fns.c:43:10: fatal error: pmix/src/include/pmix_frameworks.h: No such file or directory
   43 | #include "pmix/src/include/pmix_frameworks.h"
      |          ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
compilation terminated.

I saw in openpmix/openpmix#3133 we expected it to be in ${PREFIX}/include/pmix/src/include. I'm not sure what the disconnect is here, do we search PREFIX?

🤔 It's installed when I compile against the latest OpenPMIX. I'll try against the internal PMIX submodule and see what I find. We might need to bump the submodule pointer?

[lrbison]

It looks like pmix_frameworks.h should be generated by autogen in openpmix.

https://github.com/openpmix/openpmix/blob/9b09a2d3ee67d2dd1009dd732a629247a7a23087/autogen.pl#L292

That link should correspond to the pinned version on ompi main, in fact we are currently on the commit that introduced that change (openpmix/openpmix@22fe51c)

[rhc54]

43 | #include "pmix/src/include/pmix_frameworks.h"

You are probably hitting an issue between when you build OMPI against an internal vs external version of PMIx. The external reference should be just src/include/pmix_frameworks.h. The internal reference would be pmix/src/include/pmix_frameworks.h unless you are adding a -I option that takes care of the pmix/ prefix for you.

[rhc54]

Yeah, I confirmed that the problem is in the include statement - it should be:

#include "src/include/pmix_frameworks.h"

Your configure code already adds the necessary -I flag to point you at the correct place. It was working for @qkoziol solely because there is another -I that points to the OMPI include location that allowed the pmix/src/include path to be resolved. Doesn't work if you use an external PMIx installation.

[awlauria]

@rhc54 could you review this ?

[janjust]

@qkoziol can you open up the v5.0 cherrypick

[qkoziol]

@qkoziol can you open up the v5.0 cherrypick

#11921