Flexible CPU provisioning in clouds: A new source of performance unpredictability

Flexibility is an important characteristic of cloud computing: users can dynamically lease resources only when needed. While this feature brings along various benefits, there are several drawbacks and among them performance unpredictability is a major issue. In this paper we investigate the effects of a new source of variability introduced by the flexible CPUs allocation policy supported by several IaaS providers. According to this policy, CPUs with different computing power may be allocated to the same VM type. We will show through experiments on Amazon EC2 platform that the range of CPUs, allocated transparently to the users, may lead to very different performance experienced by an application executed with the same VM type. To perform our studies we developed a benchmark application that can either run with a predetermined workload, or scale the service demand according to the computational power of the allocated CPU. We also executed some benchmarks of the well known DaCapo suite, that is composed by real world open source applications. The collected measurements illustrate that the fluctuations of performance due to the flexible allocation policy may be very high and unpredictable.