Collaborative congestion control in parallel TCP flows

We suggest a new congestion control scheme in parallel TCP flows and compare it with single-flow based congestion control approaches in the Internet. When a node opens multiple connections that are going to the same or nearby destinations, we can assume that there is a significant correlation of congestion events among the flows. In this paper, we leverage this relationship to improve the performance provided to parallel TCP flows by having TCP senders of the related connections exchange congestion information. The usefulness of our scheme is shown by augmenting a measurement-based TCP congestion control mechanism to utilize congestion measurement information from other TCP flows at the sender. Furthermore, we compare our scheme with ECN-based TCP congestion control. Our measurement show improvements over single-flow based approaches. In terms of packet loss and delay jitter, we show that collaborative congestion control is comparable with architecturally significantly more intensive schemes, such as ECN-based TCP.