Congestion Control by Bandwidth-Delay Tradeoff in Very High-Speed Networks: The Case of Window-Based Control
Increasing bandwidth-delay product of high-speed wide-area networks is well-known to make conventional dynamic traffic control schemes "sluggish". Still, most existing schemes employ dynamic control, among which TCP and ATM Forum's rate-based flow control are prominent examples. So far, little has been investigated as to how the existing schemes will scale as bandwidth further increases up to gigabit speed and beyond. Our investigation in this paper is the first to show that dynamic control has a severe scalability problem with bandwidth increase, and to propose an entirely new approach to traffic control that overcomes the scalability problem. The essence of our approach is in exercising control in bandwidth domain rather than time domain, in order to avoid time delay in control. This requires more bandwidth than the timed counterpart, but achieves a much faster control. Furthermore, the bandwidth requirement is not excessively large because the bandwidth for smaller control delay and we call our approach Bandwidth-Latency Tradeoff (BLT). While the control in existing schemes are bound to delay, BLT is bound to bandwidth. As a fallout, BLT scales tied to bandwidth increase, rather than increasingly deteriorate as conventional schemes. Surprisingly, our approach begins to pay off much earlier than expected, even from a point where bandwidth-delay product is not so large. For instance, in a roughly AURORA-sized network, BLT far outperforms TCP on a shared 150Mbps link, where the bandwidth-delay product is around 60KB. In the other extreme where bandwidth-delay product is large, BLT outperforms TCP by as much as twenty times in terms of network power in a gigabit nationwide network. More importantly, BLT is designed to continue to scale with bandwidth increase and the performance gap is expected to widen further.