SCut: Cost-Aware Kubernetes Pod Scheduling via Spectral Graph Cuts

Provider charges can be incurred for pod traffic across cloud Availability Zones. Resource profiles are used by the default kube-scheduler, while conventional spectral cuts are based on communication alone. In SCut (Spectral Cut), a spectral partition is obtained from pod traffic and scalarized CPU and memory requests. For connected graphs, generalized Fiedler coordinates are obtained by scaling the second eigenvector of the symmetric reduction by inverse square roots of the corresponding resource weights, and unconstrained graph conductance is bounded. For disconnected graphs, whole components are grouped deterministically by resource volume, and CPU and RAM feasibility is checked separately at each worker node. Policy stages were represented in a Go prototype through local interfaces modeled on the Kubernetes Scheduling Framework. An eBPF telemetry pipeline is proposed for integration, while implementation and evaluation of live collection remain future work. In a discrete-event simulation, SCut was compared with Kubernetes LeastAllocated, an incremental ILP per pod, rules inspired by NetMARKS and TraDE, and the published and zone-aware SCORE placement steps. A disconnected DeathStarBench Social Network profile was used for component grouping, and the generalized Fiedler sweep was evaluated on two connected synthetic profiles based on Alibaba trace characteristics. For the Social Network workload, both components were kept within their assigned zones, resulting in zero modeled egress cost. The lowest mean modeled costs on both connected profiles were obtained with SCut. For the largest profile, costs were based on realized placements because some policies left pods unscheduled. Plugin deployment and cloud environments were outside the evaluation.