Worked on the kubernetes/kubernetes repository to deliver NUMA-aware device topology optimization, targeting improved performance on systems with many NUMA nodes. Developed a mechanism in Go that constrains topology hint computation to only those NUMA nodes hosting devices, reducing unnecessary enumeration and preventing kubelet stalls during startup. The approach introduced deviceNUMANodes() and updated topology hint generation to be subset-aware, while maintaining backward compatibility and safe backporting. Leveraging skills in Go, Kubernetes, and systems programming, the work enhanced scalability and efficiency for large NUMA deployments, particularly benefiting device management for GPUs and similar hardware on multi-NUMA-node hosts.