MLFD: The Implementation and Performance Evaluation of an LSTM-based, SmartNIC-Offloadable Failure Detector

This paper explores the feasibility of using machine learning algorithms for failure detection services. Our implementation and profiling results show that a DPDK-based failure detector (FD) using long short-term neural network performs well in terms of accuracy, at the cost of reasonable additional computation resources for training and inference. We also show that our implementation can be offloaded to a mainstream SmartNIC-NVIDIA BlueField-2, while retaining comparable quality of service. The viability of offloading MLFD to SmartNIC further frees up the concern that it will take up too much computation resources.