Update The Dissemination Of Research. Optimistic Replication Mechanism

With the rapid development of hardware and software technology, distributed system has become more and more important,especially in the fields where need high performance, high availability and strong scalability. Distributed file system, the core of distributed system, is usually designed using a client-server model. Files and directories are physically located in a server (or a group of servers). The server exports a file-system service to its clients. The clients thus can mount the exported file system on local mount points. Once this is done, applications on the clients can access the mounted remote file system and the local file system using the same application programmer's interface (API). In other words, they treat the remote file system as if it is local.In the beginning of this dissertation, we first discuss the replication mechanism in distributed systems.Distributed systems frequently use replication as a means to achieve a higher level of performance, reliability and availability. By replicating data on multiple servers, the probability that the data are inaccessible due to server or network failure is reduced. We have surveyed optimistic replication approach of replication mechanism. In particular, it focuses on the update propagation strategy invented for optimistic replication and studies how these strategies affect the functional characteristics of file system, including the system's performance and serviceability.This thesis addresses an important performance bottleneck: how to propagate the updates of file in a distributed file system. We design and implement an operation-based update propagation module in DPFS file system that developed by 8010 Research Lab.The operation-based update propagation comprises two stages. In the first stage, update-generated site needs to prepare file-attribute data and destination-site sets before shipping the update operation. In the second stage, the destination-site re-executes the update operation in local after receiving update data. This approach doesn't synchronize the file data to the destination-site after executing update operation in update-generated site, but directly ships the update operation to destination-site to be executed. The cost of re-executing the user operations to regenerate the files is often much cheaper than that of shipping the files across networks.The part of load generated by update-generated site transfers to destination sites, which smoothes the system load, avoids blocking the pending user's access requests in the update period.Finally, to evaluate the performance and availability of operation-based update propagation module in DPFS, We have done extensively in-depth simulation studies in these aspects, including cost model, latency and bandwidth of communication performance, variety of system-affecting load balance. The experiment results demonstrate: operation-based update propagation module meets the demands of DPFS.