Congestion control in CSMA-based networks with inconsistent channel state

In this paper, we study the performance of utility maximization congestion control over multihop CSMA-based networks. We consider decoupled vs. joint design of congestion control and medium access and consider unmodified MAC protocols such as IEEE 802.11. Networks employing such MAC protocols incur flow starvation both without congestion control and with existing TCP-based congestion control. We develop a framework to study key issues in such networks that are not incorporated by prior models, yet are critical to the performance of congestion control algorithms. We study the role of data transmission capacity that is location dependent and, even more, unknown. We show that for the case of consistent channel state, a single globally optimal data transmission capacity does not exist. Moreover, for the case of inconsistent channel state that arises due to the carrier sense mechanism itself, a data transmission capacity that provides convergence to perfectly fair rates does not exist, i.e., the congestion control algorithm converges to incorrect rates. We study the impact of inter-node collaboration within a contention region, and show that collaboration can alleviate these problems and ensure convergence to fair rates. Finally, we compare the performance of congestion control in a collaborative network with the performance of TCP, and show that TCP starves some flows, whereas congestion control with collaboration removes starvation, provides significantly better fairness, and achieves 17% higher aggregate throughput.