[mpiwg-rma] [EXTERNAL] Re: MPI_WIN_CREATE for intercommunicators

[Barrett, Brian W]

On 1/22/14 11:09 AM, "Jeff Hammond" <jeff.science at gmail.com> wrote:

>On Wed, Jan 22, 2014 at 11:53 AM, Barrett, Brian W <bwbarre at sandia.gov>
>wrote:
>> On 1/22/14 3:41 AM, "Thomas Jahns" <jahns at dkrz.de> wrote:
>>
>>>Hello Maik,
>>>
>>>On 01/22/14 10:14, maik peterson wrote:
>>>> thomas, today you should start thinking about other possibilities
>>>> to solve your problem. this mpi rma stuff is crappy.
>>>
>>>I guess that very much depends on your kind of machine. We've found it
>>>to
>>>work
>>>very nice with MVAPICH, not so great with PE and OpenMPI.
>>
>> Another way to put it; it really depends on how much work your MPI
>> implementor has put into one-sided for your hardware.  MVAPICH has put
>>in
>> a good deal of effort to use IB-natvie RDMA semantics when implementing
>> MPI one-sided.  Open MPI has not put that effort into IB (it has for
>> another network, but that's a different story), so the performance is
>>much
>> lower.  Open MPI's largely driven by squeaky wheels, and that wheel's
>>been
>> pretty quiet :)..
>
>Are you saying that if I make a big deal about OpenMPI's RMA warts on
>your list that you all will make it MPI-3-feature-complete and perform
>really well?

Chances of me personally making Open MPI perform really well over IB are
pretty minimal.  But Mellanox might have a different view ;).  My point
was that of all the things people complain about, RMA performance
generally isn't one of them.

Brian

--
  Brian W. Barrett
  Scalable System Software Group
  Sandia National Laboratories