Characterizing energy consumption of IaaS clouds in non-saturated operation

To save energy, physical machines (servers) in cloud data centers are partitioned in different pools, depending on whether they are kept on at all times and/or whether they have virtual machines instantiated. Partitioning (pooling) of servers affects the power consumption of the data center but also the performance and responsiveness to user requests. In this paper we examine the behavior of pool management scheme in different operating regions which correspond to linear operation, transition to saturation, and saturation, in particular the tradeoff between performance and power consumption. In addition, our results show that the offered load (analogous to the concept often used in networking) does not offer complete characterization of data center operation; instead, the impact of task arrival rate and task service time must be considered separately.