Re: [myrinet] Re: Registered memory in MPICH over GM

> Bill,
>
> > I have written a goodly amount of code using GM.  I am now about to use
> > MPICH-GM because I want to compare the performance of it with my code
> > while uses GM.
>
> We have MPICH-GM available on our web site. Please get the source and
> compile for your OS/platform.
>
> > I have the following question.  If you call MPI_Send or MPI_Receive, are
> > the respective buffers copied into DMA-able memory with
> > gm_register_memory or is it assumed that the buffers are already
> > gm_registered?

Patrick and Markus are right, for general user who use GM APIs,
following applies.

> GM send/receive functions need to operate memory that is DMAable.
> It is user's responsibility to prepare the buffer to be DMAable
> before calling GM send/receive functions.
>
> The user can either use GM API gm_dma_malloc/gm_dma_calloc to
> allocate DMAable memory of certain size and copy data into it, or
> using gm_register_memory to lock down user memory which make the
> memory DMAable.

Only if you want to implement MPI_Send/MPI_Receive by yourself, you need
to worry about above statements. Fortunately, these have all be taken care
of in MPICH-GM.

> In your case, you will need to implement MPI_Send/MPI_Receive using
> GM functions.
>
> Karen Wang
> ---------------------------------------------------------------
> Dr. Karen Honge Wang                    Software Development
> (626) 821-8015                          Myricom Inc.
> [email protected]                          325 N. Santa Anita Ave.
> http://www.myri.com                     Arcadia, CA 91006
> ----------------------------------------------------------------
>
>