Abstract
Cloud storage services, like Amazon S3, has become more and more popular for users and enterprises due to its simplified architecture design and reduced maintenance cost. However, due to architecture differences between POSIX-interfaced file systems and cloud storage systems, users can not transparently run most applications directly on a cloud storage system. Therefore, more and more researchers are interested in cloud-backed file systems aiming to deliver a system that leverage the best of both POSIX-interface and cloud storage. In this thesis, we discuss the challenges of supporting partial modification and random access of POSIX-interface operations on the cloud storage. Existing frameworks have proposed fixed-sized partitioning strategies on a file to improve the performance. However, deciding the size of these split files, so called blocks, is not trivial. In more detailed, too large blocks may contain more irrelevant data for serving an I/O request which introduces more transferring overhead. On the contrary, too small blocks may trigger more cloud requests to serve an I/O request which lowers the network utilization in the latency way. In our previous work, we have proposed a theoretical system architecture and a heuristic method to dynamically capture the I/O access pattern of users' requests and adjust file blocks on the cloud to achieve better performance. In this thesis, we show the limitation of the heuristic method, further improve the method by a mathematical model and introduced a model-based partitioning strategy. We show that our strategies can analysis I/O patterns and adjust the blocks on the cloud to achieve better performance. Our evaluation shows that our strategies can improve the overall performance by 7\% to 106\% compared to the static fixed-sized partitioning strategy.