Adaptive Failure Detection Model FD-LSSVR Based On Time Series Prediction

With the acceleration of the developmental process of the worldwide distributed systems and its more complex structure than before, more and more failure phenomena happen that seriously threats the high availability of systems. As one of the core technology of protecting the high availability of distributed systems, failure detection has gradually become a hot research direction. Recently failure detection has been widely used in communication protocols, Web server, cluster management and failure recovery etc. At the same time, failure detection has also been an important application in the fields of wireless network, cloud computing platform, and big data etc.Throughout the existing adaptive failure detection algorithm and in-depth research of the process of failure detection, we can understand that the nature of failure detection belongs to nonlinear time series prediction problem. The least squares support vector algorithm is one of the effective methods of nonlinear time series prediction, which has certain advantages in small sample space and forecast accuracy. The traditional adaptive failure detection algorithm is generally based on probability statistical model which is limited to some probability distribution and requires a large amount of data to calculate the arrival time of next heartbeat message, for what there are some limitations. While the use of least squares support vector regression algorithm for failure detection forecast just to make up for its shortcomings. Practical applications are often faced by small sample space, so this paper proposed and discussed FD-LSSVR (Failure Detection-Least Squares Sport Vector Regression) model, which also introduces clustering analysis method that considering the two indicators of heartbeat delays and weights to remove outliers that have a greater impact on predicted results for that the existing adaptive failure detection algorithms need to consider outliers. The weights assigned to meet the power-law distribution.The experimental results show that the model not only satisfies the OP detection level, but also performs great at both detection time and accuracy, for what it can be used to alleviate the sub-network delay for failure detection. In addition, according to actual situation, FD-LSSVR can meet the needs of a variety of different applications by adjusting the size of the parameter a values, with a certain degree of flexibility.