FaCS: Toward a Fault-Tolerant Cloud Scheduler Leveraging Long Short-Term Memory Network

Large-scale cloud datacenters often experience reduced performance and service outage. Due to the inherent complexity, heterogeneity, and multitenant architecture of these datacenters, applications (i.e., jobs and tasks) running on them are susceptible to various types of failures. In this paper, we first characterize the application failures in Google cluster trace and then propose a prediction model which can forecast the termination status of a task. Then, we introduce a task scheduler that dynamically reschedules tasks based on the predicted results. This proactive fault-tolerant scheduler improves system reliability and ensures timely execution of the applications. Simulation results show that our scheduler reduces makespan and failure rates of tasks substantially while balancing load at the same time. Moreover, early prediction along with quick scheduling adjustment improves overall resource utilization and reduces resource wastage.