Exploring collective communications on a cluster of SMPs

We investigate two strategies to improve the performance of collective communications on a cluster of SMPs. In the first part of the paper, we explore the advantages and disadvantages of the commonly used strategy "optimize shared memory send and receive only". In the second part of the paper, we show a model to design collective communications on SMP clusters which can be used to improve the disadvantages of the former strategy. Several collective operations are implemented through the later process, and the experimental results show better performance than vendor supplied implementations.