A popularity-based approach for effective Cloud offload in Fog deployments

Recent research has put forward the concept of Fog computing, a deported intelligence for IoT networks. Fog clusters are meant to complement current cloud deployments, providing compute and storage resources directly in the access network which is particularly useful for low-latency applications. However, Fog deployments are expected to be less elastic than cloud platforms, since elasticity in Cloud platforms comes from the scale of the data-centers. Thus, a Fog node dimensioned for the average traffic load of a given application will be unable to handle sudden bursts of traffic. In this paper, we explore such a use-case, where a Fog-based latency-sensitive application must offload some of its processing to the Cloud. We build an analytical queueing model for deriving the statistical response time of a Fog deployment under different request Load Balancing (LB) strategies, contrasting a naive, an ideal (LFU-LB, assuming a priori knowledge of the request popularity) and a practical (LRU-LB, based on online learning of the popularity with an LRU filter) scheme. Using our model, and confirming the results through simulation, we show that the LRU-LB achieves close-to-ideal performance, with high savings on Cloud offload cost with respect to a request-oblivious strategy in the explored scenarios.