Dynamically forecasting timeouts for group membership in WANs

In this paper, a novel method is presented for improving the performance of group membership in WANs. For avoiding an extraneous and excessive long timeout used in failed member detections, we choose timeouts dynamically according to the predicted network performance. With the excellent learning and generalizing capabilities, RBFN is used to build the timeout forecasting model. For evaluating its performance, a purely auto-regressive (AR) model is constructed for comparison. The testing result has shown that the prediction of our RBFN model is more accurate. Based on the forecasting model, multi-cycle detection algorithm is designed to keep membership from fluctuating in unstable network situation. Good membership performance can be obtained by our service, including system scalability, failure detection accurateness and group sensitivity to changed member states.