Zero-copy, Hierarchical Gather is not possible with MPI Datatypes and Collectives
Jesper Larsson Träff, Antoine Rougier · 2014
It is a desirable feature of MPI that application specific collective operations can be implemented efficiently in terms of other, more primitive operations of MPI. Recent MPI 3.0 functionality makes it possible to build portable libraries that are sensitive to and can exploit the hierarchical structure of, e.g., current shared-memory clusters, and thus in principle to implement efficient, hierarchy-aware collective operations effectively.