Edge computing is a promising cloud computing paradigm that reduces computing latency by deploying edge servers near data sources and users,which is of great importance to implement delay-sensitive applications like AR,Cloud Gaming and Auto Driving.Due to the limited resources of edge servers,task dispatching and function configuration are the key to fully utilize edge servers.Moreover,a typical task request in edge computing(called a co-task)is consisted of a set of subtasks,where the task completion time is determined by the latest completed subtask.In this work,we propose a scheme named OnDisco,which combines reinforcement learning and heuristic methods to minimize the average completion time of co-tasks.Compared with heuristic algorithm,reinforcement learning can learn the inherent characteristics of the environment without any prior knowledge,and OnDisco is therefore well adapted to varying environments.Simulations on Alibaba traces shows that OnDisco reduces the average task completion time by 58%and 76%compared with the heuristic and random algorithm,respectively.Moreover,OnDisco outperforms the baselines consistently in various data environments and parameter settings.In addition,we also designed a scheduling plugin for the Kubernetes(K8S)container orchestration framework,which allows the complex customed scheduling algorithms to be easily deployed to the actual system.OnDisco reduces the task completion time by 20.1%and 20.8%compared with the baselines in K8S system. |