Explicitly Consider Server-Attached Fans for Thermal Modeling in Edge Data Centers

Abstract

Edge data center has become an important data center type in recent years. Comparing to a conventional data center, an edge data center often lacks sophisticated cooling equipment and infrastructure support. In the resulting poor thermal environment, fans attached to individual servers have to work harder in an edge data center than those in a conventional data center. Research have shown that power generated by server-attached fans are quite significant to be ignored from thermal standpoint when fan speed is high. In this paper, as we consider power efficiency in edge data centers, we argue that power generated by fans attached to the servers should be explicitly considered for thermal modeling in the overall thermal optimization framework. We propose a design that incorporates fan power in a neural network to better predict a server's thermal state in an edge data center. We further propose a task scheduling algorithm that utilizes the improved neural network to enhance an edge data center's overall power efficiency. Based on the experimental results from a field edge data center, the improved neural network achieves better accuracy in predicting individual server's thermal state, outperforming other neural networks on precision. The proposed task scheduling algorithm, powered by the improved neural network, saves as much as 11% power consumption comparing to unoptimized algorithms.