The amount of data that are computed, stored, and shared over the Internet is rising at an unprecedented scale. This has necessitated more servers to be deployed and drastic improvement in the capacity of individual servers. However, several independent studies also reveal that resources are not optimally utilised in most existing data centres and server clusters. Since the introduction of server virtualization and cloud computing, the research community has proposed several workload aggregation and dynamic consolidation techniques, however, most of these techniques are either theoretical and rely on simulation environments or use real servers but static workloads or benchmarks. In reality, the workload of data centres fluctuates as a function of time and servers frequently experience both overloading and underutilised conditions. In this paper we introduce the HAECubie demonstrator we developed and deployed to experimentally evaluate the scope and usefulness of dynamic workload consolidation in a server cluster and to quantitatively analyse the relationship between energy/power consumption and the utility (performance) that can be achieved through workload consolidation. Our demonstrator is a video hosting platform and enables Internet users to stream videos of variable length. The number of users accessing the HAECubie as well as the duration of videos they stream are modelled as stochastic processes based on realistic estimation of the workload of existing video hosting platforms.

The power consumption of presently available Internet servers and data centers is not proportional to the work they accomplish. The scientific community is attempting to address this problem in a number of ways, for example, by employing dynamic voltage and frequency scaling, selectively switching off idle or underutilized servers, and employing energy-aware task scheduling. Central to these approaches is the accurate estimation of the power consumption of the various subsystems of a server, particularly, the processor. We distinguish between power consumption measurement techniques and power consumption estimation models. The techniques refer to the art of instrumenting a system to measure its actual power consumption whereas the estimation models deal with indirect evidences (such as information pertaining to CPU utilization or events captured by hardware performance counters) to reason about the power consumption of a system under consideration. The paper provides a comprehensive survey of existing or proposed approaches to estimate the power consumption of single-core as well as multicore processors, virtual machines, and an entire server.