Mixing ranks, tasks, progress and nonblocking collectives

摘要: Since the beginning, MPI has defined rank as an implicit attribute associated with process' environment. In particular, each process generally runs inside a given UNIX and is fixed identifier in its WORLD communicator. However, this state of things about to change rise new abstractions such Sessions. paper, we propose outline how evolution could enable optimizations which were previously linked specific runtimes executing processes shared memory (e.g. thread-based MPI). By implementing runtime-level work-sharing through what define tasks, enabling ability progress indifferently from stream context show that there potential for improved asynchronous progress. absence Session implementation, assumption validated where nonblocking Collective (NBC) implemented on top Extended Generic Requests progressed by any node thanks extension threads dynamically share their context.