VM Scaling and Load Balancing via Cost Optimal MDP Solution

Dynamic resource allocation mechanism is an essential building block in contemporary cloud computing environment, enabling the support of large variability incoming requests from enormous number applications utilizing such infrastructure. In this article, we devise a dynamic that optimizes application’s profit under set costs and revenues while maintaining performance constraints. Specifically, decision-maker (DM) agent which formulates joint admission control, scaling load balancing problem as stochastic process solvable by Markov decision (MDP) provides optimal policy. Accordingly, at each time instance, DM can determine based on system’s current state, requirements costs, whether to add or release VM (scale-out scale-in, respectively), admit reject upcoming task if admitting it, allocate it to. We explore value function structure provide insights with respect policies produced it. To address scalability issues detailed MDP solution alternative abstract consolidates multiple system states into single hence cope much larger systems expense slight degradation. demonstrate feasibility suggested scheme, designed implemented alongside two traditional auto-scalers, Amazon Web Services (AWS) ran numerous MATLAB simulations AWS-based experiments provided demonstrated superiority against compared with.