[Mpi3-rma] current version of proposal
thakur at mcs.anl.gov
Sat Mar 5 10:48:46 CST 2011
There is a sentence at 54:30 that mentions "same operation". Maybe it needs to be updated.
"The outcome of concurrent accumulate[s] operations to the same location, with the same operation and predefined datatype, is as if the accumulates where done at that location in some serial order."
On Mar 5, 2011, at 10:14 AM, James Dinan wrote:
> On 03/04/2011 11:35 PM, Rajeev Thakur wrote:
>>>> 47:28 - why is "that use the same operation" being removed. It
>>>> should be there.
>>> no, this is now in the new info key. I added a sentence clarifying
>>> this (and referencing back).
>> OK. With that deletion, the chapter doesn't say anywhere whether
>> concurrent overlapping accumulates with different operators are
>> allowed or not. Or are they allowed only if one of the two operations
>> is a no_op?
>> And what is the default if the user doesn't pass any info key
>> accumulate_ops? What should the implementation assume?
> The new default for accumulate ordering is strict ordering. I think
> that as long as the same datatype is used, the result of concurrent
> conflicting accumulate operations is defined as some atomic interleaving
> at the datatype-level.
> mpi3-rma mailing list
> mpi3-rma at lists.mpi-forum.org
More information about the mpiwg-rma