Scalable Collective Communication for Next-Generation Multicore Clusters With InfiniBand
The emergence of multicore systems poses several challenges for the next-generation cluster architecture. Due to the availability of multiple processing cores per node, more number of application processes can be run on each node. In this context, the scalability of the cluster communication middleware like MPI is important for the overall performance of large scale applications. MPI collective operations are widely used by many application codes to get good efficiency and high speed-up in performance. It is imperative that these operations scale both with respect to the protocols used and the communication resources employed for these multicore systems. SMP-aware collective optimizations have been well-researched in the past.