Efficient Multi-Cluster Scheduling for Heterogeneous Workloads

As heterogeneous multi-cluster environments become increasingly prevalent in large-scale AI and data-intensive workloads, traditional scheduling strategies struggle to balance efficiency, scalability, and adaptability. This paper proposes a unified cross-cluster scheduling framework that addresses these challenges through three key innovations: an enhanced multi-cluster resource scheduling architecture, a standardized resource capability assessment model, and a multi-factor scheduling algorithm. The architecture introduces a global coordination layer above decentralized clusters, enabling centralized scheduling decisions while preserving local autonomy. To account for hardware diversity, we define a normalized performance model based on resource capability factors across CPU, GPU, and memory. Building upon this, a comprehensive scheduling algorithm evaluates task-resource matching through affinity similarity scoring and latency-aware penalties that reflect real-time bandwidth and data locality. The proposed strategy is implemented on a multi-cluster orchestration platform and evaluated using a diverse set of workloads. Compared to baseline approaches such as Round-Robin and Resource-Aware Strategy, our method significantly improves resource utilization, reduces scheduling delays, lowers failure rates, and achieves superior load balancing across clusters. These results demonstrate the robustness and effectiveness of our solution in complex, heterogeneous computing environments.