[Mpi3-rma] RMA proposal 2 initial text

Torsten Hoefler htor at illinois.edu
Tue Nov 2 14:03:38 CDT 2010


Hubert,

Thanks for the review. First let me explain our methodology. There are
two proposals (named "1" and "2"). Proposal 1, which contains all
essential changes, is maintained by Bill and me and proposal 2, which
contains possible extensions, is maintained by Pavan. 

Both can be found on the wiki
https://svn.mpi-forum.org/trac/mpi-forum-web/wiki/RmaWikiPage

So some of your comments refer to 1 and some to 2. Pavan forked proposal
2 form an earlier version of proposal 1. So I will reply to the subset
of your comments that effects proposal 1 :-).

On Tue, Nov 02, 2010 at 01:50:44PM +0100, Hubert Ritzdorf wrote:
> Some remarks to the actual document:
> 
> (*) Page 23: MPI_COMPARE_AND_SWAP and 
>     What does ``bitwise identical'' mean for a portable
>     MPI datatype such as C int which may have different
>     sizes in heterogeneous environments.
That is a very good comment. We had the datatype discussions before and
I forgot to remove the bitwise. We shall talk about it at the next
telecon.

> (*) Page 23: MPI_COMPARE_AND_SWAP and MPI_GET_ACCUMULATE
>     Which datatypes (and alignments) are fast in hw is
>     dependent on the target rank (and possibly window)
>     in non-homogeneous systems. I can't see, how this
>     can be handled by a portable program without query
>     functions by only checking the documentation.
Yes, I agree. The MPI_RMA_QUERY could be used to query this. We shall
discuss this at the next telecon.

> (*) Example 11.15, Process A:
>     I think, that there is missing a MPI_Win_membar after 
>     X = 2 before MPI_Barrier.
>     Correspondingly, I think that MPI_Win_membar is missing after 
>     assignments to X, Y, A in Examples 11.16 and 11.17.
That should still be correct. If process a issues a win_flush(A) which
itself only returns after the data arrived at the destination (public
window) and the data is only accesses through RMA calls (as it should be
in the examples), then it should be correct. A membar should be able to
achieve the same -- but might be less elegant.

Best,
  Torsten

-- 
 bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ -----
Torsten Hoefler         | Performance Modeling and Simulation Lead
Blue Waters Directorate | University of Illinois (UIUC)
1205 W Clark Street     | Urbana, IL, 61801
NCSA Building           | +01 (217) 244-7736



More information about the mpiwg-rma mailing list