CntrMig:A High-Performance Container Live Migration System

The containerized applications bring efficient resource utilization and rapid deployment of services to the data center.When a container application happens to fail or to be preempted,existing container management tools such as Docker will restart the container application.Restarting the container application will lose the running process of the container,increase the container application completion time and reduce the resource usage efficiency.Container live migration not only enables fast recovery of the container applications but also facilitates load balance and system maintenance over the operating system layer.However,the existing open source container live migration solution suffers from significant performance overhead.The main sources of performance problems are listed below.(1)It obtains the complete container states before the container migration,these container states are usually saved as state files;(2)the transmission of the container states will bring a lot of network traffic during migration of the container application;(3)it needs to read the state files to the memory when the container is restored.We propose CntrMig,a Docker-based high performance container live migration system in this thesis.CntrMig includes six modules.The Checkpointer module gets the container's dirty memory states and provides the pre-copy function for containers;the Proxy module receives the container states and saves it in memory;the Compression module dynamically compresses the container states based on its workload characteristics;the Bypass module transmits the compressed container states data to the migration destination through DPDK(Data Plane Development Kit)and user space network protocol stack ANS(Accelerated Network Stack);the Receiver module receives container states data and decompresses it into memory,running on the migration destination host;and the Restore module resumes the container applications from the saved state files.CntrMig eliminates accessing the file system to checkpoint/restore container applications by saving container's states using in-memory data structure,cuts down container applications downtime through the pre-copy that only collects container's dirty memory state,and reduces data transfer and latency through dynamic compression and high-performance user space network protocol stacks.We evaluate CntrMig with containerized Parsec workloads.The results show that CntrMig reduces the container applications downtime by an average of 51% and a maximum of 72%,reduces the total migration time of the container applications by an average of 14% and a maximum of 40%.Finally,compared with container restart scheme,CntrMig can achieve faster recovery of the container applications and reduce the container applications completion time by 19.7%~33.3% under different workloads.