GNN-Based Deep Reinforcement Learning for Dynamic Thermal-Aware Task Scheduling in Sustainable Data Centers

With the in-depth research of data centers, the task scheduling for sustainable data centers has garnered significant attention. Since thermal dissipation power occupies half of the energy consumption of a whole data center, a large number of sustainable data centers use Thermal-Aware Task Scheduling (TATS). However, existing approaches struggle to perform well in TATS, by ignoring the dynamic characteristics of the thermal in sustainable data centers, which is highly related to the temporal and spatial relationship between servers. It might cause inaccurate scheduling timing and poor scheduling locations, which will lead to additional scheduling costs and cooling costs. To address this problem, we propose a dynamic TATS with GNN-based deep reinforcement learning (DRLGTS) for sustainable data centers. Specifically, DRLGTS first evenly distributes the tasks according to the spatial positions of servers in data center, which is the preprocessing step of DRLGTS. Then, by calling the TranSimmethod, a dynamic thermal graph generation method based on thermal transient simulation in the entire task execution process is conducted. The thermal graph generated by TranSim represents the operating status of the servers and the temporal and spatial correlations between the servers as a graph layout. Finally, based on the feature extraction of thermal graphs using GNN, the proposed dynamic thermal-aware task scheduling using deep reinforcement learning (DRL) is executed. DRLGTS minimizes the energy consumption cost and time cost while scheduling task to adjacent servers as much as possible. Experimental results show that this architecture has good effectiveness and dynamic adaptability.