In recent years, container technology gained more and more importance, especially for the deployment of cloud services. While many studies focus on the performance characteristics of running containers, only a few explicitly address container start times. However, a good understanding of those can help performance engineers and cloud service providers in various domains. For example, autoscaling and scheduling algorithms can profit from additional knowledge about container start times, if they explicitly consider start delays for diferent containers. Another related research field is serverless computing, where the optimization of container startups and the analysis and mitigation of cold starts are vital research topics. Therefore, container start times and their influencing factors need to be analyzed in a systematic way. State-of-the-art literature focuses mainly on comparative analyses of diferent container runtimes and their influence on the container start time [ 1, 2]. Only few consider additional influencing factors like storage driver [ 3], base images [3, 4] as well as particular deployment aspects [5]. Moreover, the majority of these studies have only short evaluations with few container images and measurement repetitions as well as non-standardized methodologies. In this work, we perform a systematic and extensive study to identify influencing and negligible factors for container start times. The start process of a container begins when the container engine receives the task to start a new container. It ends with the start of the execution of the container's start command defined in the image manifest. The container start time is the time elapsed between these two events. During the design and execution of this study, we explicitly work with the recent ACM SIGSOFT Empirical Standards [6]. We group potential influencing factors for container start times in two categories: image factors and deployment factors. The image factors are all information that is extractable from the container's image manifest. This includes the image size, number and size of layers, ports, volumes, and more. Deployment factors include the node OS, kernel and architecture, container runtime, storage driver as well as other metrics, like the number of running containers on the node. The pull or download time of the container image will not be investigated further, as this is mostly a matter of network conditions. Our goal is to predict container start times based on image and deployment information. Our study is divided into three stages. In stage one, we will perform extensive measurements of container start times with a large number of images from the prominent public repository
Docker Hub. A study analyzing the characteristics of container images on Docker Hub was
published in 2019 [
In this talk, we present the study design, experiment setup, and preliminary results of our study. Moreover, we elaborate more about the motivation and potential use cases. We hope that the discussion will help us to discuss the strengths and weaknesses of this study regarding both the methodological and technical aspects.