Load-distributing algorithm using fuzzy neural network and fault-tolerant framework

Abstract

Load-Balancing System is widely used for difference transaction/data processing systems in many years. Now, that is one of the most important techniques to provide multiple servers architecture to the web system. Dynamic distributing algorithm in the Load-Balancing system provides more effective and fair job distribution to the servers in the system, so that the total performance of the system is enhanced. However, most of them cannot be applied in the practical Web System because of the characteristics of the web traffic. In additional, the fault-tolerant issue to the stateful protocol such as TCP-IP is another problem for Load-Balancing System using in the Web System. In this thesis, new models of load distributing algorithm and fault-tolerant framework to the Load-Balancing System are proposed. These two models are mainly for solving the problems arisen in the traditional Load-Balancing Algorithm used in a Web System and for optimizing the performance of the system. The proposed work scheduling algorithm of a web based Load-Balancing System is based on a Fuzzy Neural Network and Kohonen’s Algorithm. In the proposed algorithm, the real-time system usage is feedback to the system with an updating mapping rules based on different requested objects categorized into different servers groups with different cache sizes and according to their input frequencies to enhance the cache hitting rate of scheduling. Simulation results show that the proposed technique obtains 92% to 99% cache hit-rate and it finely balances the backend server resource usage. In the proposed framework for the fault-tolerant in Load-Balancing System, a new method based on the group backup is used. In this method, the connection data is stored in peer servers effectively. It also provides a fast server switching when a backend server becomes faulty. Comparing with the traditional fault-tolerant methods, this framework provides a more safe and a cost effective solution to the system.