Dynamic Load Balancing With Uncertain Factor Based On Cluster

Cluster computing is one of the important branches in computer science. Using cluster computing to resolve large data problem or high time complexity problem is the first choice in other scientific fields, as well as in the computer field. Meanwhile, load balancing is the key technology of cluster system, which is to determine the mapping between tasks and nodes and to improve the cluster's performance. However, DLB(dynamic load balancing) is needed to deal with some uncertain factors in load balancing, such as the time of task executing, the number of tasks, etc. Therefore, the research of DLB in this paper is proposed based on the uncertain time of task executing, which aims to gain a load scheduling strategy and a load estimate method to make the scheduling evaluation attain a better result. The main content of the paper is as follows.(1) On the base of four-tuple DLBM(dynamic load balancing model) proposed by other scholar, this paper introduces a five-tuple DLBM. DLB ranges over many aspects that are physical environment, software environment and all kinds of strategies. They are influenced by each other. The five-tuple DLBM involves hardware environment, scheduling environment, task allotment, load estimate, scheduling strategy and scheduling evaluation. They are all expressed through math definition. The five-tuple DLBM interprets the lull aspects of DLB and makes the logical relationship between the five tuples more clear by using formalized language.(2) According to the five-tuple DLBM and the corresponding solution of each tuple, the paper realized the model under a certain experimental conditions based on the Linux cluster system. Task allotment is completed by transferring array's index. It's simple and is no need to access kernel files. This paper adapts the Max-Min method, which is better than before in the aspect of estimate accuracy. Furthermore, the machine statuses are reduced into two ones, one is idleness and the other is overloaded. First of all, the former sends a request to the control node. On receiving the request from the idle node, control node calculates the most overloaded node that is to send the first index of tasks to be executed to idle node. In this way, all the nodes spend less time on scheduling the tasks.(3) Experiments are made according to the five-tuple DLBM based on tasks that are independent, having a little or no relativity, uncertain executing time. The result shows that the strategies proposed in this paper can make the load balancing better. What's more, scheduling evaluation shows that the strategies are reasonable.