Skip to Main Content
This paper investigates the performance of a single TCP-Reno connection over a lossy, congested link as a function of TCP packet size on the forward link and the ratio α of forward-to-reverse link capacity measured in terms of packets/sec. α>1 constitutes the primary scenario of interest since the reverse link (for the ACK) is a bottleneck, which directly impacts the forward link throughput. Congestion at the forward buffer is induced by assuming buffer size to be significantly less than the bandwidth-delay product of the round-trip connection. Further, forward link error characteristics are modeled as a continuous-time two-state Markov chain, where packets are either transmitted free of errors (good state) or lost (bad state). By tuning the model parameters, the link error process can be made to vary from independent loss to (highly) correlated loss scenarios. Results show that TCP-Reno performance benefits from larger packet sizes for connections dominated by link errors; throughput is less sensitive to packet size when the connection is dominated by forward path congestion. Additionally, performance sensitivity to variation of forward buffer size and the round-trip propagation delay is discussed and supported by simulation results using ns-2 network simulator.