Design Of Raft Cluster Distribution Strategy Based Erasure Coding

In order to provide fault tolerance and data consistency in global or local distributed data centers,distributed data services usually rely on consensus protocols such as Paxos and Raft to make multiple server nodes reach a consensus.Compared with traditional multi copy replication,the consistency protocol based on erasure coding data replication can better save storage and network costs,and help to achieve low latency,high fault tolerance and high throughput of data access.However,the direct application of erasure coding in consistency protocol will reduce the activity level of the system.To solve this problem,when the number of failed servers reaches a certain threshold,CRaft stores full replication rather than erasure coding replication.However,it increases storage and network costs.In addition,the using of erasure coding leads to the increase of write set and higher write latency in Wan.If the number of nodes in the write set could reduce,using a server with lower wide area latency for data commission will bring a better experience to users.To solve the above problems,this paper improves the CRaft protocol from two aspects: adjusting the distribution of coding blocks and the size of write set.The main research contents are as follows:(1)A consistency protocol HRaft for adaptive erasure coding data maintenance is proposed.The opportunity of erasure coding data replication in CRaft is that the system has more healthy servers.By analyzing the characteristics of erasure coding data block reconstruction,it is found that fewer data blocks can be used to ensure the recovery of erasure coding data in case of server failure.Therefore,how to adjust the location of erasure coding data according to the state of the system has become a key challenge.According to the current state of each server of the system,this paper uses the secondary supplement method to supplement the necessary redundant data.The key idea is to use the result information of the first round of data distribution to adjust the placement of coding blocks under the condition of ensuring the recovery of erasure coding data to maintain the consistency of data.Experimental evaluation shows that when a server failure occurs in a cluster composed of five servers,HRaft has better performance than craft,the write delay is improved,the leader's network bandwidth consumption is reduced by 44.8%,and the throughput is 118% higher than CRaft.(2)An adaptive quorum based on latency grouping is proposed.Using erasure coding data distribution leads to the increase of write set and needs to wait for more node responses,which means higher delay in Wan.This paper studies how to adjust the data distribution scheme according to different requests.According to the distribution scheme determined by the number of health servers,the deployment in the Wan has a fixed write latency.In contrast,a distribution strategy is designed by using the delay of the leader's last heartbeat to select the appropriate write set.This strategy can use the server nodes with low wide area latency to complete the data placement,and save storage and network cost while ensuring the latency limit.