Abstract: Abstract: The purpose of the multiple data centers task scheduling strategy is to allocate computing tasks to different servers in each data center to promote resource utilization and energy efficiency. A deep reinforcement learning-based integrated scheduling strategy for multiple data centers is proposed, which is divided into two stages: data center selection and task allocation within data centers. In the multiple data center selection stage, the scheduling strategy integrates computing power resources to improve the overall resource utilization, firstly, a Deep Q Network with Prioritized Experience Replay (PER-DQN) was used to obtain the communication paths to each data center network in the data center node network, then the resource use cost and network communication cost were calculated, and the optimal data center was selected according to the principle that the sum of the two costs is minimum. In the task allocation phase, the tasks in the data center were divided and the tasks were added to the scheduling queue according to the First-Come-First-Served (FCFS), the task assignment algorithm based on Double Deep Q Network (Double DQN) was used to get the optimal assignment strategy, which can select the server to perform the computing task, avoid hot spots and reduce the energy consumption of refrigeration equipment. Experimental results show that the average total cost of PER-DQN-based datacenter selection algorithm is reduced by 3.6% and 10% compared with the Computing Resource First (CRF) and Shortest Path First (SPF) path selection methods, respectively. Compared with Round Robin scheduling (RR) and greedy scheduling (Greedy), the Double DQN-based task deployment algorithm reduces the average Power Usage Effectiveness (PUE) by 2.6% and 1.7% respectively. This strategy can effectively reduce the total cost and data center energy consumption, and realize the efficient operation of multiple data centers.

Key words: deep reinforcement learning, multiple data centers, task scheduling, temperature aware, power usage effectiveness

摘要： 摘 要: 多数据中心任务调度策略的目的是把计算任务分配到各个数据中心的不同服务器上，以促进资源利用率和能效的提升。提出了基于深度强化学习的多数据中心一体化调度策略，分为数据中心选择和数据中心内部任务分配两个阶段。在多数据中心选择阶段，整合算力资源以提高总体资源利用率，首先采用具有优先经验回放的深度Q网络(PER-DQN)在以数据中心为节点的网络中获取到达各个数据中心网络的通信路径；然后对资源使用成本和网络通信成本进行计算，并依据这两个成本之和最小的原则，选择最优的数据中心。数据中心内部任务分配阶段，在所选数据中心内部，对计算任务进行划分并遵循先来先服务原则(FCFS)，将任务添加到调度队列中，结合计算设备状态和环境温度，采用基于双深度Q网络(Double DQN)的任务分配算法获得最优分配策略，以选择服务器执行计算任务，避免热点产生，降低制冷设备能耗。实验结果表明，基于PER-DQN的数据中心选择算法分别相比于计算资源优先(CRF) 、最短路径优先(SPF)路径选择方法的平均总成本下降3.6%、10%；基于Double DQN的任务部署算法分别相比于较轮询调度(RR)、贪心调度(Greedy)的平均电源使用效率(PUE)下降2.6%、1.7%。该策略能够有效降低总成本和数据中心能耗，实现多数据中心的高效运行。

关键词: 深度强化学习, 多数据中心, 任务调度, 温度感知, 电源使用效率