Research Of Replica Management Mechanism In Cloud Storage System

With the rapid growth of information in the network environment, replica managementmechanism of the cloud storage system should meet the need of strong reliability, good load balance,high replica utilization, low network congestion probability and etc.This thesis proposes two strategies about replica management—an Replica Placement Strategybased on Node Grouping and Distributed Hash Tables and an Data Replication Strategy based onData Classification and Popularity Prediction, and module of management mechanism based on thetwo strategies is designed.The main work of the thesis is as follow:1. An Replica Placement Strategy based on Node Grouping and Distributed Hash Tables. Nodesof the cloud storage system are divided by disaster district into groups, with the nodes and replicasmapped by distributed hash tables technology; so as to prevent overheating, only nodes with lowload could storage an new replica. Compared with Lightweight Adaptive Replication, the ReplicaPlacement strategy bring the cloud storage system with higher data reliability，and higher loadbalance if compared with Hot Ranges/Rings strategy of Data and QoS-aware Replica Placementstrategy of Core-Selection.2. Data Replication Strategy based on Data Classification and Popularity Prediction. In order toreduce network congestion peak of access to information, with better control of the overall numberof replicas, the law of access from users is summarized, and replicas are detected for previouslycreation and placement under different access laws. The work of the replication strategy containsthe determination of current data type, calculating the expected number of new replicas and dealingwith hot replicas. Compared with Pure Popularity Replication strategy, the replication strategycould satisfy information need from users, and maintain a low number of replicas with hightemperature when there come the access peak period.3. The modular designation of replica management mechanism. The replica managementmechanism is composed by five modules:resource logically organization module for markingnodes and replicas; initial replication module for creating two replicas of source data; forecastedreplication module for creating replica needed in future; replica placement module for placingreplicas on target nodes; hotspot detection module for nodes detecting local load by themselves.