Design And Optimalization Of Concurrent Transmission System For Parallel Computing

Parallel computing can improve the computing performance of the system by distributing a computing task to multiple distributed processing units and excuting simultaneously.The processing unit or computer which is equipped with multiple network interfaces has the capablity of concurrent transmission.This thesis studies the speed-up method of parallel computing and fault tolerant performance of networks for the data communication advantages of multi-path concurrent transmission.Firstly,the thesis discusses the related technologies of parallel computing and multi-path concurrent transmission.The characteristics of MPTCP and SCTP are analyzed in detail.Then the PingPong test,high-performance Linpack and test method of performance benchmarks are briefly discussed.Secondly,the thesis analyzes the process of solving the large-scale dense linear equations by LU.The influence of parameters such as matrix size,number of processes,processor grid and matrix block on computational performance are studied in single and four node cluster environments.The results show that the optimal parameters of single machine environment are N≤100000,np = 2,P×Q = 1×2 and NB = 256.In the four nodes environment,the optimal parameters are N≤200000,np = 8,P×Q = 2×4 and NB = 256.Finally,the thesis analyzes the data transmission mechanism of OpenMPI through BTL module.The communication optimization scheme of introducing MPTCP,SCTP and Infiniband are proposed.The system implementation is completed,and its performance is tested.According to the results of HPL,the maximum speed-up ratio is 1.30 for dual-path Gigabit MPTCP,which is comparable with the 40Gb/s Infiniband whose maximum speed-up ratio is 1.61.Compared with the SCTP,MPTCP has better fault tolerance because it has shorter path switching time and is unrelated with the number of processes.