Research Of Parallel Communication Technique On Cluster Network

With the increasing development of the Parallel Computing technique, and owing to its great price/performance ratio, outstanding scalability and availability, Cluster system is becoming the mainstream of the High Performance Computer Architecture. In Clusters, the performance of inter-node communication is the critical factor, which affects the actual performance of Clusters. In addition, an efficient design and implementation of Cluster communication system have been a hotpot of such area in recent years.Because the CLUMPS has become the main trend of Cluster systems, and due to the coming of Clusters of large nodes and super nodes, the performance of Cluster communication system plays a more and more important role while the high processing speed of Cluster nodes is continuously advancing. In conventional Cluster networks with single NIC within nodes and single communication path among nodes, the capacity of Cluster communication subsystem has become a performance bottleneck for the restriction that comes from the limit of network devices. To resolve this problem, a new technique called Parallel Communication, which uses multiple NICs and multiple network paths to transmit messages among Cluster nodes, has been brought up. We introduced this mechanism into the Dawning 4000L Cluster to increase the bandwidth performance of its network.Based on the former works done by others, this dissertation firstly discussed the essential issues of Parallel Communication and investigated some typical implementations of such technique. Then we proposed a Parallel Communication scheme based on the message striping mechanism, presented its designed and implementation on the BCL basic protocol in detail, including the extended implementation of it. The experimental results showed that this implementation has a famous bandwidth performance, satisfying our design goal by and large. Finally, we concluded the thesis and the work done by this author, and gave some aspects of research work that we might do next.