Research On Causal Consistency Model Based On Grouping Strategy And Adjacency List

To achieve high availability in a distributed system,we create multiple copies for each piece of data,and then these copies are assigned to multiple nodes for management.How to maintain the consistency between these copies is particularly critical.Among the many consistency schemes,causal consistency is favored by researchers because of its better performance.At present,the proposed causal consistency schemes rarely consider the excessive management overhead and communication overhead when there are a large number of nodes in the system under the environment of full geo-replication.At the same time,they failed to make full use of the advantages of partial geo-replication.To this end,this paper studies the causal consistency in the distributed system,and proposes the following two innovative solutions:(1)In the environment of full geo-replication,we propose a causal consistency model based on grouping strategy which solves the problem of excessive overhead caused by managing a large number of nodes in the system.We group the nodes that are close to each other in a logical position,and then we elect a leader for each group through specific rules.The leader coordinates the scheduling of other nodes in the group.In this way,we achieve hierarchical management of all nodes,which reduces system overhead to a certain extent and improves system scalability.At the same time,we introduce a client migration mechanism to ensure that the client in the remote data center can read the data written by the local client as soon as possible within a period of time.(2)In the environment of partial geo-replication,we propose a data causal consistency model based on the adjacency list,which solves the trade-off between throughput and update visibility that existed in the previous solution.We introduce the adjacency list to represent the copy storage relationships between data centers,and build a stable vector to indicate the stable state of each data center with adjacency relationships.When the data is synchronized,the update is propagated along the single list of the current node.While ensuring the update visibility,we increase the throughput of the system.