Kubernetes-aware load balancing for leaf-spine EVPN data centers

Data centers and Kubernetes are fundamental to modern cloud computing, both relying on effective load balancing for optimal performance. Kubernetes services distribute workloads across nodes and pods, while data center networks using leaf-spine topologies distribute traffic across multiple paths. However, a significant issue arises: Kubernetes requests can be redirected between nodes, suffering extra routing hops and adding unnecessary load and latency. This occurs because Kubernetes and network load balancing operate independently of each other. This paper proposes a solution to this problem: making the data center network Kubernetes-aware. By integrating real-time Kubernetes API data into network load balancing via distributed agents at leaf routers, we dynamically adjust ECMP weights using EVPN and BGP extended communities, aligning network routing with pod distribution. We demonstrate our solution experimentally using SR Linux, leveraging its non-native agent capabilities. Results validate that our Kubernetes-aware routing adapts to cluster workloads, eliminating inter-node redirection and optimizing performance.