Hound: A Parallel Image Distribution System for Cluster Based on Docker

Current applications, consisting of multiple replicas, are packaged into lightweight containers with their execution dependencies. Considering the dominant impact distribution efficiency gigantic images on container startup (e.g., distributed deep learning application), image “warm-up” technique which prefetches these replicas to destination nodes in cluster is proposed. However, current solely focuses identical distribution, fails take effect when distributing different nodes. To address this problem, paper proposes Hound, a simple but efficient system based Docker. support diverse requests nodes, Hound additionally adopts node-level parallelism (i.e., downloading parallel) further improve distribution. The experimental results demonstrate outperforms Docker, kubernetes runtime interface (CRI-O), and Docker-compose terms performance request images. Moreover, high scalability evaluated scenario ten