[mvapich-discuss] truly one-sided communication
Thiago
thiago at sci.utah.edu
Wed Jan 19 11:21:42 EST 2011
Ah pity! I need this for a paper due in a few days. Hmm, all I really
need is an RMA read from the target memory which is read-only and never
modified. I take it this means that synchronization is unneeded? Aside
from waiting for MPI-3 to hopefully have this as an option (please!),
how painful would it be to hack MVAPICH2 to simply throw out the
synchronization so that when I call MPI_Get it only does a single RMA
read over Infiniband and the lock/unlock skip the communication? If
that's too complicated, what would be the next best option? I have 64
nodes, each running 8 threads where each thread does computation and
random MPI_Get communication. Is there something, (perhaps unportable
and only works on 1.6-rc2!) that I could do that would let me call
MPI_Get without having to have the target know this is happening?
Perhaps using MPI_MODE_NOCHECK and doing "undefined" behavior that
actually ends up working just this time? If I must do a standard
passive communication, what's the best way to ensure my performance
doesn't tank? Is there a cheap MPI function I can have a thread do in a
while loop to ensure that progress is made?
Thanks,
Thiago
On 1/19/2011 7:22 AM, sreeram potluri wrote:
> Dear Thiago,
>
> Thanks for your email. Currently, MVAPICH2 supports truly one-sided
> communication for the active modes of synchronization (Fence and
> Post-Wait/Start-Complete) available in MPI-2 RMA. The truly one-sided
> designs for passive mode are in the pipeline for the next (1.7)
> release of MVAPICH2.
>
> Thank you
> Sreeram Potluri
>
More information about the mvapich-discuss
mailing list