Abstract:
Fog computing is an extended cloud computing technology allowing services embedded into virtual machines or containers to be placed at the edge with closer proximity to the end devices. Nevertheless, one of the main difficulties adding up to the complexity of the on-demand fog placement topic is deciding on the proper time and place of the fog deployment process, and deriving the adequate container and service distribution over the available fogs at a certain location. Several techniques have been suggested in the current literature as potential solutions,
in which some consider users preferences or random-based approach for
fog deployment, while others reside on threshold-based mechanisms. However, due to the huge increase in the number of requests coming from end devices including IoT users, fog deployment and container placement must be scheduled to serve locations with high service requesting profiles while decreasing the cloud processing load. In this context, we first propose a fog placement model that allows to produce an adequate scheduling decision by using a hybrid technique combining time series forecasting and reinforcement learning. The proposed technique learns the intensity of service invocations and behavior of end devices at different locations over time for predicting the localization plan. Second, we
propose a K-Means based clustering model embedded within a multi-objective optimization scheme for fog and container placement. A comparison of our proposed solution with random-based and threshold-based fog placement scheduling approaches show that the number of processed requests performed by the cloud decreases from 100% to 29% compared to 93% and 67% in the other two models.
The aforementioned results explore the efficiency of our proposed scheme in scheduling the fogs in their rightful place, which helps in decreasing the cloud’s load, decreasing the network congestion, and increasing the overall quality of service.
Moreover, experimental results illustrate that the proposed optimization
model and clustering technique further improve the pre-existing heuristic-based solutions for container distribution.
