PowerCap: Leverage Performance-Equivalent Resource Configurations for power capping

Abstract

Cloud providers use power under-provision policy to reduce the cost of infrastructure. Power capping technology is one of common approaches of power under-provision mechanism. It limits the maximal power consumption of server. DVFS and CPU capping are widely used approaches for power capping. However, they bring performance degradation due to lower frequency and less CPU slices. Here, we discover that applications under different resource configurations have similar performance but distinct power consumption. We call this phenomenon as “Performance-Equivalent Resource Configurations (PERC)”. Based on PERC, we propose a heuristic algorithm PowerCap to reduce power consumption when total power dissipation of server exceeds power budget. The basic idea of PowerCap is to reallocate resources by selecting configuration with the lowest power consumption from its PERC. Our experimental results show that PowerCap reduces power consumption below power budget with 50% less application reconfiguration and 12% more power comparing to DVFS approach. It also has none performance loss comparing to original configuration.