Image Preheating: Enables Fast Startup of Large-Scale Parallel Containerized Jobs

Containers and Kubernetes are becoming the normative way to deploy services in the data center. This rich virtualization ecosystem simplifies application development and management, resolves complex dependencies between applications and execution environments, and improves the efficiency of application deployment. Unfortunately, the startup and deployment delays of containerized jobs significantly limit the efficiency of job execution. In particular, the larger the container image, the longer the image download time, which seriously affects the service processing capability of the system. In this paper, we propose the image preheating technique to accelerate the startup of large-scale distributed parallel containerized jobs. According to the different working scenarios, image preheating is divided into layer preheating and image preheating. First, layer preheating is applied in the scenario where a single machine pulls a single image to start a single container, and we explore the parallel nature of multiple threads in pulling different layers of images and propose a more efficient heuristic algorithm to accelerate parallel pulling of images. Second, image preheating is applied in the scenario where the entire parallel job is deployed and started on the cluster. As soon as one task image of the job is downloaded, the preheating is completed and the entire job starts. In addition, we propose a very practical approach called manual image preheating to provide a better experience for users executing the job. Simulated and real experiments show that the method we designed can improve the download speed of images by 23.54% and the start of jobs in the cluster by 3.73%.