A Dynamic Scheduling Algorithm with Time Varying Resource Constraints in Colocation Data Centers

The colocation of online services and offline tasks has been widely used in data centers, which improves the resource utilization. Due to the interference between different types of workloads, scheduling co-located offline tasks becomes very complex and challenging. Existing colocation scheduling algorithms try to find the optimal combination of different workloads to avoid performance interference and maximize the utilization. Different from existing works, which usually assume the amount of resources for offline tasks is predefined and keeps unchanged, this paper takes the time varying resource constraints caused by dynamic changes of the workload of online services into consideration. We propose a dynamic scheduling algorithm based on Monte Carlo Tree Search(MCTS) and use the simulation of subsequent scheduling to obtain the best scheduling action at each moment. In order to handle the dynamics of resource constraints, we firstly extend MCTS by adding a slide window mechanism, and get a new search tree named Monte Carlo Tree with Slide Windows(MCTSW). With time changing, MCTSW reconstructs the scheduling tree to include new resource constraints, and the reconstruction can reuse subtrees so as to reduce overhead. The experiment results show that, compared existing works, our algorithm can obtain better scheduling plan and improve the resource utilization of the colocation system.