Optimizing Multi-core MPI Collectives with SMARTMAP

We recently enhanced a lightweight compute node operating system with a simple page table mapping strategy that allows parallel application processes within a node to share effectively a single address space. This strategy allows for each process to maintain a normal, private address space, but also allows for processes to behave like threads that can freely read and write the memory of all processes on a node. In this paper, we demonstrate the benefits of this approach for improving the performance of MPI collective operations. In particular, we describe a new multi-threaded MPI reduce algorithm that outperforms existing approaches by as much as a factor of seven on a quad-core processor.