Idle Compute Sharing in Dedicated Kubernetes Clusters Using Karmada

Overview: With the launch of dedicated Kubernetes clusters, many business lines have underutilized resources. To improve utilization, the platform introduced an idle compute sharing feature that allows sharing spare CPU/memory from dedicated clusters to offline workloads.

Use Cases: (1) Fully managed dedicated clusters with low CPU usage can share idle compute to generate cost offsets. (2) Adding existing nodes to improve utilization of non‑containerized machines, e.g., DB services with low CPU usage, by joining them to dedicated K8s clusters and configuring shareable CPU/memory.

Feature Details: Users configure mixed‑workload policies, set shareable resource amounts, and define share plans (scope, time slots). The UI path is Cluster Management → Cluster Details → Idle Compute Sharing Plan. Plans can be created, edited, disabled, or deleted.

Technical Implementation: The solution builds on Karmada with custom extensions. It uses cgroup cpuset to bind CPUs, defines compression and eviction watermarks, and introduces a resource interpreter to treat offline pods as batch‑cpu/batch‑memory for scheduling without modifying original deployments. A push member‑cluster integration is used for the sharing scenario.

Current Status and Future Plans: The feature is in production, improving CPU utilization from ~15% to ~35% for DB services. Ongoing work includes expanding to more services, refining public‑cluster fallback logic, and further integration with search cache and resource registry.