Distributed Load Balancing Strategy Research

With the computer technology entering the period of network computing, researches and applications on distributed system have been wilder and wilder. Apparently, there must be other computer structures substitute the traditional Von Neumann structure. The parallel and distributed system depresses the bottleneck of processing and provides better ratio of performance to price. And distributed system can continue to fuction when there are fault in parts of the system. So there are much more development in distributed system in future than today.The optimization of distributed system refers to load balancing. Though people have researched on load balancing for about 20 years, there are few systems are satisfied because load balancing is a NP problem.The paper presents theory of distributed system and technology of distributed computing at first, then researches on algorithms and models of load balancing particularly and roundly. In this paper, we analyzis the causations of the extra spending in distributed system and the difficulties in the load balancing researches.Our first innovation is presents a new load balancing policy which is based on the research on typical load balancing policies such as distributed^ centralizedN sender-initiated and receiver-initiated policy. Our load balancing policy synthesizes above policies' merits and gets over their shortcoming. It is more applicable and practicable, and can be the basis of the development of the industrial products.Secondly, most current researches have been concentrating on the proposing of the policies. Researches on the implementation model based on the policy are few. After researching several current load balancing policy, we builds several implementation models such as chain model, reticulation model and chain-reticulation model increased by performance and researches them comprehensively based on network topology. Each model is given corresponding algorithm and estimation. The advantages and shortcomings and applicable ranges of each model are also given. The paper especially emphasized the chain-reticulation model' s superiority and high performance on dynamic load balancing. These models we presented base on above policies and visualize abstract polices. They can be used for reference by other researches.At last, our experimentation uses JAVA RMI but not PVM used by much many parallel and distributed experimentations. We combine distributed theories with realization technology that would be basis of the industrialization in future.