Cluster computing systems are becoming increasingly popular for providing cost-effective and affordable computing environments for day-to-day computational needs for a wide range of applications in various disciplines. The emerging InfiniBand Architecture (IBA) standard is being positioned as a strong candidate for high-speed interconnect for next generation clusters. Compared to previous generation cluster interconnects, IBA provides many novel features such as multiple transport services, Remote DMA (RDMA) read/write support, atomic operations, hardware multicast support, and multiple service levels (priorities). These novel features provide unique challenges and opportunities to design scalable and high-performance communication subsystems for next generation clusters. This research will focus along the following two directions to efficiently support message passing programming paradigm (using the Message Passing Interface (MPI) standard) on clusters: 1) Low-latency, high-bandwidth, and scalable point-to-point communication operations; and 2) High-performance and scalable collective communication operations. Optimal protocols and algorithms along these two directions will be developed while taking advantages of the novel features of IBA. The derived solutions will be evaluated through a set of micro-benchmarks and applications for performance and scalability. Best solutions will be integrated into an MPI implementation and will be available for public use. This will enable architects to design next generation high-end distributed-memory systems with IBA-based clusters. It will also facilitate the end-users to run their MPI applications on IBA-based clusters to harness best performance with scalability.
