Pipelining and overlapping for MPI collective operations

Collective operations are an important aspect of the currently most important message-passing programming model MPI (message passing interface). Many MPI applications make heavy use of collective operations. Collective operations involve the active participation of a known group of processes and are usually implemented on top of MPI point-to-point message passing. Many optimizations of the used communication algorithms have been developed, but the vast majority of those optimizations is still based on plain MPI point-to-point message passing. While this has the advantage of portability, it often does not allow for full exploitation of the underlying interconnection network. In this paper, we present a low-level, pipeline-based optimization of one-to-many and many-to-one collective operations for the SCI (scalable coherent interface) interconnection network. The optimizations increase the performance of some operations by a factor of four if compared with the generic, tree-based algorithms.