Cloud computing is a computing method based on the Internet designed to share resources through virtualization technology. For a large number of requests waiting to be processed, task scheduling is used to reasonably allocate computing resources to requests. With the rapid development of computer hardware and software, deep reinforcement learning (DRL) provides a new direction for better solving task scheduling problems. In this paper, we propose a novel DRL-based dynamic load balancing task scheduling algorithm under service-level agreement (SLA) constraints to reduce the load imbalance of virtual machines (VMs) and task rejection rate. First, we use the DRL method to select a suitable VM for the task and then determine whether to execute the task on the selected VM violates the SLA. If the SLA is violated, the task is refused and feedback a negative reward for DRL training; otherwise, the task is received and executed, and feedback a reward according to the balance of the VMs load after the task is executed. Compared with three other task scheduling algorithms applied to randomly generated benchmark and Google real user workload trace benchmark, the proposed algorithm exhibits the best performance in balancing VMs load and reducing the task rejection rate, improving the overall level of cloud computing services.

云计算是一种基于Internet的计算方法，旨在通过虚拟化技术共享资源。对于大量等待处理的请求，使用任务调度为请求合理分配计算资源。随着计算机硬件和软件的飞速发展，深度强化学习（DRL）为更好地解决任务调度问题提供了新的方向。在本文中，我们提出了一种在服务级别协议（SLA）约束下基于DRL的动态负载平衡任务调度算法，以减少虚拟机（VM）的负载不平衡和任务拒绝率。首先，我们使用DRL方法为任务选择合适的VM，然后确定是否在所选VM上执行任务违反了SLA。如果违反了SLA，任务被拒绝，并且对DRL培训反馈了负面奖励；否则，接收并执行任务，执行任务后，根据虚拟机负载的平衡反馈奖励。与应用于随机生成基准和Google实际用户工作负载跟踪基准的其他三种任务调度算法相比，该算法在平衡VM负载和降低任务拒绝率，提高云计算服务的整体水平方面表现出最佳性能。