Theoretical analysis of performances of TCP/IP congestion control algorithm with different distances

According to current TCP/IP implementations, the acceleration in additive-increase phase depends on the distance of connection. In this paper, the performance of Additive Increase and Multiplicative Decrease (AIMD) congestion control algorithm in TCP is analyzed in two ways, both focusing on effects of the heterogeneity or the mixture of different accelerations caused by different distances. First, we analyze flow time minimization, extending the competitive analysis by Edmonds et al. to heterogeneous case. We show (a) the performance loss of TCP/IP in Long Fat Pipe Networks (LFNs) is caused by the heterogeneity rather than long distance itself. Next, we step forward to more realistic single-drop model, where upon each congestion only one, instead of all, connection drops rate, and analyze asymptotic total and per-connection bandwidth utilizations. We show (b) increasing the number of connections makes total utilization better as opposed to common model, (c) in homogeneous environments, victim policies or choice of which connection drops do not affect total utilization, and (d) in heterogeneous two-connection environments, maximum total utilization is achieved by certain victim policy which leads to unfair share, whereas fair utilization is achieved by certain random victim policy.