MPI_File_set_view datatype check

[ggouaillardet]

Following the issue initially reported on Stack Overflow i opened this issue on MPICH.

It turns out the MPI standard 3.1 chapter 13.3 requires

An etype (elementary datatype) is the unit of data access and positioning. It can be any MPI predefined or derived datatype. Derived etypes can be constructed by using any of the MPI datatype constructor routines, provided all resulting typemap displacements are non-negative and monotonically nondecreasing.

And we do not check such a thing in Open MPI.

What would be the best way to achieve that ?

• one option would be to write an adhoc check in MPI_File_set_view()
• an other option would be to add yet an other flag (OPAL_DATATYPE_FLAG_MONOTONIC ?) and update it when modifying datatypes.

@bosilca i can take a crack at it, but i would like you to give me a direction first.

FWIW

• ROM-IO hangs in and endless loop or crash in some cases.
• ompio seems fine with that (well, from v3.1.x) , but portable code should not rely on this (current) behavior.

@edgargabriel FYI

[bosilca]

For derived datatypes it seems expensive (and unnecessary) to compute this property for all datatypes. I would add 2 flags, one to mark the datatype as MONOTONIC, the other to mark the data as already ANALYZED for monotonicity (thus providing a cache mechanism for future checks). We are using 16 bits for the OPAL datatype flags, and there is enough room left to use 2 more.

In MPI_File_set_view if the ANALYZED flag is set, react accordingly to the MONOTONIC flag. Otherwise, one can use opal_convertor_raw to check for the monotonic property (an example of usage is in test/datatype/ddt_raw.c) and set the MONOTONIC flag accordingly, and always set the ANALYZED flag.