A Container Acceleration System Based On Non-volatile Memory

In recent years,cloud computing technology has developed rapidly,in which container plays an important role.However,in scenarios such as serverless computing,slow deployment and cold-start of containers are still a performance bottleneck.Whenever a function is invoked,the container is started and is destroyed immediately after handling the request.In serverless computing platforms,languages such as Python and JavaScript,require heavy runtimes,making cold-start over 10× slower than launching an equivalent C program.These technologies make container cold-start more frequent and expensive.The emerging non-volatile memory provides a new opportunity to improve container performance due to its excellent read-write characteristics.In order to speed up the deployment and cold-start of containers,134 images downloaded more than 1 million times on Docker Hub are analyzed in depth and two findings are obtained: 1.Image storage is one of the reasons for the slow deployment of containers;2.Only a part of the image file is required for container cold-start.The core design idea of N-Docker,a container acceleration system based on non-volatile memory,is based on these two findings.During the container deployment phase,N-Docker can detect the storing layers that are too slow and store them in non-volatile memory.In the cold-start phase of container,N-Docker captures the hot image files necessary for container cold-start by tracking the process of container cold-start,stores them in non-volatile memory,and writes the cold image files back to the hard disk.The experimental results show that compared with the original Docker,the deployment of the N-Docker container is 1.21× faster and the cold-start of the container is 2.96× faster.Compared with NVM-Docker,which stores all images in non-volatile memory,NDocker achieves the same performance,but the usage of non-volatile memory is only 11.8%.