TotalCOW: Unleash the Power of Copy-On-Write for Thin-provisioned Containers

Modern file systems leverage the Copy-on-Write (COW) technique to efficiently create snapshots. COW can significantly reduce demand on disk space and I/O bandwidth by not duplicating entire files at the time of making the snapshots. However, memory space and I/O requests demanded by applications cannot benefit from this technique. In existing systems, a disk block shared by multiple files due to COW would be read from the disk multiple times. Each block in the reads is treated as an independent one in different files and is cached as a sperate block in memory. This issue is due to the fact that current file access and caching are based on logic file addresses. It poses a significant challenge on the emerging light-weight container virtualization techniques, such as Linux Container and Docker, which rely on COW to quickly spawn a large number of thin-provisioned container instances. We propose a lightweight approach to address this issue by leveraging knowledge about files produced by COW. Experimental results show that a prototyped system using the approach, named TotalCOW, can significantly remove redundant disk reads and caching without compromising efficiency of accessing COW files.