Aurelien: Re:
I have not made myself clear here. The implementation I am talking about supports the entire interface. It just does so without reporting errors and lets MPI in an undefined state after process failures, as is
I understood that was what you meant. While you may not see providing that implementation as a "quality of implementation" issue, others would. So many implementors would feel compelled to provide an implementation that does more. It is the fault-free overhead of the one that does more that is critical. Bronis
specifically allowed by the draft of #323. This is not a "quality of implementation" issue, as there are very valid and justifiable reasons for an implementation to choose not to support fault tolerance, such as when the target hardware is reliable. However, in such an implementation, an FT application is still portable and supported, but it will not survive failures (due to lack of support from the MPI layer). Cost on implementors is ridiculous (implementing empty stubs mostly, the most "complex" function is agree, which is a straight remaps to allreduce). The cost in performance is null, zero.