Research On Key Issues Of Communication System On Cluster Of SMP's

Cluster is a widely used architecture in modern super-computer and super-server. Since cluster of SMP's (CLUMPS) has high ratio of performance and costand high fault tolerance ability, it becomes a hot point in current research fields.Communication system is a key factor that affects the actual performance ofCLUMPS. Since the architecture of CLUMPS is different to uni-processor cluster,the communication system is different, too. The research on communication systemof CLUMPS plays an important role in the research and development of CLUMPS.After analyzing all aspects of cluster architecture and cluster communicationsystem, this paper introduced a system hierarchy model by analyzing connectionperformance. A multi-protocol mechanism, which consists of a shared memorybased protocol and a semi-user-level protocol, is discussed under this model.Then this paper concentrates on the key issues of the two protocols of themulti-protocol system. Some key issues are discussed to improve the performance ofshared memory based protocol. The security, reliability, portability andheterogeneous network environment support of the semi-user-level protocol are alsodiscussed. Process Algebra is introduced to describe and verify the protocol.Performance is a key factor to evaluate a communication protocol. This paperpresents the evaluation model and performance analysis of BCL-3, which is theBasic Communication Library of DAWNING-3000. Point to point communicationperformance, LogGP communication parameters and collective communicationperformance are used to evaluate the whole system. The tests showed goodperformance and scalability of the protocol.Finally, this paper discussed the factors that affect the communicationperformance. A multi-card protocol model and a double-card protocol prototype areintroduced to improve the performance.