A load balance service based on probabilistic neural network

For widely supporting the group communication in the WAN environment, client/server cluster architecture is introduced in our protocol. In order to balance the loads on servers, a novel method for characterizing the workload is presented to replace the traditional workload descriptor - the CPU queue length. The combination of the CPU queue length and other five workload indices are used to determine the load levels. A modified probabilistic neural network is adopted to classify the server load states into six types with the index combination as the input vector. The experiment proves that the shorter mean response time can be obtained in the new method. Based on it, load balancing policy is executed in the system to balance loads on servers.