Total Acknowledgements: A Robust Feedback Mechanism for End-to-End Congestion Control

End-to-end data transport protocols have two main functions: error recovery and congestion control. The information required by the sender to perform these functions is provided by acknowledgements (ACKs) from the receiver. The Internet transport protocol, TCP/IP, uses cumulative acknowledgements (CACKs), which provide a robust but minimal mechanism for error recovery which is inadequate for heterogeneous networks with random loss. Furthermore, TCP's congestion control mechanism is based on counting ACKs, and is therefore vulnerable to loss of ACKs on the reverse path, particularly when the latter may be slower than the forward path, as in asymmetric networks. The contributions of this paper are as follows: (a) We show that a simple enhancement of CACK provides suucient information for end-to-end congestion control. We term this ACK format total ACKs (TACKs). Modiication of TCP-Reno to exploit TACKs leads to signiicant performance improvements for asymmetric networks, eliminating timeouts during Reno's fast recovery phase. TACKs do not improve the error recovery capability of TCP, however, so that the modiied TCP-Reno is still vulnerable to random loss. (b) We devise a novel ACK format that uses TACKs for congestion control, and negative ACKs (NACKs) for eecient error recovery. Typically, the main concern with NACKs is that of robust-ness to ACK loss, and we address this using an implementation that provides enough redundancy to provide such robustness. (c) We use the TACK+NACK acknowledgement format as the basis for a new transport protocol that provides eecient error recovery and dynamic congestion control. The protocol provides large performance gains over TCP in an environment with random loss, and is robust against loss of ACKs in the reverse path. In particular, the protocol gives high throughput upto a designed level of random loss, independent of the bandwidth-delay product. This is in contrast to TCP, whose throughput deteriorates drastically if the random loss probability is higher than the inverse square of the bandwidth-delay product.