Integrating Apple Silicon and macOS as Kubernetes Worker Nodes : Evaluating User-Space and Virtualization for Workload Execution

Sammanfattning

Kubernetes serves as the de facto standard for orchestrating containerized workloads, exposing limitations regarding diverse hardware integration. Specialised architectures such as Apple Silicon remain isolated from these mainstream cluster environments. This research addresses a key operational gap by demonstrating and evaluating macOS worker-node integration in a Kubernetes-based research cluster. Achieving hardware acceleration within isolated environments, such as virtual machines, on Apple Silicon required implementing and evaluating two distinct execution models. The first model leverages user-space for direct execution, granting unrestricted access to the underlying hardware. The second utilises Apple's virtualization framework, specifically building upon the macos-vz-kubelet project to provision macOS virtual machines. macOS lacks native compatibility with Linux-based distributed storage solutions, specifically Ceph Persistent Volume Claims. Resolving this limitation required deploying an intermediary NFS server hosted within a standard Linux pod. Evaluating these implementations revealed clear operational trade-offs. The user-space approach delivers superior computational throughput, demonstrating performance gains of approximately 20% for CPU and 5% for GPU workloads compared to the virtualization implementation. Performance metrics fluctuate significantly depending on the specific framework and workload evaluated. The virtualization implementation successfully mimics standard Kubernetes workflows, offering significantly enhanced usability and lower operational complexity than the user-space model. A single-node, quantized LLM inference benchmark evaluated the energy efficiency of the user-space implementation against an NVIDIA H100 MIG 40GB configuration. The macOS implementation achieved roughly 8 to 19 times higher tokens-per-watt than the H100 MIG configuration. The results demonstrate technical feasibility in the evaluated environment, while automated deployment, scaling, and security remain open challenges.