1 Load Characterization and the Implementation of Service Levels in the Scalable Proxy Architecture

The TranSend transformation proxy is an example of a new class of middleware services being deployed in the internet. TranSend performs lossy compression on Web data before delivering it to the end user, resulting in a 3-7 increase in surfing speed in return for degraded image quality. In this paper, we perform detailed measurements of the traffic and load that TranSend will experience from the Berkeley home-IP modem pool population; this load analysis is performed using the results of a 2 month long Web trace gathering experiment. We then use the results of this traffic analysis to design queuing disciplines that will allow us to provide differentiated levels of service to different TranSend traffic classes. We compare FIFO queuing, Strict Priority queuing, and Lottery Scheduled queues, and we show that lottery scheduled queues in the “distillers” (the transformational workers in TranSend) result in the best overall service.