[Mpi3-hybridpm] New draft of the EI chapter uploaded
balaji at mcs.anl.gov
Sun Feb 6 20:00:43 CST 2011
On 02/06/2011 06:25 PM, Rajeev Thakur wrote:
> MPI_Comm_shm_alloc probably needs an Advice to Users about the input
> communicator argument to this function. Just passing any communicator
> (e.g. MPI_COMM_WORLD) may not result in successful shared memory
> allocation. One use case is the communicator should represent
> processes on the same node. Other use cases could be implementation
> defined. Not sure of the exact language to put here, but it would be
> useful to make users aware of this issue in some way, rather than say
> nothing at all.
Right. There was also a suggestion that we provide a new predefined
communicator called (MPI_COMM_NODE or MPI_COMM_SHM) that allow shared
memory regions to be created on them, their dup's or subsets. We didn't
decide on what the right thing to do is (could be as simple as an advice
to users and say that MPI does not handle it). I'll add a discussion
item for it.
> This proposal is a shrunk version of earlier proposals, with only
> helper threads and shared memory allocation functions included.
> Endpoints are not included. Is the endpoints proposal going to be put
> forth separately later? I hope it is in good shape because it was a
> promising idea.
The endpoints proposal is not removed. We were just bandwidth limited to
bring text for all parts to this particular Forum meeting. We'll be
adding it as soon as this meeting is done, hopefully by the next one.
More information about the mpiwg-hybridpm