A Resource Isolation Approach In File System For Docker Container

With the rapid development and popularization of cloud computing,applications in cloud are becoming more and more complex and diverse.How to reduce resource competition among applications in virtualization environment has become a research hotspot.Docker promotes the development of container.Comparing with virtual machine,container,an operating system-level virtualization technology,has lower performance overhead.However,sharing the kernel leads to a worse isolation.Specifically,in traditional centralized file systems,there is a lot of competition of resources,such as the locks to protect global data structures,which will degrade the isolation of file system operations in containers.Besides,containers share limited soft resources(such as inode,block and fd)in host file system without a mechanism to isolate them.FXMARK is an open source benchmarks suite that measures the isolation of file system operations.However,its existing benchmarks are mainly for the host file system.Extend FXMARK to support testing in Docker container,after analysis with a series of experiments,it is found that the lock competition in the file system operation of the container comes from the host file system,but the Docker storage driver also introduces new lock competition.Then,the lock competition introduced by storage driver is analyzed in several concurrent scenarios.The result shows that a new lock function can be implemented for containers to avoid the competition and improve the performance of file system operation in the container.Since the Docker file system only forwards file system operations to the host file system,each container can be assigned a weight to use the soft resources.The file system operation of the container is forwarded or intercepted according to the usage weight of the soft resource and the current state information of the soft resource to realize the allocation of the soft resource between containers.Experiments show that the optimized file system operations can achieve at least 3 times performance improvement in 10-core.Besides,the optimized system can achieve a proportional allocation of soft resources between containers with minimal overhead.