一种多层网络下动态负载均衡算法

A Dynamic Load Balancing Algorithm in a Multi-level Network

分布式系统由若干个独立的节点组成,一些节点由于接收到大量请求而过载,还有一些节点却负担较少的请求任务。通过负载均衡技术可以使节点间的负载分配更加合理,最大化利用服务器集群的处理能力,达到扩展服务器集群的带宽和增加吞吐量,加强网络数据处理能力,提高网络的灵活性和可用性的目的。传统的集中式负载均衡方案采用静态负载均衡算法,由控制器全权负责任务分配。它的优点是功耗低而且稳定性强,缺点则是负载均衡效果不是最佳,总体处理速度较慢,中央控制器节点由于负担重容易成为系统瓶颈。同时,它的系统扩展在大规模集群中表现差。相比之下完全分布式方案是可扩展的,由于所有节点既是处理节点,也是分发器,而调度器只负责任务调度,从而减轻了控制器的负担,避免成为系统瓶颈。提出了一种异构分布式计算系统集群的负载均衡策略。该算法采集各个节点CPU使用率、存储器使用率两个系统参数,以决定各节点的工作量。同时,设计两层结构,解决全局通信负担较重的问题。仿真结果表明,该算法有效提高了负载均衡的效率。

A distributed system consists of several independent nodes,in which some nodes may be overloaded due to massive requests arrivals,and another some are idle without any requests. Load balancing techniques can be used to effectively distribute the load between nodes to reach the purpose of extending bandwidth of server clusters,increasing its throughput,enhancing network data processing capability,improving network flexibility and availability. Traditional centralized load balancing adopts static load balancing algorithm,solely responsible for the tasks assigned by the controller.Its advantage is lowpower consumption and high stability,and its disadvantage is not the best in the load balancing effect and slowoverall processing speed.Due to the heavy burden,the central controller node can easily become a bottleneck.At the same time,its system scalability is poor,with bad performance in large scale cluster.By contrast,a fully distributed solution is scalable,because all nodes are both processing nodes and the dispatcher,while the load scheduler is only task scheduling,thereby reducing the burden on the controller to prevent it from becoming a system bottleneck.A heterogeneous distributed computing systems in the cluster load balancing strategy is proposed.The algorithm requires the CPU usage and memory usage to determine the workload of each node.At the same time,two-level structure is designed to solve the problem of heavier global communications burden.Simulation results showthat the algorithm can effectively improve the efficiency of load balancing.