Experiments with adaptive error recovery strategies

We report results on the efficiency of error recovery strategies over heterogeneous networks with both wired and wireless components. We investigate the relative impact of end-to-end delays and error patterns on protocol performance for three broad categories of protocol behavior: aggressive, conservative and adaptive. We monitor the network behavior using varying error conditions and we set end-to-end delays ranging from 10 to 200 ms in order to capture delay characteristics of wired, wireless and satellite networks. We show that (i) the end-to-end delay has varying impact on protocol performance. This impact depends on the recovery strategy, the error rate and the error type. (ii) Standard TCP Reno and Tahoe mechanisms that implement their respective recovery strategies display conflicting patterns of behavior, that might not conform to network conditions nor to their strategic goals. (iii) Adaptive behavior yields better results when a protocol flexibly alternates conservative and aggressive strategies based on the error pattern detected and the end-to-end delay. We Support this conclusion with the results of TCP-Probing.