Edge computing paradigm is prone to failures as it trades reliability against other quality of service properties such as low latency and geographical prevalence. Therefore, software services that run on edge infrastructure must rely on failure resilience techniques for uninterrupted delivery. Unique combination of hardware, software, and network characteristics of edge services is not addressed by existing techniques that are designed or tailored for cloud services. In this work, we propose a novel method for evaluating the resilience of replicated edge services, which exploits failure dependencies between edge servers to forecast probability of service interruption. This is done by analyzing historical failure logs of individual servers, modeling temporal dependencies as a dynamic Bayesian network, and inferring the probability that certain number of servers fail concurrently. Furthermore, we propose two replica scheduling algorithms that optimize different criteria in resilient service deployment, namely failure probability and cost of redundancy.

Top