Re: alternate proposal

Eric Salo (salo@mrjones.engr.sgi.com)
Mon, 4 Mar 1996 13:26:17 -0800

> I thought that I understood your proposal until I got to the example:
>
> > MPI_PUT(buf, 100, MPI_INT, 3, 200*sizeof(int), comm)
> >
> >The data is copied by MPI into the shared page that corresponds to rank 3.
> >While the other ranks also receive a copy of this data, none of them can
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> >access it directly.
>
> Are you implying some sort of replicated shared memory model with
> consistency enforced by implicit broadcasts ? Probably not, but you might
> still consider rewording this paragraph.

Sorry, this was poorly worded. All I meant was that in this example, we're
implementing the put/get windows with real shared memory pages and so we have
to be careful. Let's say we had a 16KB window and four MPI processes. If the
MPI implementation had only allocated a single page of shared memory for them,
we would get different behavior than what we wanted because any data that was
written into the window of any one process would also appear in the windows of
every other process. So, we instead allocate four shared pages and partition
them up between the four processes. By writing into the window of one process,
we are still causing changes that are visible in the address spaces of the
other processes but these changes are hidden by MPI.

-- 
Eric Salo         Silicon Graphics Inc.             "Do you know what the
(415)933-2998     2011 N. Shoreline Blvd, 7L-802     last Xon said, just
salo@sgi.com      Mountain View, CA   94043-1389     before he died?"