Forecasting Heartbeat Delay for Failure Detection over Internet Using Nonlinear System

To overcome Internet dynamic characteristics and accurately predict next heartbeat message delay for failure detection service, a novel learning machine is proposed to predict next heartbeat arrival time. We use a nonlinear autoregressive network with exogenous inputs to learn nonlinear and linear characters of heartbeat messages, perform one-step-ahead prediction to estimate future heartbeat delay. The inputs are two moving window observations of past heartbeat delays and heartbeat sending time, the output is next heartbeat delay, the network is trained by standard back-propagation algorithm, its weights and basis are adjusted by approximate steepest descent rule. Simulation result shows that this adaptive algorithm can accurately capture heartbeat dynamics over Internet and make minimum prediction error under different network environments such as bottleneck link, link down and up.