Autonomic power management with self-healing in server clusters under QoS constraints

The increasing use of server clusters has made their energy consumption an important issue. To address it, several power management techniques are being developed. In order to be useful, these techniques must address the performance and availability implications of reducing energy consumption. This paper presents a power management technique that maintains the quality of service (QoS) levels specified with service level agreements expressed as a threshold for a percentile of the response time. In addition, it provides self-healing by identifying when servers fail and automatically provisioning new servers. The technique is based on balancing the load so that it is concentrated in a small number of servers. For this, it only requires two utilization thresholds and models of performance and power consumption for the application executed in the server. It works in heterogeneous servers and provides overload protection. Several experiments carried out on a prototype show that the technique reduces energy consumption (up to 57.59 % compared to an always-on policy) while providing self-healing and maintaining the QoS.