Design And Implementation Of Index Structure Based Container Image System

Container-based virtualization has been widely used in the cloud due to its lightweight features.Docker is currently the most popular container framework because it can package applications and their dependencies into a self-contained image.Through images,users can easily store and deploy containers.Docker takes a layered structure to build its images,so that the identical layers will only be stored and pulled once when storing and deploying containers.However,layered image format introduces redundant data and useless data into different images,resulting in inefficient container storage and deployment.After investigating the top 50 images on the Docker Hub,we summarize the problems in layered image format: 1)Coarse layer-level deduplication: layer-level deduplication cannot detect and remove redundant data between different layers;2)Lack of on-demand data retrieving capability: when deploying a container,Docker needs to download the entire image in advance,but the container requires only a small part of the data in the image.Therefore,there is an urgent need for an image format that can provide fine-grained deduplication and can retrieve data on demand.Gear can simultaneously remove redundant data in the layered images and avoid downloading unnecessary data when deploying containers.It uses an index structure to replace the traditional layered structure to build images.The basic idea of the index structure is to separate the metadata and data of the container file system.Specifically,in the index structure,the file system of the container is no longer split into multiple layers,but is split into two parts: the metadata of the file system(Gear index)and the data saved in the form of independent files(Gear file).A Gear image has two parts: a Gear index and a series of Gear files.Through the Gear index,all Gear files belonging to the image can be indexed.When saving Gear images,the tiny Gear indexes are stored in a single-layer layered image,which can be conveniently stored and distributed using the Docker framework,while Gear files deduplicate at file-level through content hashing.When deploying Gear containers,containers can be started after only downloading Gear indexes.Gear file are downloaded as needed.The experimental results show that the index structure based images can reduce storage overhead by 54% compared to the layered images,increase the container deployment speed by 1.64 times under high bandwidth condition,and up to 4.82 times under low bandwidth condition.