[mvapich-discuss] Problem with mulitiple gpus' send/recv task

[khaled hamidouche]

Hi Feiyu,

MVAPICH2-GDR does not support Multiple GPUs per process. Each process can
use only 1 GPU, however multiple processes can share the same GPU.

Is there a reason to not use a process per GPU rather than 1 thread ?

Thanks a lot

On Sun, Jan 25, 2015 at 3:43 AM, <feiyulv at mail.ustc.edu.cn> wrote:

> All
>   I'm trying to use two MPI processes to control 4 gpu's send/recv between
> two servers, the procedure is shown as follows:
>
> Server1
>  Server2
>
>
>
> (GPU1)  tid1 – MPI_Send  ----------------------à       (GPU2)  tid2 –
> MPI_Recv
>
>
>
> (GPU2)  tid2 – MPI_Recv  ß----------------------       (GPU1)  tid1–
> MPI_Send
>
>    However, I got a cuda error 33 when implementing it on mvapich2.0-gdr.
> Does one mpi process can just control one gpu to send/recv message? Or some
> other reasons caused this problem.
>    Any suggestions will be appreicated
>
> --feiyu
>
>
> _______________________________________________
> mvapich-discuss mailing list
> mvapich-discuss at cse.ohio-state.edu
> http://mailman.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cse.ohio-state.edu/pipermail/mvapich-discuss/attachments/20150126/1c8e6fa1/attachment.html>