Cost-Efficient, Reliable, Utility-Based Session Management in the Cloud

We present a model and system for cost-efficient and reliable management of sessions in a Cloud, based on the von Neumann-Morgenstern utility theorem. Our model enables a web application provider to maximize profit while maintaining a desired quality of service. The objective is to determine whether, when, where, and how long to store a session, given multiple storage options with various properties, e.g. cost, capacity, and reliability. Reliability is affected by three factors: how often session state is stored, how many stores are used, and how reliable those stores are. To account for these factors, we use a Markovian reliability model and treat the valid storage options for each session as a von Neumann-Morgenstern lottery. We proceed by representing the resulting problem as a knapsack problem, which can be heuristically solved for a good compromise between efficiency and effectiveness. We analyze the results from a discrete-event simulation involving multiple session management policies, including two utility-based policies: a greedy heuristic policy intended to give real-time performance and a reference policy based on solving the linear programming relaxation of the knapsack problem, giving a theoretical upper bound on achievable utility. As the focus of this work is exploratory, rather than performance-based, we do not directly measure the time required for solving the model. Instead, we give the computational complexity of the algorithms. Our results indicate that otherwise unprofitable services become profitable through utility-based session management in a cloud setting. However, if the costs are much lower than the expected revenues, all policies manage to turn a profit. Different policies performed the best under different circumstances.