FYI, we are discussing allowing the user to ask for a "required" value for the memory model, in MPI-3.1. The idea is that if you are only going to do PUT/GET/ACC accesses and not direct load/store, there's no semantic difference between UNIFIED and SEPARATE. In that case, if the user specifies that "required" is only SEPARATE, the implementation can have a bit more flexibility in choosing UNIFIED vs. SEPARATE. For MPI-3, ch3 will return UNIFIED. As Jeff mentioned, there was a bug in mpich-3.0 to mpich-3.0.4, so we were returning SEPARATE. That's fixed in the git master. In the next release, we'll correctly return UNIFIED. -- Pavan On 07/13/2013 11:05 AM, Jeff Hammond wrote:
UNIFIED means that the public and private window are the same. SEPARATE, which is equivalent to MPI-2 RMA behavior, requires explicit synchronization of the public and private windows, e.g. with MPI_WIN_SYNC. If you are in the unified model, SYNC is a no-op.
You cannot choose this because UNIFIED requires the hardware to provide something that is not always present. There is little to no value in asking for SEPARATE if UNIFIED is available. I can imagine scenarios where it might be useful but they are mostly contrived.
MPICH's ch3 device is going to provide UNIFIED, btw, at least in the trunk. Pavan committed that fix a few days ago.
Jeff
On Sat, Jul 13, 2013 at 10:57 AM, Ronny Brendel <[email protected]> wrote:
Hey,
Can I choose between the two memory models, or is this provided by the mpi implementation or underlying hardware?
If I can choose: How can I? If I cannot choose: Why is that?
I tried using MPI_Win_set_attr to set it, but Mpich throws an error that the key is invalid.
(I am using mpich 3.0.4)
Reading the standard I couldn't find out wether this is supposed to be choosable by the user. I just assumed it is. Because, if I don't query MPI_WIN_MODEL, there is a chance my program works on one but not on the other. So I have a correct program depending on which model there is and the code looks exactly the same. Which is behavior I would not expect.
I hope you can help, cheers, Ronny _______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss
-- Pavan Balaji http://www.mcs.anl.gov/~balaji