Optimizing Kubernetes Cluster Load: From Static Scheduling to Advanced Resource Management

Kubernetes uses a static scheduler that matches Pod request resources with Node allocatable resources. While static scheduling is simple and efficient, it often leads to a situation where the cluster appears fully allocated but the actual workload is low and node loads are imbalanced.

The article first explains why Kubernetes adopts static scheduling: dynamic scheduling that satisfies all enterprise workloads is extremely difficult, so a static approach is preferred for simplicity.

It then describes the composition of cluster resources, using CPU as an example. Each node reserves resources (system‑reserved, kube‑reserved, eviction‑hard) which are subtracted from the node’s allocatable capacity. In practice, about 5%–10% of resources become fragmented across nodes, making it hard to match user‑requested container specs with the available fragments.

To address low cluster utilization, several technical solutions are proposed:

Pod Resource Compression : Reduce the Pod request resources by a configurable ratio (stored in a workload annotation such as stke.platform/cpu-requests-ratio ) at Pod creation or recreation. The compression ratio is dynamically adjusted by a custom reconciler based on historical workload metrics.

Node Resource Oversell : Apply a configurable oversell ratio (e.g., stke.platform/cpu-oversale-ratio ) to Node allocatable and capacity resources via a Mutating Admission Webhook that patches Node status. The ratio is periodically tuned by a reconciler using node‑level load history.

Enhanced Autoscaling (HPAPlus‑Controller) : Replace the built‑in HPA controller with a per‑HPA goroutine model, allowing independent scaling policies, custom response times, workload‑aware disabling, and scaling based on pod limits and historical metrics. It also integrates with multiple monitoring back‑ends via an Extension API server.

Additional ideas such as scheduler extenders, business‑level quota management, and online/offline workload mixing are mentioned as future work.

The article concludes that these dynamic resource‑management techniques rely heavily on robust container monitoring and that the author’s team is collaborating with Tencent Cloud Monitoring to improve large‑scale cloud migration.

At the end of the article, a recruitment notice for senior cloud‑native engineers is included.