The Evaluation And Optimization Of The Scalability For Container File System

Container is a kind of lightweight virtualization.With its characteristics of convenient deployment and speedy launch,the applications of Container on cloud platform have been increasing for recent years.Container needs special storage driver(e.g.,Aufs,Overlay)to manage its images hierarchically.However,the introduction of storage driver also increases the complexity of the I/O path,which limits the I/O scalability in a multi-container scenario.In order to research the scalability problem of container file system,a performance evaluation about Overlay's file operations is designed.According to the result,while executing file operations,Overlay introduces extra performance overhead compared to host file system.Moreover,the performance overhead increases as the number of cores increases.Base on the analysis of this result,the scalability bottlenecks of container file system are mainly from lock contention and file redirection: 1)although the mutex of host file system(VFS mutex)is able to avoid deadlock,it also severely restricts the scalability and parallelism of several operations(e.g.,rename,Copy-on-Write or deletion),because these operations need to compete for and obtain this lock before they can continue to perform.2)the file path held by user mode cannot be aware of the layers of Overlay,and while opening a file,Overlay can access its real file by utilizing this hierarchical information for redirection.Based on the two scalability bottlenecks above,two optimizations——Free-lock Rename and Memorable Redirection——are proposed,respectively.The main idea of Freelock Rename is that file operations from container will no longer take obtaining the VFS mutex as a necessary condition,and avoid deadlock scenario only by the file lock provided by Overlay and the characteristics of container(unshared workspace and one-way rename).Specifically,in the Lock-free Rename,rename from a same container will only compete for the file lock provided by Overlay,but not for VFS mutex.However,only after obtaining this file lock,the rename can continue to perform.Though this way,Lock-free Rename is able to avoid a lock contention and deadlock effectively.Memorable Redirection will utilize the real file found in the search process to avoid redundant cross-layer access.At the end of search process,Memorable Redirection Modules will save the result,and the redirection process can obtain this result directly.To evaluate the effect of two optimizations,a contrast experiment is applied to the file operations which exist the scalability problems.The result shows that compared to the native file operations,the performance of most optimized operations has been improved.Free-lock Rename improves the performance of rename,Copy-on-Write and deletion by 90%to 280%,and Memorable Redirection improves the performance of Opening by 14%.