On the performance of TCP over throughput-optimal CSMA

An interesting distributed throughput-optimal CSMA MAC protocol, called adaptive CSMA, was proposed recently to schedule any strictly feasible rates inside the capacity region. Of particular interest is the fact that the adaptive CSMA can achieve a system utility arbitrarily close to that is achievable under a central scheduler. However, a specially designed transport-layer rate controller is needed for this result. An outstanding question is whether TCP Reno (one of the most mature versions of TCP) is compatible with adaptive CSMA and can achieve the same result. The answer to this question will determine how close to practical deployment adaptive CSMA is. Our answer is yes and no. First, we observe that running TCP Reno directly over adaptive CSMA results in severe starvation problems. Effectively, its performance is no better than that of TCP Reno over legacy CSMA (IEEE 802.11), and the potentials of adaptive CSMA cannot be realized. We then propose a multi-connection TCP solution with active queue management and prove that it can work with adaptive CSMA to achieve optimal utility. NS-2 simulations demonstrate that our solution can alleviate starvation and achieve fair and efficient rate allocation. We remark that multi-connection TCP can be implemented at either application or transport layer. Application-layer implementation requires no kernel modification, making the solution readily deployable in networks running adaptive CSMA. Our results show that adaptive CSMA can work well with only light-weight TCP modifications, bringing it a step closer to practicality.