Application type awareness pod-level and system-level container scheduling

Abstract

Kubernetes, as a powerful tool for managing containerized applications, is considered a promising tool for supporting cloud computing platforms. The default scheduling scoring strategy only considers seeking an optimal node for the current pod and ignores the availability of subsequent nodes. Additionally, the node with the highest overall score may not necessarily be the most suitable node for the current task when the scoring process is performed. Therefore, a new Container scheduling strategies based on application type awareness at the pod and system levels (ATASL) is proposed. Firstly, in order to address the computational waste caused by the need to sequentially traverse and score all nodes in traditional node filtering methods, ATASL binds labels for Pods and nodes based on the required resources of Pods and the remaining resources of nodes, corresponding to “Compute” and “Memory”. So the subsequent scheduling of Pods is restricted to the corresponding groups of nodes only, avoiding the traversal of all nodes for scoring. Moreover, before scheduling each new task, ATASL adjusts the node roles to accommodate dynamic load changes based on the real-time resource status of the nodes. Secondly, when calculating the node score, not only the Pod-level score that matches the resource demand of the Pod is considered, but also the “system penalty score” mechanism is introduced to avoid the performance bottleneck caused by the over-utilization of a certain resource. This mechanism imposes a penalty on nodes where the utilization of a particular resource significantly exceeds the overall average utilization of the cluster, preventing resource imbalance and performance degradation (i.e., preventing overburdened nodes from being selected). Finally, a Kubernetes cluster was built using VMware to evaluate system performance. The experimental results show that ATASL can significantly improve the overall throughput and system resource utilization of the cluster, and also lead to a substantial improvement in node balance.