Machine learning applied to failure detection
Wenbin Qi, Jingyi Zhao · 2023
The operation of distributed systems highly depends on reliable and continuous communication between nodes. The failure detection phase thus plays an important role in system unavailability. Previous studies have shown potentials for LSTM based failure detectors but the heavy computation cost remains a problem. We propose a technique to reduce the computation time of LSTM based FD by performing forecasting instead of making one prediction at a time (i.e. predicting multiple timestamps at once). Our method achieves better computation time compared to previous LSTM FDs but falls short of BFD. We discuss the heavy computation cost of back-propagation of LSTM and the high frequency nature of FD tasks and conclude that LSTM isn’t likely to be the best fit for Machine Learning based FD.