Research On Systemic Structure And Load Balancing Methods For Cluster System

With the development of computer and computer network, the price of computer and its component become lower. It is possible now to utilize a lot of computers to finish some work which cannot be finished or completely done by single powerful computer. In this paper, we will introduce a architecture of a cluster system from viewpoints of software and hardware, emphatically discuss some key technologies used when deploying a cluster system, such as scalability, availability, resource management and load balancing.We all know that one of the most important factors hampering the application of Web site is the high response time to the user request, which hindering the development of applications based on Internet.Cluster systems have been widely recognized by current research and industry. Cluster system is very useful in most of the situation such as super mainframe, distributed database system and some big network Web server. In this paper, we instance an example of 4-node cluster system to illuminate a way to improve the performance of the traditional web servers and use membership group to reduce the information amount which broadcast the system state of all the cluster system node. At first, we use a request-distributing node to dispatch the request from the web, and the request-dispatching node dispatch the user request to all cluster system web server nodes according the number request and the sum of all request serve time of every cluster system node. In this way, the request-dispatching node dispatch the first arrived request to the node which has little user request and serve time. When all user requests are dispatched or all server request queues are full, the request-dispatching node stop dispatching user request to wait new user request. Whenever a new user request arrive and no less than one cluster system node's request queue is not full, it start working again. Second, we use membership group to reduce the information amount of the all cluster system. Without using membership group, the cluster system will broadcast the state information at each system clock and it will transmit about(n-1)n pieces of node state information at each system which will brought greatsystem burden. Through using membership group, a node only communicate with its membership node and share the request with its membership node when it is in heavy system burden. In this way, we can balance the load of a heavy system burden node to its membership node. Because we choose the membership node according the factors of physics and group them by some rules, the heavy burden node will share the burden with the node which is its neighbor to him. When the first membership node of the heavy burden node is full of request, the heavy burden node will share the request with its second membership node, and so on. According to some rules, we arrange each cluster system node in different order which make different node's burden request queue different. In this way, we can avoid this situation that all heavy burden nodes dispatching the user request to the same node and make the request-receiving node's request queue full. Owing to make the membership node group be overlap in some way, we can make the cluster system load balance in some part of the systemfirstly. When all the nodes in the same membership's request queue are full, all nodes in the membership group will share their burden in their membership group respectively, which will make the burden shared in all cluster system.This paper discusses and provides the request dispatching and selecting schemes for Web server clusters and the stochastic high level Petri net models for those schemes. This paper focuses on these schemes, their performance models and analysis. Moreover, we applied approximate analysis technique to simulate the performance of the Web cluster system models, the request dispatching and select schemes; it has improved the performance of the Web Server greatly.