The article A Deep Dive into HighNodeUtilization and LowNodeUtilization Plugins with Descheduler discusses two Node utilization plugins of the descheduler in the Kubernetes community. Both of these plugins use requests to calculate the resource usage of nodes, which cannot address the problem of node overheating (where a small number of nodes have higher resource usage than the majority of other nodes). However, this article introduces the LowNodeLoad plugin of koordinator descheduler, which solves this problem. It distinguishes between high-watermark nodes, normal nodes, and low-watermark nodes based on the actual resource usage of nodes.
The koordinator descheduler is compatible with the community descheduler while adding two plugins, MigrationController and LowNodeLoad (added in Koordinator v1.1).
The LowNodeLoad plugin of koordinator is similar to the lowNodeUtilization plugin in that it evicts pods from high-watermark nodes to low-watermark nodes. However, unlike lowNodeUtilization, it classifies nodes based on their actual load, effectively addressing the issue of node resource overheating.
The MigrationController plugin provides resource reservation and arbitration mechanisms (interception mechanisms) to ensure application stability when pods are evicted by descheduler.