[Mpi-forum] MPI International Survey

Atsushi HORI ahori at riken.jp
Sun Dec 9 18:17:27 CST 2018


Thank you, Jeff and Bill,

I believe this survey topic should be discussed out of this mailing list.  So, I will create another mailing list to discuss on this.  The mailing list will include the members who are not involved in MPI Forum.  So, if you want to join the mailing list, just let me know.  If you want to argue on this point, you may continue discussing on this mailing list. 


Now I am going to show my idea on what Jeff and Bill talked about.

I have noticed that the question in the survey is vague.  Let me clarify things on the term "MPI+X."

I think the operator '+' here is vague and not appropriate.  In many cases 'X' means OpenMP (CUDA or something).  However this means that OpenMP (or CUDA, for example) provides the complemental programming model to MPI and basically the programming models of MPI and "X" are orthogonal.  So, with the mathematical sense, it should be noted as "MPI * (times) X" instead of "MPI+X" simply because MPI and X are different things and we cannot add them.  The '*' symbol may also mean outer product of them and its spanning space gets bigger, and thus it gets harder for users to program with the combination.

Contrastingly, what Jeff talked about is "Y over (on top of) MPI."  We may express it as "Y/MPI" which comes of the English expression of "over."  In this case, "Y" is implemented with using MPI to provide something which MPI does not provide.  For example, many users use pnetCDF or HDF5, implemented with using MPI-IO, instead of using MPI-IO directly (this is one of the reasons why I did not include any MPI-IO related questions).

The intent of the original survey question is to ask if users are willing to move from using MPI to using some other thing.  In the draft survey I asked you to answer already has the question on "MPI+X" and I think we should have another question to ask on "Y/MPI" instead of "Z or MPI."

Finally I thank Jeff and Bill for letting me know on this issue. 

-----
Atsushi HORI
ahori at riken.jp
https://www.sys.r-ccs.riken.jp






More information about the mpi-forum mailing list