This document describes performance measurements taken in the MAGIC
gigabit testbed relating to the performance of TCP in wide area
ATM networks. In particular, the behavior of TCP with and without
ATM cell level pacing is studied. DEC OTTO interfaces, which support
ATM/SONET at 155 Mb/s link rates were used as the basis for these
studies. These interfaces support cell level flow control as well
as schedule-based cell pacing.
Some of the conclusions that might be drawn are: (1) traffic shaping (for example, the OTTO scheduling) is critical when bottlenecks such as OC-3c to TAXI rate mismatches occur in the network, even when only single hosts are involved, (2) traffic shaping substantially improves performance when traffic from multiple hosts is multiplexed across a single link; significant packet losses were observed even with relatively slow sources with no pacing, while cell pacing lead to no observed losses and higher throughput, (3) cell level pacing is necessary because the TCP rate control mechanism does not control traffic burstiness sufficiently to avoid congestion-induced cell losses in wide area networks, primarily due to large TCP segment sizes and small switch buffers.
Send comments or questions to Joseph B. Evans, evans@ittc.ku.edu.