TCP over large buffers: When adding traffic improves latency

Excessive buffer sizes in access networks may result in long delays, the effect known under the name of bufferbloat. There exist mechanisms that can solve the problem, however their deployment may be difficult or not feasible. In this paper, we explore other approaches that may bring better performance if the known solutions cannot be applied. Our main finding is the paradoxical observation that adding low intensity traffic of short packets improves latency at the price of a slight decrease in throughput. The result comes from the fact that buffer sizes are generally limited to a number of packets, so that filling them with small packets improves the queueing delay. Such a countermeasure does not require any modification at the head of the bottleneck link so an end host can use it to improve its performance. We also notice that the reverse download traffic has beneficial influence on performance, which explains why the negative effect of bufferbloat sometimes appears as alleviated. The paper presents the results of extensive experiments in a realistic setup over an ADSL link as well as on a dedicated testbed to illustrate the performance gains and the effect of several TCP variants.