Throughput-smoothness tradeoff in preventing competing TCP from starvation

In this paper, we systematically study a fundamental tradeoff for TCP friendliness: in order for a congestion control protocol to maintain a certain degree of TCP friendliness, the longer its smoothness timescale is, the lower its average sending rate should be. This throughput-smoothness tradeoff exists not only for the traditional definition of TCP friendliness but also for any congestion control protocols that are designed to prevent competing TCP flows from complete starvation (which we believe is a basic requirement for a traffic control protocol) and then maintain a certain degree of TCP friendliness. Specifically, we derive the TCP-friendly sending rate requirements for three different smoothness timescales (i.e., milliseconds, seconds, and minutes) with a flow-level queueing model, and verify the analytical results with extensive packet-level simulation results. Finally, we propose a new TCP-friendly congestion control protocol for applications that prefer a smooth predictable sending rate on a multi-minute timescale, called Long-Time-Scale TCP-Friendly CBR-Like Rate Control (L-TFCBR).