[mvapich-discuss] MPI and CUDA Streams

[Jeff Hammond]

Hi Brody,

I wonder if generalized requests could be used in this context.  I'm
not suggesting that these do exactly what you asked for here but one
might be able to do something that accomplishes the overall goal
coupling asynchronous functions in MPI with asynchronous aspects of
CUDA.

Best,

Jeff

On Sun, Sep 16, 2012 at 9:53 AM, sreeram potluri
<potluri at cse.ohio-state.edu> wrote:
> Hi Brody,
>
> MVAPICH2 does use streams/events internally to pipeline copies from the GPU
> and transfers over the network. However, it does not have a way to expose
> these streams/events to the application or for the application to indicate
> which streams/events MVAPICH2 should use.
>
> The only way to synchronize between MPI_Isend/recv and GPU kernels is to use
> MPI_Wait/Waitall which I assume is what you are using right now.
>
> Best
> Sreeram Potluri
>
> On Sat, Sep 15, 2012 at 11:34 PM, Brody Huval <brodyh at stanford.edu> wrote:
>>
>> Hi,
>>
>> Is there a way to have MPI_Isend/recv operations that are synchronized
>> with cuda streams and cuda events? Such that we can queue them with other
>> GPU kernels?
>>
>> Thanks,
>> Brody Huval
>> _______________________________________________
>> mvapich-discuss mailing list
>> mvapich-discuss at cse.ohio-state.edu
>> http://mail.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss
>>
>
>
> _______________________________________________
> mvapich-discuss mailing list
> mvapich-discuss at cse.ohio-state.edu
> http://mail.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss
>



-- 
Jeff Hammond
Argonne Leadership Computing Facility
University of Chicago Computation Institute
jhammond at alcf.anl.gov / (630) 252-5381
http://www.linkedin.com/in/jeffhammond
https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond