The Available Bit Rate (ABR) service has been developed to support data applications over Asynchronous Transfer Mode (ATM). It is hence interesting to study the performance of reliable data transport protocols like TCP over ABR. We study the effect of running large unidirectional file transfer applications on TCP over ABR with an explicit rate algorithm (ERICA) implemented at the ATM switches. The study shows that it is possible to get maximum TCP throughput when there are enough buffers at the switches and the sources are rate-limited by ABR. However, when the number of buffers is smaller, there is a large reduction in throughput even though the Cell Loss Ratio (CLR) is very small and the sources are window-limited by TCP. The primary reason for the large throughput degradation is the TCP timer granularity (which affects the timeout interval) and retransmission of duplicate packets which are discarded at the destination. We show that switch buffers should not be dimensioned based on the ABR Source parameter Transient Buffer Exposure (TBE). When ABR capacity is varied, we find that CLR exhibits high variance and is not a good indicator of TCP performance. Further, larger buffers and a larger number of TCP sources increase TCP throughput. Even when buffers are small, dropping of End of Message (EOM) cells should be avoided. This avoids merging of packets at the destination AAL5 and improves fairness.
Complete Paper in Adobe Acrobat (739,676 Bytes)