It's prohibited by the standard to read send buffer before a send
operation completes, and I understand the theoretical rationale behind
I'm currently layering a protocol stack on top of MPI, and this
protocol allows a buffer to be read by multiple peers concurrently.
Thus for strict conformance I can either serialize read access to such
buffers or just make memory copies - neither approach is optimal. I'm
wondering that whether there is any practical exploitation of this
restriction - Is there an Open MPI BTL driver (or whatever it is
called in the Open MPI architecture) that really unmaps pages that are
totally contained inside the communication buffer from the application
address space while doing DMA?
If the answer is none or very rare, then I'll perhaps just dismiss it
as a red herring and big headache can be avoided.
() ascii ribbon campaign - against html e-mail
/\ - against microsoft attachments