Skip to main content
Kubernetes container optimization policies in Kubex define how workload data is analyzed and how resource requests and limits are recommended. The policy dialog groups its controls into operational windowing, resource settings, GPU settings, and advanced settings.

Figure: Container Optimization Policy Dialog

The values below are based on the default policy and they may differ from the values configured in your environment.

​
CPU Settings

CPU Settings contains separate Request and Limit policies. Each policy can be based on the Average Container or Busiest Container, and can use the Sustained, Average, or Peak statistic. Maintain Usage Between sets a target utilization range. The CPU allocation range defines the minimum and maximum recommended CPU values and the step between values.

​
Memory Settings

Memory Settings contains separate Request and Limit policies. Select the memory measurement (Total, RSS, or Working Set), the container basis, the statistic, and the target usage range.

​
Ephemeral Storage Settings

Ephemeral Storage Settings contains separate Request and Limit policies for ephemeral storage.

​
GPU Settings

GPU Settings provides policies for GPU and GPU Memory. Each has separate Average Container and Busiest Container settings, with Sustained, Average, or Peak available as the statistic. The GPU Allocation Range applies to GPU sizing. The screenshot shows an increment of 1 GPU and a range of 0.125, 0.25, 0.5, 0.75, and 1–100 GPUs. No separate allocation range is shown for GPU Memory.

​
Advanced Settings

​
Operational Windowing

Operational Windowing specifies the analysis period and the samples used for analysis.

​
ML Model Strategy

  • Combined looks at the historical workload range and generates a 24-hr representative model
  • Outer envelope will react quickly to any increase in usage

​
Guaranteed QoS

  • Off: sizing recommendations will not consider Guaranteed QoS
  • Preserve: will detect Guaranteed QoS if limit = request, and will preserve it in the recommendations
  • Force: resizing recommendations assuming Guaranteed QoS is required

​
HPA Settings

Choose an HPA Recommendation Mode:
  • Ignore HPA Metrics: Request recommendations are not generated for resources used by HPA scaling metrics.
  • Optimize with Headroom: Continue generating recommendations for HPA-controlled resources.
When configuring headroom, the dialog provides HPA Target Headroom (%) values for CPU and Memory. Default value is 10% for each.

​
Policy Constraints and Allocation

The resource policy sections include these additional controls where shown:
  • Min Hours for Reclaim sets a minimum number of hours before reclaiming resources; the defaults are 2 hours for requests, 100 hours for CPU, Memory, and Ephemeral Storage limits, and 168 hours for GPU and GPU Memory.
  • Max Increase and Max Decrease can be left blank for no limit.
  • Allow Downsize controls whether the policy can recommend a smaller allocation.
  • Max Limit-to-Request Ratio is available for CPU and Memory requests; leave it blank for no ratio limit.
  • Recommend Limit when Unspecified is shown for CPU and Memory limits.
  • Min, Max, and Increment define allowed allocation sizes for CPU, Memory, and Ephemeral Storage.
Review policy changes against workload requirements and cluster capacity before applying them.