The high costs of infrastructure construction compel managers to optimize the power utilization of existing data centers. Even with capacity oversubscription applied, the power under-utilization problem still remains. Some approaches enhance resource utilization by leveraging idle resources, but they also increase the risk of overload.
To solve the problems, a research team from Emerging Parallel Computing Center (EPCC) published their new research on 15 May 2025 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.
The team proposed a fluctuation-aware power auction strategy to reduce the power overload probability while maximizing the power utilization. The strategy is verified and tested in a simulation testbed using real cluster traces from Google. Compared to the state-of-the-art power management algorithms, FLAPS reduces the power overload by 9%-10% on average while maintaining a resource utilization rate difference of less than 2.8%.
In the research, they identify the potential abrupt fluctuations in power usage and carry out targeted optimization during both the prediction phase and the resource allocation phase. The power usage prediction module improves the vanilla LSTM model by incorporating a feature selection layer. This addition highlights the decile features of CPU utilization that are most relevant to the prediction results, thereby enhancing the training convergence speed and inference accuracy of the prediction model. Upon acquiring accurate power data, the fluctuation-aware auction strategy takes power variance into account and prefers scheduling power-stable jobs for spot power.
The prediction module forecasts the power usage of different jobs based on historical records. Compared to the conventional features, decile features are more related to the average CPU usage. Our model introduces decile feature in two ways to maximize its impact. Based on the prediction results, the calculated idle power usage is then utilized in the resource auction phase. The auction algorithm takes into account both supply and demand, as well as the jobs’ load fluctuation characteristics. Jobs’ prices change exponentially based on their coefficient of variation. This approach can also be viewed as a penalty for jobs with high fluctuation. Experimental results show that the fluctuation-aware auction strategy proposed in this paper can maintain high resource utilization while avoiding a high overload ratio.
Future work could further explore the impact of workload characteristics on power resources in specific scenarios, including factors like workload granularity, virtualization level, and heterogeneity.
DOI: 10.1007/s11704-024-40184-5