[Mpi3-rma] RMA WG discussion

[Jeff Hammond]

> I have a further question to MPI_NO_OP and MPI_{R}Get_accumulate.
> I assume that origin_count is not significant for MPI_NO_OP
> (such as origin_datatype)

Like origin_addr, origin_count and origin_datatype should be ignored
when MPI_NO_OP is used.  I'm not sure that the text needs to be
changed for this though.  Pavan said that there is some text that
asserts that buffers (i.e. pointers) be valid (i.e. not NULL) unless
otherwise stipulated.  Is there text to the same effect for count and
datatype?  If one uses count=0 in MPI_SEND, does the datatype have to
be valid?  If not, then we can ignore that argument at least.  Since
we don't have precedent for MPI_NO_OP, it seems that the notion that
count can be ignored is implicitly obvious but nonetheless should be
stated (or we could require count=0 for MPI_NO_OP...).

> Which count variable defines the number of basic elements to be returned ?
>   (result_count, result_type) or (target_count, target_type) ?

result_addr is the buffer that is written at the origin so I believe
the answer is (result_count, result_type).

> Should an error code be returned if the other number is too small ?

MPI-3 11.3.2 (MPI_GET_ ends with "The target buffer must be contained
within the target window or within attached memory in a dynamic
window, and the copied data must fit, without truncation, in the
origin buffer."  MPI-3 11.3.4 (MPI_ACCUMULATE) says something similar.

I would contend that the same behavior applies to
MPI_(R)GET_ACCUMULATE, but it might be appropriate to add text to this
effect, just to be clear.  We need to say slightly more than for
MPI_GET though, because there are three (buffer, count, datatype) that
need to be compatible.

If we combine the restrictions for GET and ACCUMULATE,
result_size<=target_size and origin_size<=target_size - size is
derived from (count, datatype) - but I don't know that there has to be
any restriction relating result_size and origin_size that isn't
included in the former.

Best,

Jeff


-- 
Jeff Hammond
Argonne Leadership Computing Facility
University of Chicago Computation Institute
jhammond at alcf.anl.gov / (630) 252-5381
http://www.linkedin.com/in/jeffhammond
https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond