[Mpi3-rma] Draft Plenary Presentation for MPI RMA

Jeff Hammond jhammond at mcs.anl.gov
Fri Feb 4 16:24:21 CST 2011

If it is not obvious, ARMCI_Fence is bulk remote completion on a
single target, not global ala shmem_quiet (?).  Of course, ARMCI
AllFence completes at all targets what was sent from the calling

ARMCI is sequentially consistent between the source and target so
ARMCI_Fence ensures that the remote completion is visible to everyone.

Sorry for being pedantic.  Maybe everyone already knew all this.


On Fri, Feb 4, 2011 at 3:42 PM, James Dinan <dinan at mcs.anl.gov> wrote:
> On 02/04/2011 03:10 PM, Torsten Hoefler wrote:
>> On Fri, Feb 04, 2011 at 02:35:45PM -0600, Pavan Balaji wrote:
>>> On 02/04/2011 01:14 PM, Torsten Hoefler wrote:
>>>>> Slide 27: ARMCI has per-operation non-blocking calls.  Blocking calls
>>>>> and waits give you local completion and only fence gives you remote
>>>>> completion.
>>>> Yes, ARMCI has per operation remote completion which we don't support at
>>>> this point. Proposal 2 had this a while ago but we didn't find a strong
>>>> use-case (as opposed to the per-operation local completion). Added a
>>>> comment.
>>> No, ARMCI doesn't have per operation remote completion; only local
>>> completion. Attached the slides I had distributed earlier about this for
>>> your reference.
>> Ok, I got this impression from our previous discussion. So Jim, what did
>> your comment then mean?
> I meant the same thing as what Pavan wrote, but had a language fail: ARMCI
> gives per-operation local completion for non-blocking operations.  Blocking
> operations are locally complete when they return.  Bulk remote completion is
> accomplished through fence.
>  ~Jim.
> _______________________________________________
> mpi3-rma mailing list
> mpi3-rma at lists.mpi-forum.org
> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma

Jeff Hammond
Argonne Leadership Computing Facility
jhammond at alcf.anl.gov / (630) 252-5381

More information about the mpiwg-rma mailing list