Resource Management And Scheduler System On Private Cloud Paas Service Based On Container

As an excellent container orchestration and management system,Kubernetes has been widely used to construct private cloud Paa S service based on container.There are many kinds of applications running on the private cloud,and cooperation may exist among these applications because of their dependencies on others.However,the default scheduling algorithm of Kubernetes does not pay attention to the cooperation among applications,leading to that the platform cannot guarantee Qo S of the applications.Faced with such a problem,this paper tasks the actual project as research background.After analyzing container technique and computing resource management technique,this paper designs and implements a computing resource management system based on the open source project Kubernetes.This system consists mainly of an event-driven resource monitor module,a scheduling module based on game theory and a resource utilization prediction module based on time series analysis.The experiment shows that the system is viable and valid.The main work of this paper includes:1)In Kubernetes,every status message of a container contains the resource status information of its host machine,and the average changing frequency of resource status of host machine is much slower than query frequency of container status,leading to that repeated host resource status information jams the network.After deeply analyzing the instant demand of container status and the characteristic of host status changing with events,this paper proposes an event-driven resource monitor model called Ed RMM,then designs and implements a monitor module based on Ed RMM model.The experiment shows that Ed RMM model can decrease 52% transmission data on average on the 3 tests.2)The default scheduler of Kubernetes cannot guarantee the cooperation constraints among applications,which may cause the QoS of applications decrease and even make these applications unavailable.To solve the problem,this paper proposes a kind of scheduling algorithm called Gm Sch based on the maximum profit in game theory.This algorithm analyzes the cooperation constraints of applications and uses cooperative game theory to allocate resource for applications to gain better Qo S meanwhile guarantees the resource utilization of the system.In the experiment,an index calculated by the normalized response time and priority is used to represent the Qo S of applications.Compared with the default scheduler of Kubernetes,GmSch can deal with the cooperation and improve 25.6% QoS.3)The resource demand of applications may fluctuate with time,and static resource allocation may lead to resource wastage or performance decrease.Besides,dynamic allocation on instant demand will introduce latency.This paper proposes a resource utilization prediction algorithm called CRUPA based on ARIMA time series model.CRUPA uses data collected by the monitor module to train a prediction model,and uses this model to predict the resource utilization of application in near feature,then scales the resource allocated to the application.The experiment shows that the average deviation of CRUPA algorithm on the test dataset is 6.9% and its accuracy is higher than the widely used prediction method based on threshold.