Research And Implementation Of Resource Scheduling Strategy Based On Kubernetes

Since the advent of Kubernetes in 2015,it has received very high evaluations in various professional communities.In just a few years,it has become the most used container orchestration system in the industry.It has very rich system features,but it can not meet all application scenarios in use.First,the default scheduling strategy of Kubernetes cannot evenly schedule applications based on network bandwidth resources,which cannot meet the scheduling requirements of network-intensive applications.Secondly,the default scheduler of Kubernetes will perform a complete pre-selection process and optimal process for all Pods in the same Replica Set.There is a lot of redundant work,scheduler performance will be affected in large clusters.Thirdly,the ResourceQuota configured for multi-tenant in the cluster is static,and the resource number that can be scheduled in user space cannot change with the change of cluster node resources.To address these problems,this article has done the following:1)Aiming at the problem that the default scheduling strategy of Kubernetes does not consider network bandwidth resources,this dissertation designs and implements an optimal scheduling algorithm BalancewithNetAllocation named BWNA that can evenly schedule applications based on network bandwidth resource.The experimental results show that the BWNA algorithm can evenly schedule the application based on network bandwidth resources and make full use of the network bandwidth of nodes in Kubernetes cluster,which meets the scheduling requirements of network-intensive applications.2)Aiming at the problem that the default scheduler of Kubernetes has a lot of redundant work when scheduling some Pods under the same Replica Set,this dissertation designs and implements a Pseudo-Parallel-Scheduling algorithm named PPS.The experimental results show that the PPS algorithm not only can realize pseudo parallel scheduling of Pods under the same Replica Set,but also can reduce the redundant workload and the average scheduling time of Pods in the scheduling process without affecting the normal operation of Pod.3)Aiming at the problem that the ResourceQuota of multiple tenants in the cluster cannot be changed dynamically with the change of cluster resources,this dissertation designs and implements a controller-DynamicQuota.The experimental results show that the controller can monitor the change of node resources in the cluster and dynamically modify the ResourceQuota of multi-tenants,so that the resource number that can be scheduled in user space can change dynamically.In summary,Kubernetes cannot meet the needs in some application scenarios.This dissertation designs and implements corresponding solutions to these problems and verifies the effectiveness of the solutions through experiments.