Owing to the potential for fast deployment, containerization technology has been widely used in web applications based on microservice architecture. Online container placement aims to improve resource utilization and meet other service quality requirements of cloud data centers. Most current heuristic and hyper-heuristic methods for container placement rely on single allocation rules, which are inefficient in heterogeneous cluster scenarios. Moreover, some container placement tasks often have similar characteristics (e.g., resource request types and physical machine types), but traditional single-task optimization modeling cannot exploit potential common knowledge, resulting in repeated optimization during resource allocation. Therefore, a new multi-task genetic programming method is proposed to solve the online multi-objective container placement problem (MOCP-MTGP). This method considers selecting appropriate allocation rules according to the types of resource requests and cluster status. MOCP-MTGP can automatically generate multiple groups of allocation rules from historical workload patterns and different cluster states, and capture the similarities between all online tasks to guide the transfer of general knowledge during optimization. Comprehensive experiments show that the proposed algorithm can improve the resource utilization of clusters, reduce the number of physical machines, and effectively meet resource constraints and high availability requirements.

由于容器化技术具有快速部署的潜力，因此已广泛应用于基于微服务架构的 Web 应用程序中。在线容器放置旨在提高资源利用率，并满足云数据中心的其他服务质量要求。当前大多数用于容器放置的启发式和超启发式方法都依赖于单个分配规则，这些规则在异构集群场景中效率低下。此外，一些容器放置任务往往具有相似的特征（例如，资源请求类型和物理机类型），但传统的单任务优化建模无法利用潜在的常识，导致资源分配时出现重复优化。因此，提出了一种新的多任务遗传编程方法来解决在线多目标容器放置问题 （MOCP-MTGP）。该方法考虑根据资源请求的类型和集群状态选择合适的分配规则。MOCP-MTGP 可以从历史工作负载模式和不同的集群状态中自动生成多组分配规则，并捕获所有在线任务之间的相似性，以指导优化过程中的一般知识传递。综合实验表明，所提算法可以提高集群的资源利用率，减少物理机的数量，有效满足资源约束和高可用性要求。