Design and Implementation of a High-Performance Storage Server

摘要

This project investigated the design and implementation of storage servers toaccount for costs imposed by current network constraints. The activity had two main focuses. The first was on the design of a storage server that was based on the use of asynchronous protocols and techniques where ever possible to cut down on the message traffic associated with synchrony. The second was on the use of push-based data broadcast to alleviate the problems of high network asymmetry. In the first part of our study, our objective was to explore the use of asynchronous protocols in the construction of a distributed object storage manager. We have shown that we can achieve better performance than a system that uses highly synchronized techniques such as standard locking. We have also shown that such an approach can facilitate extensibility in the sense that we can decouple basic system components making it easier to add new access methods. We have implemented a version of our storage server and have tested it under various workloads to experimentally verify our theory. We have shown that we can approach the transaction throughput of commercial object oriented database systems. Since our prototype is not highly tuned, we take this to be good news. Our experiments have also led us to a new cache coherency protocol called invalidate-on-abort that performs quite well. In the second part of the effort, we have also developed a new technique that we call broadcast disks (BD) that can be used to deliver data to clients in asymmetric communication environments. We have demonstrated that under certain assumptions, this technique can significantly improve performance. We have also shown that this performance gain is most striking when a cost-based caching policy is used. We have designed such a policy as well as an implementable algorithm that approximates the ideal case.