Performance Benchmarking of Feather Virtual Nodes for Edge Computing
摘要
Using lightweight orchestration platforms that have low startup latency and resource overheads on constrained nodes is critical in edge computing. Feather, a lightweight orchestration framework for edge environments, has been designed with this challenge in mind. Feather has some attractive design characteristics but still has not been rigorously evaluated under realistic edge workloads. This paper introduces the first overall benchmarking suite for Feather across multiple runtimes. The evaluation includes: (i) cold start latency for container images of various size and complexity, (ii) image-pull and caching, and (iii) CPU and memory usage for stress workloads. The benchmarks include both containerd and the OSv unikernel backends. In general, for medium-to-large container images, the study found that the OSv backend can lower startup latency while Feather’s unique virtual-node design can achieve consistent resource isolation in the face of load. These results provide new knowledge and understanding about Feather’s scalability, runtime concerns, and its role in multi-tenant edge computing environments.