Utilize PMIx type checking for PML check #12481
Conversation
It appears that sometimes we get a bogus value back for the PML selected by a remote proc. Unfortunately, the current code for exchanging that info removes all type information from the exchange. Replace the use of those macros with native PMIx calls so we can check that the data type being returned is the one we expected. True, it should not have changed - but this will help detect and debug any errors. Signed-off-by: Ralph Castain <[email protected]>
There was a problem hiding this comment.
This is not a solution to anything, especially not to PMIX returning incorrect values. It was helpful once, because we got lucky and the returned values had a different type, but that was a random event, not worth a patch that only fix one single instance of retrieving data from PMIX.
|
So...would it be better for the user to blindly segfault when there is an error, instead of detecting that something is wrong, reporting the error, and cleanly exiting? Remember, in the case in question, you were just being lucky - you were treating a blind object as a It isn't just a bug in PMIx that could cause the situation - could happen between two OMPI procs in different jobs, each running a different version. Where do you guarantee someone will always post the same type of data for a given key? |
|
This PR is just adding some protection. Someone else when they have time can go through the entire ompi code base and add similar protection. If this passes CI i''m merging. |
|
It appears that sometimes we get a bogus value back for the PML selected by a remote proc. Unfortunately, the current code for exchanging that info removes all type information from the exchange.
Replace the use of those macros with native PMIx
calls so we can check that the data type being
returned is the one we expected. True, it should not have changed - but this will help detect and debug any errors.