Design And Implementation Of Container Cloud Platform Component Supporting Accelerator Card

With the rapid development of artificial intelligence,the demand for computing resources in related applications is accelerating,more and more applications will be deployed in the cloud.Cloud computing technology centralizes and unifies the management of computing resources,storage resources,and network resources,renting them out on demand,greatly reducing the threshold for users to access computer resources.Container technology represented by Docker containers has been increasingly chosen by cloud service providers as the virtualization technology for cloud platforms due to its advantages of low resource consumption,fast startup speed,and high resource utilization.Container orchestration technology represented by Kubernetes automates container deployment and management.At present,the cloud platform solution that combines Docker container and Kubernetes has been widely applied,greatly improving the efficiency of research and development operations.The accelerator card mentioned in this thesis is a heterogeneous computing device used to accelerate neural network computing and brain like computing.However,native Kubernetes and Docker do not manage heterogeneous computing resources such as accelerator cards,and cannot incorporate accelerator card resources into the management of container cloud platform to allocate and schedule accelerator card resources.In addition,through the analysis of the scheduling strategy of the Kubernetes scheduler,it was found that the default scheduling strategy of Kubernetes still has limitations in resource scheduling and cannot cope with user needs in heterogeneous and complex scenarios.In response to the above existing problems,this article mainly conducts research from the following aspects:1.On the basis of investigating the device loading principle provided by Docker for heterogeneous computing devices,the Docker adapter component of the accelerator card is designed and implemented,which enables the Docker container to support and manage the accelerator card,so that applications can use the specified accelerator card in the Docker container to accelerate computing.2.On the basis of investigating the device loading principle provided by Kubernetes for heterogeneous computing devices,the Kubernetes adaptation component of the accelerator card is designed and implemented,which enables Kubernetes to support and manage the accelerator card,so that Kubernetes can discover and monitor the accelerator card,which is a heterogeneous computing device,and incorporate the accelerator card resources into Kubernetes’ management,laying the foundation for Kubernetes to schedule the accelerator card resources.3.Researched and analyzed the resource scheduling strategy of Kubernetes,and analyzed the limitations of Kubernetes’ default scheduling strategy.Then,facing various resource scheduling scenarios in Kubernetes,design a DDRF optimal scheduling strategy based on DRF strategy.This strategy schedules Pods based on the proportion of dominant and sub dominant resources of the corresponding nodes,reducing resource fragmentation caused by imbalanced resource utilization within cluster nodes and maximizing cluster resource utilization.The scheduling effects of Kubernetes default scheduling strategy,DRF strategy,and DDRF strategy were tested by simulating the Pod application resource volume and cluster node resource volume of actual production clusters.The experimental results showed that DDRF strategy can improve the balance of multiple resource utilization rates within Kubernetes cluster nodes,reduce node resource fragmentation,and improve cluster resource utilization.Finally,tested the Docker adaptation component of the accelerator card and the Kubernetes adaptation component of the accelerator card.Verified that the application that needs to use the accelerator card can operate normally in the Kubernetes cluster where the Docker adaptation component of the accelerator card and the Kubernetes adaptation component of the accelerator card are deployed.