Modeling price and demand in cloud computing systems

Abstract

The purpose of this thesis is to explore the behavior of pricing and demand within cloud computing systems. As cloud computing has become increasingly popular, research has been focused on finding a mechanism to optimize social welfare. Current schemes typically charge users at a fixed price per time unit. Theoretical research alternatives to this include combinatorial auctions. Analyses of these systems often utilize a simplistic model of a user. This thesis aims to introduce a more realistic user model and analyze optimization from the user's perspective in an environment similar to that of the Amazon EC2 spot instance market. It is shown that the optimal strategy of a user with a deadline and sequential load can be derived from a set of dynamic programming equations, the results of which can be implemented without knowledge of the current price. Additionally, it is shown that this strategy is closely approximated by an expected rate user model. Finally, it is shown that these types of users are elastic and the load on the system can be controlled through the price mechanism.