[Mpi3-hybridpm] Shared memory segments proposal

[Ron Brightwell]

On 11/05/2010 04:40 PM, James Dinan wrote:
> Hi All,
>
> I wanted to follow up on my questions during today's hybrid telecon.
> Thanks, Ron, for taking the time to explain the shared memory segments
> proposal and sorry if I took us off on a tangent.  I just joined Argonne
> in September so I wasn't able to arrange travel in time to join you all
> in Stuttgart.

No problem.

>
> I took a look at the slides and overall I think the shared memory
> segments proposal is a good idea.  It will provide a good alternative
> for programmers who want to create node-level shared structures but want
> to use processes instead of threads.  I think that in order for this to
> be more generally useful, we should consider adding something like
> semaphores, mutexes, or some other data-driven (as opposed to send-recv
> which will require an MPI rank) synchronization mechanism.

I don't necessarily disagree. I just don't know what the "right" 
synchronization mechanism(s) should be. I was hoping we would get a 
better idea of this from experience, but we haven't had enough 
experience with the work we've done so far to know. My first inclination 
is to see if there's something in the MPI RMA operations that we could 
possible leverage rather than having something separate.

>
> Another comment: Given that we are doing allocation collectively, could
> we actually guarantee that the shared memory is mapped at the same base
> pointer on every process?  We do have MPI at our disposal so we could,
> in principle, do some communication up front to coordinate a base
> address that works for everyone before we map the shared segment.
>

Good question. The anticipated usage model certainly makes this easier 
too. I'll have to think about this...

-Ron