The effectiveness of end-to-end congestion control mechanisms
JÃ ⁄ rg Bolliger, Urs Hengartner, Thomas K.R. Gross · Repository for Publications and Research Data (ETH Zurich) · 1999
TCP’s success is in part due to its ability to deal with congestion, yet congestion control remains an important topic for today’s Internet protocols. Over time, several enhancements have been proposed that improve TCP Reno’s congestion control mechanism, e.g., FACK TCP, Rate-Halving, and TCP Vegas. Most of these enhancements are motivated and evaluated by simulations or small-scale experiments. To assess the effectiveness of these improvements in practice, we performed an Internet experiment during six months using a set of hosts in North America and Europe. We measured protocol dynamics of about 25,000 bulk-data transfers using various congestion control mechanisms. Although this study is limited in that it provides only a snapshot of a small part of the rapidly chang ing Internet, it allows us to draw the following conclusions when comparing the proposed TCP enhance ments to the baseline Reno protocol: Overall, SACK-enhanced protocols are more robust against packet loss (for loss rates smaller than 10 %). As a consequence, they achieve significantly higher bandwidths. Furthermore, these protocols utilize the available network resources more efficiently as they cause fewer unnecessary retransmissions. The global nature of our experiment allows us to see that the benefit of SACK-enhanced protocols varies widely, e.g., it is almost a factor two bigger for intra-continental than for inter-continental connections. Vegas-style congestion window management during the congestion avoidance phase moderately improves on the number of (multiple) packet loss events, but it does so at the expense of lower throughput. These findings are supported by protocol micro-measurements: SACK-enhanced protocols (i) are highly effective in avoiding timeouts due to burst losses and due to non-trigger of recovery and (ii) are able to detect and repair lost (fast) retransmissions. Furthermore, we show that Rate Halving’s congestion control strategy is a win in situations with small congestion windows, because it is able to evoke additional duplicate acknowledgments and thus reduces the number of non-trigger timeouts. Nonetheless, non-trigger timeouts still account for a significant fraction of the timeouts experienced by Rate Halving. Although Rate-Halving’s retransmission strategy is more conservative than FACK’s, overall, Rate-Halving coupled with lost retransmission detection experiences fewer timeouts.