Experimental Performance Evaluation of ATP in a Wireless Mesh Network

It is well known that TCP performs poorly in wireless mesh networks. There has been intensive research in this area, but most work uses simulation as the only evaluation method; however, it is not clear whether the performance gains seen with simulation will translate into benefits on real networks. To explore this issue, we have implemented ATP, a transport protocol designed specifically for wireless ad hoc networks. We choose ATP because it uses a radically different design from TCP and because reported results claim significant improvement over TCP. We show how ATP must be modified in order to be implemented in existing open-source wireless drivers, and perform a comprehensive performance evaluation on a mesh testbed under different operating conditions. Our results show that the performance of ATP is highly sensitive to protocol parameters, especially the epoch timeout value. To improve its performance we design an adaptive version that utilizes a self-adjustable feedback mechanism instead of a fixed parameter. We illustrate its advantages by using a measurement study to compare it with the original ATP and to a standard TCP Tahoe implementation.