[Mpi3-rma] Alternative Proposal for Shared Memory Support

Bronis R. de Supinski bronis at llnl.gov
Tue Mar 15 20:44:49 CDT 2011


I suggested a joint meeting because that will make it
clear on the schedule. I was not thinking about adding
administrative overhead. I would like to be present.
I will likely miss the Tuesday morning sessions due to
a long telecon that I need to attend.

I think some sort of "best effort" wording will be needed.
It is possible that processes are on the same node but
that no shared memory is available. In that case distinct
physical memory regions could be returned.

I am not quite following the group comments. Are you saying
that the different windows are returned to different shared
memory domains? I thought the window would be available to
all processes but physically shared memory is used when possible.
How is the group returned? How do you query it?


On Tue, 15 Mar 2011, Torsten Hoefler wrote:

> Bronis,
>> What happens if the processes do not have shared physical
>> memory?
> If the communicator had P processes, then the new window will have P
> windows (or groups) with one process in each.
>> Is it just that physical shared memory is used
>> when possible? What if multple nodes are included in comm?
> Yes, the call is similar to a comm split where all nodes that share a
> common memory have the same color and all nodes that don't have
> different colors. It will create multiple distinct windows. The user can
> query the group to see which processes share a window.
>> Will each node use shared memory? That seems fairly
>> reasonable.
> Yes, it's basically the same semantics as Ron's original proposal where
> a new communicator MPI_COMM_SHM (?) was defined. This functionality will
> be assimilated into one call and group query (which is more logical
> because one doesn't necessarily need a full communicator). A new use for
> groups :-).
>> Overall, I think this proposal is good. I do think a number of details
>> have to be worked out before it is ready but I agree that discussing
>> it at a Forum meeting is the way to go.
> Yes, there will be some discussion, but it could be worth it.
>> Perhaps we could request a joint RMA/Hybrid WG session to discuss it?
>> I think it is a mistake to cover it just within one group or the
>> other.
> Sure, that would be best. We could also just pick one or the other and
> make sure all key people from the both working groups are present ;-).
> Seems like less administrative overhead.
> Best,
>  Torsten
> -- 
> bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ -----
> Torsten Hoefler         | Performance Modeling and Simulation Lead
> Blue Waters Directorate | University of Illinois (UIUC)
> 1205 W Clark Street     | Urbana, IL, 61801
> NCSA Building           | +01 (217) 244-7736

More information about the mpiwg-rma mailing list