Estimating Power Consumption of Collocated Workloads in a Real-World Data Center

Carbon-aware computing is emerging as one of the sustainable ways to meet the projected rise in IT energy demand. Carbon-aware computing methods require workload-level power consumption as one of the inputs. To this end, we conducted a novel study involving a real-world data center (DC) to attribute power consumption to individual workloads collocated on servers. In this paper, we propose a machine learning-based modeling approach to estimate server-level power consumption and then adapt the server-level model to estimate the power consumption of individual workloads running on the server. The proposed method aims to solve two problems: the diversity of workloads in a real-world data center concerning the magnitude of power consumed and the lack of ground-truth for workload-level power consumption. For the first problem, we use short-range power models instead of a single model for the entire power range, which reduces the estimation error by about 28 to 30 percentage points on average compared to the conventional method. We use a top-down modeling approach for the second problem, using dominant feature-based scaling to transform server-level model estimations into workload-level power attributions. According to our knowledge, this study is a first of its kind where we proposed a method to attribute power consumption to individual workloads in a real-world DC.