边缘场景下基于DDQN的容器组调度策略

Container Group Scheduling Optimization Strategy Based on DDQN in Edge Scenarios

工业互联网中存在大量部署于边缘服务器上的在/离线容器服务,这些容器服务一方面承载着低延时,高响应的需求,另一方面又具有错综复杂的调用关系。通常边缘集群的调度策略并未考虑到容器服务之间的依赖关系,这导致具有依赖关系容器服务可能在调度过程中被分散到不同的边缘节点上,并由此产生大量跨节点调用造成额外资源损耗。针对具有依赖关系的容器,该文提出面向边缘场景的容器组调度优化策略。首先通过容器聚类算法CDSC(Container Dependency Spectral Clustering)将有依赖关系的容器划分为一个或多个容器组,使得组内容器依赖强度尽可能的大,组间依赖强度尽可能的小,以减少其进行跨节点调用的频率;再通过引入双深度Q网络模型(Double DQN)将容器组作为基本调度单位,以容器依赖开销,集群和节点内部负载为优化目标,根据边缘节点实际情况自适应学习优化调度策略,使其能应对复杂多变的边缘集群情况。经实验表明,相比于传统的启发式算法和其他深度强化学习算法,该算法在容器服务响应时间、集群和节点负载方面具有明显的优势。

The industrial Internet is populated with a large number of on/offline container services deployed on edge servers.On the on hand,these container services bear the demand for low latency and high response,and on the other hand,they have intricate invocation relationships.The usual scheduling strategies for edge clusters do not take into account the dependencies between container services,leading to dependent container services possibly being dispersed across different edge nodes during scheduling,thereby generating a large number of cross-node calls and causing additional resource loss.We propose an optimization strategy for container group scheduling in edge scenarios for containers with dependencies.Firstly,the CDSC(Container Dependency Spectral Clustering)is used to divide dependent containers into one or more container groups,maximizing the dependency strength within groups and minimizing it between groups,to reduce the frequency of cross-node calls.Then,by introducing the Double Deep Q-Network model(Double DQN),the container group is used as the basic scheduling unit,with container dependency overhead,cluster and intra-node load as optimization targets.The strategy adaptively learns and optimizes scheduling strategies according to the actual situation of edge nodes,enabling it to cope with complex and changing edge cluster situations.Experimental results show that compared to traditional heuristic algorithms and deep reinforcement learning algorithms,the proposed algorithm has significant advantages in terms of container service response time,cluster and node load.