Toward Adaptive Workload Scheduling in Kubernetes Across the Edge-Cloud Continuum

Abstract As the digital landscape evolves, the demand for high-performance, low-latency computing has surged, pushing traditional cloud computing to its limits. This has led to the conceptualization of the edge-cloud continuum, which integrates and leverages the strengths of the cloud, fog, and end-devices layers. However, realizing this continuum remains a challenge. This paper presents an architecture and evaluation of a Kubernetes-based platform that provides seamless integration between the cloud and fog layers. Key contributions include the reasoning for choosing Kubernetes as the base, a comparative analysis of single versus multi-cluster Kubernetes architectures, federation solutions, and cluster network interconnection strategies. The introduction of workload descriptors enables users to convey the edge-cloud continuum requirements and preferences of various workloads to the platform. By integrating the CosmoSpan Scheduler with the CosmoSpan Data Collector and workload descriptors, workload instance scheduling is significantly improved compared to the standard Kubernetes scheduler. This is demonstrated by a 46% improvement in the average Edge-Cloud Continuum Placement Score without sacrificing scheduling time. This score evaluates factors such as CPU, memory, node uptime, network bandwidth, latency, fault tolerance, and unscheduled workload instances. Additionally, CosmoSpan Deschedulers maintain a high and stable Edge-Cloud Continuum Placement Score despite environmental changes, ensuring that edge-cloud continuum objectives are met at all times.