The Research On Load Balance Policy For Large Scale VoD Server

The large scale VoD server can support vast concurrent streams, cover wide district and provide cooperative programs with the replicas of multimedia object placed the node near to user so that any program could be viewed at any time and any place. Because the each server load is varying with the server's place and request arrival, the recently research work more focus on the dynamic resource management. Based on the analysis of existed research work, the load balance mechanism is described generally in advanced and then the innovated research work is carried on the replica locating and load balance method about the dynamic replication, server selection and load migration as follows:Construct a logically hierarchical cluster for the replica locating in the dynamic scalable method. With the number and place of the replica varies with user request and server load, the server nodes which store same replica are organized into a logically hierarchical cluster in the view of replica connection as node connecting and replica as node. With the apportionment of management overhead to all nodes, the decentralized maintenance mechanism can meet the demand for the multi-replica access capacity, scalability, robust and adaptability. The mathematical analysis with cover rate indicates that the fault tolerant of logically hierarchical cluster is affected by the single cluster size as the number of whole node is fixed. The simulation result shows that the per-stream management overhead is extraordinary smaller than the delivery bandwidth of multimedia object, so the management overhead can be neglected.The node load is insufficient for the server selection because the multi-replica scatter at different place and the backbone is expensive. In order to alleviate the stress of backbone, a minimal latency server selection method is designed base on the retrieve cyclicity and delivery continuity. An end-to-end server response time is deduced according to the demand for retrieve cyclicity and delivery continuity. As the user requests for an object, the idle node with minimal response time is selected. The simulation result indicates the request acceptance rate and startup delay of the minimal latency server selection is superior to the minimal load server selection. It is also show the minimal latency server selection can improve the request acceptance rate and decrease the startup delay.The dynamic replication trigger policy is proposed where the early replication and instant replication is adopted according to node load state for the tradeoff of the expansion overhead and delivery load of logically hierarchical cluster. A request probability estimation function is derived for the target node selection and object for early replication. As the complement of early replication, the instant replication is used to balance load between nodes with the source node reserves more resource to server the local user. The simulation result shows that the dynamic replication improves the request acceptance rate and decreases the startup delay. The results also can be concluded that the replication overhead is mainly affected by the replication and its trigger threshold and then impacts the request acceptance rate and startup delay.The cooperative nodes deployed in a local region maybe belong to different logically hierarchical cluster, so the local load migration is used to further balance their load. The local load migration is based on the reciprocity capacity which is defined with the number of successful and failing migration. Not statically partition, the reciprocity node group is formed with the estimation and periodic exchange of reciprocity to adapt the varied node load state. An object which request number is maximal and not in migration process will be selected, then the source node will send message to other group nodes in turn according to cooperative capacity estimation until accept an acknowledgement. The simulation data supports that the local load migration can improve the request acceptance rate and reduce the startup latency. The difference of the accepted threshold and trigger threshold should be proper set indicated by the simulation result.