The Research Of Hybrid Schedulling Model Based On Sparse Matrix And Parallel Algorithm

Since Freud created the first ship model test tank more than 140 years ago,the fluid properties of ship and marine structures rely on physical model test detection.In the background of China's ship and marine industry innovation in urgent need of Internet plus ship/offshore engineering equipment hydrodynamic simulation of advanced design platform,the numerical tank is born.Many solution models in the numerical tank involve a large number of sparse matrix vector operations.While the existing sparse matrix vector parallel algorithm based on MPI parallel framework brings the gospel for solving model to improve computing efficiency.Aiming at the problem of existing parallel algorithm and scheduling policy model can not effectively solve the load balancing exposed by large sparse matrix vector multiplication,the paper studied the following aspects: optimization sparse matrix storage structure,static partition sparse matrix initialization,dynamic scheduling load tasks.And proposes a sparse matrix parallel algorithm based on hybrid scheduling model.The main work of this paper are as follows:(1)Analysis of existing representative storage formats for sparse matrix.for line format compression storage method was used to study from the principle and data structure,and the design of the storage structure of two index storage structure on the basis of.Optimizing sparse matrix storage space in order to reduce the amount of communication in the parallel cluster.(2)Analysis the traditional parallel partition when the existing matrix operation,and according to the essence of sparse matrix structure and vector computation,design a non zero element for equal partition algorithm to make the non-zero element allocation of sparse matrix is more balanced.According to the MPI parallel framework of inter process communication and process of realization of one dimensional design of zero elements for equal partition algorithm.(3)According to sparse matrix vector multiplication may be have the load balance problem in parallel computing,analysis and comparison of different scheduling start mode and cluster load balancing scheduling algorithm,this paper proposes a hybrid scheduling model for sparse matrix vector multiplication load balancing problem.(4)Analysis of the factors in the cluster effect calculation of load capacity of parallel processing nodes,the proposed hybrid scheduling model for load design,construct the corresponding ability to calculate the dynamic evaluation of cluster node load balance cluster expression and threshold,and the specific design of the mixed load scheduling model process to load tasks in the cluster.In this paper,the mixed load balancing strategy model through the use of a non zero element for equal partition and weighted round robin scheduling strategy of combining static and dynamic,and based on the experimental environment of MPI cluster,by comparing the experimental of the parallel hybrid model has a high load balancing strategy.