\a Concurrent Le System for a Highly Parallel Mass Storage Subsystem," in a Prototype of Swift Was Built Using Unix and an Ethernet-based Local-area Network

17 When a second Ethernet path was added between the client and the storage agents, the data-rates measured demonstrated that the Swift architecture can make immediate use of a faster interconnection medium. The data-rates for write almost doubled. For read, the improvements were only on the order of 25% because the client could not absorbe the increased network load. Second, simulations show how Swift can exploit more powerful components in the future, and where components limiting I/O performance will be. The simulations show that data-rates under Swift scale proportionally to the size of the transfer unit and the number of storage agents when suucient interconnection capacity is available. Even though Swift was designed with very large objects in mind, it can also handle small objects, such as those encountered in normal le systems. The penalties incurred are one round trip time for a short network message, and the cost of computing the parity code. Swift is also well suited as a swapping device for high performance work stations if no data redundancy is used. The distributed nature of Swift leads us to believe that it will be able to exploit all the current hardware trends well into the future: increases in processor speed and network capacity , decreases in volatile memory cost, and secondary storage becoming very inexpensive but not much faster. The Swift architecture also has the exibility to use alternative data storage technologies, such as arrays of digital audio tapes. Lastly, a system like our prototype can be installed easily using much of the existing hardware. It can be used to fully exploit the emerging high-speed networks and the large installed base of le servers. Acknowledgements. We are grateful to those that contributed to this work including Daniel Edelson for his assistance with the simulator, Aaron Emigh for his work on the prototype and in measuring its performance, Laura Haas, Richard Golding and Mary Long for their thoughtful comments on the manuscript, Dean Long for building the custom kernels necessary for our measurements, and John Wilkes for stimulating discussions on distributed le systems. 16 Swift incorporates data management techniques long present in centralized computing systems into a distributed environment. In particular, it can be viewed as a generalization to distributed systems of I/O channel architectures found in mainframe computers 21]. 6.1 Future Work There are two areas that we intend to address in the future: enhancing our current …