Rename Name-Shifted Profiling Interface (PMPI) to Interception Interface

[wesbland]

Problem

As part of developing the proposal for QMPI, we needed to differentiate between the existing interface and the newly proposed interface. This pointed out the strange existing name.

PMPI is used for more than just profiling and it would be best if the name reflected that.

Proposal

Replace “name-shifted profiling interface”, “profiling interface”, etc. (or similar text) anywhere it is used to say “name-shifted interception interface”.

Changes to the Text

https://github.com/mpi-forum/mpi-standard/pull/724

Impact on Implementations

None

Impact on Users

Might break PDF searches

References and Pull Requests

[mahermanns]

I wonder whether we should also think about a new prefix IMPI_ instead of QMPI_. I know the Q as the "next after P" has a long internal tradition, yet it may not be an intuitive prefix for user. However, may we have potential name conflicts with Intel MPI internal naming?

[wesbland]

I think that's a fair point (though it's probably part of https://github.com/mpiwg-tools/mpi-standard/pull/10 😄).

I haven't thought through it that much, but IMO, FMPI might make more sense (function pointer interception) and wouldn't cause confusion with IMPI. While not part of the standard, it's been around for a long time and has a lot of relatively famous uses of the prefix. It's kinda the same as us not wanting to use OMPI to me.

[devreal]

I am not sure I like the idea of renaming something that has been in the standard for decades only because there is now an alternative mechanism for achieving the same goal. The term has been used in countless books, publications, and presentations. And renaming the PMPI prefix would break far too many codes. Granted, "profiling" was the wrong term to begin with but we have to live with what we have I guess.

If, for some reason, the Forum wants to go down that road I suggest using "interposition" instead of "interception". For one, "library interposition" is an established term in computer science and the PMPI interface enables that. Second, it has at least a P in it to match the P in PMPI. And third, it sounds less bellicose (which arguably is a highly subjective argument).

[schulzm]

Marc-Andre can correct me, if I am wrong, but I think two things collided here in the discussion: this ticket is intended to clarify the usage of the existing PMPI API, for which profiling has been increasingly confusing due to its usage. Actual profiling is only one of many more cases, as we use the wrapping/interception much more generically these days. Nevertheless, it is of course a term that has been used for very long (even it is incorrect).

However, the ticket does not change the naming - PMPI stays PMPI, as this would indeed break codes. That discussion is about choosing the right prefix for the new interface: the current name QMPI is only a placeholder and not intuitive at all - for this IMPI or FMPI could make sense, but that should probably be taken to a new threads.

[devreal]

@schulzm You're right, I misread the discuss about changing prefixes. That discussion should indeed go into a different thread to avoid further confusion (esp since QMPI is not part of MPI yet). The rest of my arguments still stand though :)

[mahermanns]

@devreal Yes, no actual API is changed, just how we name and refer to it. However, as @wesbland already noted, the users knowing the term "profiling interface" would have to get used to looking up "interception interface" in the index. Whether that is too big of a hassle, and how we might mitigate such problems should be discussed broadly in the forum, I guess.

The discussion of the new prefix, as you realized, is only about the so-called QMPI interface, which will have a different set of functions to query and register the callback functions for the tools. And the idea here is to come up with an intuitive prefix.