Research Of An Adaptive Failure Detection Model Based On Message Delay Prediction

Failure detection is a key technology to realize high availability in distributed systems. It has been widely used in many distributed application systems, such as network communication protocols, distributed algorithms, group communication protocols, group membership management protocol, computer cluster management and grid computing and it is always used to solve coherence and atomic multicast issues, etc. which are fundamental issues in distributed systems.In the actual network environment, the systems are relatively large and the network situation is complex, so the traditional failure detection method which is realized by setting a fixed time-out value can not meet the QoS of the system requirements. In the adaptive failure detection algorithms, predicted theoretical time-out values are obtained by calculating and analysing the historical records of message delay. The time-out values, which are used to judge that whether the next message is failed or not, can change with the actual message delay. So the adaptive failure detection algorithms have strong adaptability.A double detection model based on message delay prediction is proposed in this article and a PA-AFD adaptive failure detection algorithm is proposed on the base of this model. The PA-AFD algorithm uses the Zipfs law to weight-calculate the time-out value and then judges that whether a message is failed or not.The failure detection level which the PA-AFD algorithm satisfies is proved in this article. Experimental results and analysises show that PA-AFD algorithm has detection accuracy and detection completeness and can ease the impact of network delay on failure detection.