Scalable Parallel-Access for Mirrored Servers
A. Zeitoun · 2002
To increase a client's perceived throughput while transferring large multimedia files, a parallel-access approach can be used to simultaneously transfer different pieces of a file from different mirrors. Previously proposed approaches, however, either employ very aggressive implementations with poor scalability or suffer from deployment difficulties on the current Internet. Using trace-driven and real Internet experiments, we show the effect of poor scalability and performance degradation for some of the previous parallel-access implementations. In this paper, we propose a hybrid scheme that (1) minimizes the added load on all mirrors while (2) maximizing the client-perceived throughput. Our scheme uses informed mirror selection based on round-trip time (RTT) measurements as well as dynamic monitoring. While using less than 10% of the available mirrors, we show a decrease in transfer time by more than 85% over traditional single-server transfers and 35% over other parallel-access implementations.