We have presented TCP-Santa Cruz, which implements a new approach to end-to-end congestion control and reliability, and that can be implemented as a TCP option. TCP-Santa Cruz makes use of a simple timestamp returned from the receiver to estimate the level of queueing in the bottleneck link of a connection. The protocol successfully isolates the forward throughput of the connection from events on the reverse link by considering the changes in delay along the forward link only. We successfully decouple the growth of the congestion window from the number of returned ACKs (the approach taken by TCP), which makes the protocol resilient to ACK loss. The protocol provides quick and efficient error-recovery by identifying losses via an ACK window without waiting for three duplicate acknowledgments. An RTT estimate for every packet transmitted (including retransmissions) allows the protocol to recover from lost retransmissions without using timer-backoff strategies.
Simulation results show that TCP-Santa Cruz provides high throughput and low end-to-end delay and delay variance over networks with a simple bottleneck link, networks with congestion in the reverse path of the connection, and networks which exhibit path asymmetry. We have shown that TCP-Santa Cruz eliminates the oscillations in the congestion window, but still maintains high link utilization. As a result, it provides much lower delays than current TCP implementations. For the simple bottleneck configuration our protocol provides a 20% - 39% improvement in end-to-end delay (depending on the value of n) and a delay variance 3 orders of magnitude lower than Reno. For experiments with congestion on the reverse path, TCP-Santa Cruz provides an improvement in throughput of at least 47% - 67% over both Reno and Vegas, as well as an improvement in end-to-end delay of 45% - 59% over Reno with a reduction in delay variance of 3 orders of magnitude. When we examine networks with path asymmetry, Reno and Vegas achieve link utilization of only 52% and 33%, respectively, whereas Santa Cruz achieves 99% utilization. End-to-end delays for this configuration are also reduced by 42% - 58% over Reno.
Our simulation experiments indicate that our end-to-end approach to congestion control and error recovery is very promising, and our current work focuses on evaluating the fairness of TCP-Santa Cruz, its coexistence with other TCP implementations, and its performance over wireless networks.