TCP congestion control macroscopic behaviour for combinations of source and router algorithms

Abstract

The network side of Transmission Control Protocol (TCP) congestion control is normally considered a black-box in performance analysis. However, the overall performance of TCP/IP networks is affected by selection of congestion control mechanisms implemented at the source nodes as well as those implemented at the routers. The paper presents an evaluation of macroscopic behaviour of TCP for various combinations of source algorithms and router algorithms using a Dumbbell topology. In particular, we are interested in the throughput and fairness index. TCP New Reno and TCP Cubic were selected for source nodes. Packet First-in-First-out (PFIFO) and Controlled Delay (CoDel) mechanisms were selected for routers. The results show that TCP New Reno performs well, in terms of throughput, in a low BDP scenario. However, as expected in high BDP scenario, TCP New Reno deteriorates and TCP Cubic performs better. CoDel in the network side further deteriorates TCP New Reno flows in high Bandwidth-Delay Product (BDP) scenario, while considerably improving TCP Cubic. PFIFO deteriorates both TCP Cubic and TCP New Reno in high BDP. Almost in all cases CoDel seems to improve fairness.