[mpiwg-rma] shared-like access within a node with non-shared windows
balaji at mcs.anl.gov
Fri Oct 18 16:29:28 CDT 2013
On Oct 18, 2013, at 4:25 PM, Jim Dinan wrote:
> This is only correct if you assume that the remote NIC can't translate a displacement.
> Allowing all processes on the node direct access to the window buffer will require us to perform memory barriers in window synchronizations, and it will cause all lock operations that target the same node to block. I understand the value in providing this usage model, but what will be the performance cost?
I think Jeff is asking for this model to "also" be available, rather than this being the "only" model. For example, the memory allocation will not be symmetric for all processes if it's shared memory, but maybe the application is willing to take that cost.
However, I'd still appreciate some discussion on the semantic benefit of this approach compared to the two window approach. Yes, two windows are ugly, but we need to be clear if we are doing this to clean up the interface (which is a valid reason) or we are gaining additional semantics (e.g., we can now do exclusive locks).
More information about the mpiwg-rma