A game theoretical method for auto-scaling of multi-tiers web applications in cloud

Abstract

Cloud computing is a newly emerging reliable and scalable paradigm in which customers pay for cloud resources they use on demand. However, current auto-scaling mechanisms in cloud lack the critical self-adaption policy which helps application providers decide on when and how to reallocate resources. Furthermore, virtualization techniques can not ensure an absolute isolation between multiple virtual machines sharing the same physical resource, which leads to some customers paying unfairly for heavy-loaded resource under a widely-adopted fixed pricing scheme. In this paper, we present a global performance-to-price model based on game theory, in which each application is considered as a selfish player attempting to guarantee QoS requirements and simultaneously minimize the resource cost. Then we apply the idea of Nash equilibrium to obtain the appropriate allocation, and an approximated solution is proposed to obtain the Nash equilibrium, ensuring that each player is charged fairly for their desired performance. First, each player maximizes its utility independently without considering the placement of virtual machines. Then based on the initial allocation, each player reaches its optimal placement solely without considering others' interference. Finally we propose an evolutionary algorithm which step by step updates the global resource allocation based on the initial optimal allocation and placement.