Quality of Service Channelling for Latency Sensitive Edge Applications

Edge data centers (EDCs) typically provide lower availability rates than Cloud counterparts since they lack expensive support systems such as air conditioning units and power generators. To avoid this limitation deteriorating response time which is critical for Edge applications, use of proactive optimization algorithms is essential. Such proactive algorithms, however, require an accurate method for estimating availability of a VM that runs on a certain EDC. Estimation is complicated by several dependent and independent probabilistic events (e.g. hardware, software or network failures, power outage, etc.) that affect VM availability in different levels. In this paper, we propose a Bayesian Network model of QoS related parameters to estimate availability level of a VM in Edge infrastructure. We compare our approach to other machine learning methods and to the common practice in reliability theory that is the use of prior failure probability. According to experimental results, proposed method can identify VMs that satisfy user defined availability objectives with up to 94% accuracy and decrease SLO violations by nearly 44%. In addition, we empirically investigate the trade-off between running time and accuracy of proposed approach.