D-PALD: A Dynamic Power-Aware Load Dispatcher with Response Time Percentile Guarantee in Heterogeneous Clusters

Abstract

The power consumption of a server is not linear to its activeness, i.e., a server with 10% load may still draw as much as 60% of its peak power, therefore, a significant amount of energy has been used to keep servers active even under very light or idle loads. The resulting effect has been increased low power-effectiveness in data centers which elevate ownership costs and put more pressure on rack and enclosure densities. This motivates us to design a scheme to dynamically dispatch load to achieve energy efficiency while maintaining the quality of service (QoS) by exploiting a fundamental characteristic of data centers: heterogeneity. In this work, we propose D-PALD to guarantee response time percentile while dynamically dispatch the workload among the servers to maximize the power efficiency in heterogeneous data centers. We develop a power efficiency model to characterize properties of servers while providing the percentile guarantee and also design a power aware load dispatching algorithm. Our experiments demonstrate that DPALD can save a significant amount of power without sacrificing user performance compared to the baseline power management algorithms.