Impact of ATM ABR control on the performance of TCP-Tahoe and TCP-Reno

We study the effect of ATM ABR control on the throughput and fairness of running large unidirectional file transfer applications on TCP-Tahoe and TCP-Reno with a single bottleneck ATM link. The various ATM ABR control are characterized by different switch queue threshold (Qth) values which define when the bottleneck link declares congestion and starts marking RM cells and, the reduction factor (RDF) values which define how quickly the source reduces its cell transmission rate. The key results of this study are the following. First, our results show that TCP-Reno is slightly better than TCP-Tahoe in terms of the throughput when the round-trip delay is short, specially with a large packet size. Second, with a medium round-trip delay (1 ms), TCP-Reno gives a better throughput with long packets, while TCP-Tahoe performs better with short packets. Third, when the round-trip delay is long (30 ms), TCP-Tahoe clearly offers higher throughput than TCP-Reno. TCP-Tahoe consistently outperforms TCP-Reno in terms of fairness. Fourth, for a short or medium round-trip delay, tight ABR control (specially with lower RDF) results in higher throughput than loose control; while the opposite is true for a long round-trip delay. Finally, a strong positive correlation between poor throughput and poor fairness is only valid for short and medium round-trip delays for TCP-Tahoe