FDKeeper

Abstract

Failure detection is a basic service for building dependable cloud computing system. Unfortunately, the natural complexity of cloud computing system makes failure detectors much harder to build. In this work, we present the design, implementation, and evaluation of FDKeeper, a failure detector for cloud computing system with several features. First, FDKeeper is a failure detector with short detection time, high accuracy and low disruption. Second, FDKeeper is scalable, with communication cost and number of manager nodes scaling linearly with the number of physical nodes, which makes FDKeeper very suitable for large-scale cloud computing system. Third, FDKeeper provides a protocol for failure detection among many cloud computing systems. FDKeeper achieves these features by providing a layered failure detection mechanism, a cluster architecture and a cloud failure detection protocol. The evaluation results show that the detection time for app/kernel layer is less than 600 milliseconds, and 2 seconds for vmm/hardware layer, the accuracy is 100% for all layers, and the disruption is less than 0.27 times per year. As for scalability, the theoretical analysis and experimental results show that there is a linear relationship between the number of nodes used to accomplish the failure detection and the number of nodes in the computing systems, and the growth factor is less than 0.001. Thus we believe FDKeeper is the first failure detector that is fast, reliable, scalable, and viable. As such, it could change the way that a dependable cloud computing system is built.