Research On The Optimal Placement Strategy Of LRC's Nodes Under The Hierarchical Structure Of Cluster

The heavy workload of the data center resulting from the explosive data growth can be greatly relieved by the distributed cluster storage system(DCSS).DCSS has been widely used in high-performance computing,big data analysis,and high-volume video-cloud applications due to its high reliability,availability and scalability.DCSS place storage nodes in different clusters,and the distribution of bandwidth resources between inner-cluster and cross-cluster is extremely uneven.The inner-cluster bandwidth is usually 20 times more than that of the cross-cluster.In some extreme cases,the ratio reaches 240 times.With the gradual increase of the data scale and the node number,the problem of node failure appears in the distributed cluster storage system accordingly.Local Repairable Code(LRC)has been widely used in distributed storage systems because it can perform the repair localy.LRC can improve the fault tolerance of distributed storage systems and improve the reliability of data storage.Therefore,the placement strategy of the nodes of LRC in DCSS according to the hierarchical structure of the cluster has become a current research hotspot.This dissertation proposes a novel placement strategy for two types of LRC nodes in distributed storage clusters according to the cluster hierarchy.The proposed placement strategy can reduce the number of code blocks transmitted across the cluster during the repairment of single-node and multi-node failures.In other words,it can reduce the crosscluster bandwidth where resources are in a shortage.The cluster bandwidth overhead has been simulated and verified using MATLAB.The main contributions of this thesis are as follows:(1)A placement strategy for regular(k,l,g)LRC code nodes in the cluster has been optimized.In the case of ensuring the minimum scaling overhead when hot data becomes cold data,the global check nodes are placed in the cluster where the specific data node is located,instead of being placed in a cluster alone.At the same time,the cross-cluster bandwidth overhead of two-node repair under this placement strategy has been analyzed.According to the simulation results,the optimized placement method can reduce the cross-cluster repair bandwidth overhead by up to 51.4%,as the number of data blocks increases,the gap between the cross-cluster bandwidth overhead before and after optimization is smaller.(2)A general expression of the parity check matrix of(n,k,r,t,d)New BLRCs codes suitable for storing hot data with a locality of 2 and uneven availability has been obtained.According to the expression,a placement strategy to place all nodes in k clusters has been proposed.The cross-cluster bandwidth of single-node repair and multi-node repair under this placement strategy have been analyzed.It has been contrasted with the methods of flat placement,random placement and order placement.According to the experimental results,under this strategy,both the repair cost of a single-node failure and the two-node failure are far less than other placements strategy.Compared with flat placement,the single-node repair cost under this placement strategy is reduced by more than 91.7%,and the repair cost of two-nodes is reduced by more than 90.8%.