Attached is a slightly updated version of the slides. I had forgotten to include one additional item (in the last slide of the attached slide deck). -- Pavan On 12/02/2012 09:30 AM US Central Time, Pavan Balaji wrote:

On 12/02/2012 05:27 AM US Central Time, Torsten Hoefler wrote:

...

The only thing I disagree to is slide 6. Win_sync really does nothing in the separate case. Your example would work on shared memory if flush issues a memory barrier *and* MPI_Barrier() is an optimization barrier (which is is in C and needs to be ensured in Fortran anyway). But we can discuss next week.

The problem here is that MPI_BARRIER is a process barrier, not a memory barrier. FWIW, with current compilers, which do not reorder instructions across functions, this should work correctly. But this is a correctness concern when compilers do more than that, e.g., IPC optimizations.

An alternative would be to leave it as-is, and add a note that we assume the compiler will not reorder instructions here because it's a function call, though I don't prefer this alternative.

-- Pavan

-- Pavan Balaji http://www.mcs.anl.gov/~balaji