Availability, as a part of Service Level Agreement (SLA), is a critically important issue in cloud services, as an application may not be able to run after certain server or network failures. Cloud service providers seek to not only fulfill the SLA, but also simultaneously minimize their operating costs, which are dominated by the energy consumption. In order to minimize the impact of a server/switch failure inside the datacenter on a single application, one would like to spread out the Virtual Machines (VM) for the application across different racks. However, in doing so, the power consumption may increase significantly. In this paper, we develop a variance-based metric to measure the risk of violating the availability requirement. We then propose two heuristic algorithms to place VMs in online and offline manners, respectively. These algorithms aim to strike a balance between minimizing the risk of violating the availability requirement and minimizing the energy, in order to reduce the overall cost.

A central challenge in developing cloud datacenters Service Level Agreements is the estimation of downtime distribution of a set of provisioned servers over a service window, which is compounded by three facts. First, while steady-state probabilities have been derived for birth-death processes involving server failures and repairs, they could be highly inaccurate under transience. Furthermore, steady-state cannot be assured under typical service windows. Therefore, estimation of transient distributions is essential. Second, the processes of failures and repairs may follow any distribution and hence need to be extracted using system log data and modeled using appropriate general distributions. Third, downtime distributions over service windows depend on the number of servers and their deployment structure for a contract. We develop an efficient and generalized sample path randomization approach to precisely estimate transient probabilities under three different checkpointing strategies and three flexible failure distribution models. The estimators are unbiased, consistent, efficient and sufficient. Their asymptotic convergence is established. The estimation algorithms are computationally efficient in solving practical problems and yield rich information on transient system behaviors. The methodology is general and extensible to various server failure and repair processes characterized using birth-death modeling.