On Mar 15 2011, George Bosilca wrote:
Nobody challenged your statements about threading or about the correctness of the POSIX standard. However, such concerns are better voiced on forums related to that specific subject, where they have a chance to be taken into account by people who understand them.
That would be true if I were still trying to get the shambles fixed; I am not, because I tried and failed at the appropriate time. I am warning OpenMPI people that the infrastructure threading support it NECESSARILY relies on is NOT RELIABLE.
This particular topic was about MPI level threading support, more specifically about the threading support Open MPI would like to provide. In this limited context, people interested in using thread and MPI together are well aware about the limitations imposed on their application, as well as the pitfalls they have to avoid. Moreover, with the new endpoint proposal (the one I was pointing on my previous email), the threads will become first-class citizen in MPI.
If you think that even relative experts are well aware of the limitations and pitfalls, you are living in a fantasy world. But that is not the real point. It is that OpenMPI CANNOT make them first-class citizens, no matter what it does, because the infrastructure does not support that. At best, it can make them APPEAR to be so, with the specification that sometimes they will work and sometimes they won't, nobody will know where the boundary lies - and almost all failures will be rare, system-, configuration- and workload-dependent, unpredictable and non-repeatable. I would a thousand times sooner not have a 'facility' like that. Regards, Nick Maclaren.
