Report on "Scalable TCP: Improving Performance in High Speed Wide Area Networks"

In this paper Tom Kelly proposes another flavor of TCP congestion control algorithm, which arguably is incrementally deployable and results in better utilization of link capacity, when flows are few and heavy duty. The motivation behind the paper is the fact that traditional TCP reacts slowly in lossy networks, where packet drops are frequent even before the full link capacity is reached. High loss rates result in many premature drops in the window size, which results in poor utilization . Authors propose that instead of using the traditional "additive increase and multiplicative decrease", "multiplicative increase and multiplicative decrease" enables congestion window to react quickly in face of random losses. Recommendations for choosing "good" multiplicative constants for increase and decrease of the window sizes are also made and deeply analyzed.

In the group discussion, this paper has seen a group of strong supporters, while another group was skeptical about the benefits. The advocates argued that the proposed modification is reasonably implementable and also results in large performance gains in real networks. Argument that performance is improved only when there are few heavy duty flows sharing the entire link is also acceptable, because there is a large community which needs to regularly transfer bulk data across the dedicated links. Furthermore it is argued and supported by experiments that scalable TCP interoperate well with the legacy traffic. Several other benefits like small instantaneous rate variation, etc has been pointed out.

Another group who was skeptical about the real benefits argued that paper doesn't make a strong case by itself. It was argued that the experiments conducted were not sufficient . Experiments (for fairness) where multiple scalable TCP flows compete for the link capacity should have been conducted. Also the results of performance improvements, when the scalable TCP operates in networks with large buffers, should be presented. This is necessary because traditional TCP tends to utilize link capacity efficiently as the amount of network buffer is increased. Furthermore performance of scalable TCP in complex network topology with heterogeneous RTTs needs to be quantified.

In conclusion, scalable TCP proposes a small and intelligent modification to the traditional TCP. However, in order to make a strong case, it needs further analytical and experimental results.

Sailesh Kumar