Vidhyashankar Venkataraman (Cornell University, US); Paul Francis (Cornell University, US); M. Kodialam (Bell Labs, US); T. V. Lakshman (Bell Labs, Lucent Technologies, US)
High-speed organizational networks running over leased fiber-optic lines or VPNs suffer from the well-known limitations of TCP over long-fat pipes. High-performance protocols like XCP require changes in the network. Other protocols like FastTCP assume nothing about the network but may not perform as well as network-aware protocols. In this paper, we present a new transport protocol that exploits the fact that these networks can offer priority queuing, thus finding the sweet spot between assuming too much and too little about the network. Our protocol splits a given transport flow into two prioritized flows. The higher priority flow operates with the legacy congestion control while the lower priority flow aggressively exploits spare capacity in the network while not interfering with the other participating flows. This isolation of the aggressive flow into strictly lower priority queues gives us more latitude in how to operate the aggressive component. We provide a detailed description of our protocol based on our user-space implementation. We show through Emulab experiments as well as simulations that this protocol can produce near-perfect goodputs in lossy networks, can considerably improve the completion time of short flows, and can sustain a high bottleneck utilization even in changing network conditions.