Research on Distributed Matrix Computing Based on LT Code
Distributed computing systems have emerged as necessary tools for processing large amounts of data in the context of the continuous expansion of big data and machine learning applications.For computing clusters with a certain scale,their performance are inevitably affected by system noise.Therefore,it is necessary to consider encoding technology in distributed computing systems to enhance their robustness.Most of the existing encoding schemes used in distributed matrix computing are fixed rate codes,which cannot adapt to the actual situation of dynamic changes in the number of nodes.Meanwhile,owing to deadlines for some tasks,the average cost should be minimized completely to reduce latency while ensuring smooth task completion.To address the above issues,this paper proposes the application of the Luby Transform(LT)code to distributed matrix computing in fog computing scenarios,and designs Remo2 algorithm.Based on the rate-free characteristics of LT codes,adaptive channel state changes can be achieved through an appropriate degree distribution function design,bidirectional cutting,and degree factorization methods to reduce latency and enhance the robustness of distributed computing systems.This paper lets k1 be the row value of the submatrix after the partition of the A matrix and k2 be the column value of the submatrix after the partition of the B matrix.The experimental results indicate that under a fixedk,value precondition,compared with the Factored LT(FLT)code and Block-diagonal Coding-LT(BDC-LT)algorithm,the average cost of the Remo2 algorithm can be stably reduced by 33.3％compared to that of the former,and the redundancy can be reduced by 7.7％compared to that of the latter.In addition,when the code length of k1k2 is fixed,a lower degree of discretization of k1,k2,and lim|k1-k2|→ 0 results in a smaller average overhead.
万方数据
维普
