High Performance Broadcast Support in La-Mpi Over Quadrics

LA-MPI is a unique MPI implementation that provides end-to-end reliable message passing between application processes. LA-MPI collective operations are implemented on top of its point-to-point operations, using generic spanning tree-based collective algorithms. The performance of the collective operations scales in a logarithmic order over that of the point-to-point operations. Thus, it is desirable to provide more efficient and more scalable collective operations while maintaining the end-to-end reliability. To this end, we investigate the feasibility of utilizing Quadrics hardware broadcast in this paper. We explore several challenging issues such as broadcast buffer management, broadcast over arbitrary processes, retransmission and reliability. Accordingly, a low-latency, highly scalable, fault-tolerant broadcast algorithm is designed and implemented over Quadrics hardware broadcast. Our evaluation shows that this implementation reduces broadcast latency and achieves higher scalability relative to the generic version of this operation. In addition, we observe that the performance of our implementation is comparable to that of the high performance implementation by Quadrics Supercomputers World for MPICH, and HPs for Alaska MPI, while providing fault tolerance to network errors not provided by these.