Modeling of Multiple TCP Segment Losses in Slow Start and Congestion Avoidance Phases

摘要

This paper presents analytical modeling to systematically capture the characteristics of TCP fast retransmission and recovery. One of the novel contributions of this paper is to evaluate the conditions under which fast recovery can be successful in the slow start and in the congestion avoidance phases, respectively. Unlike previous works where they simply assume in a RTT round the packets following the first lost packet are all lost, we consider a more realistic case in which the packets following the first lost packet in the same RTT round may not be all lost. This consideration makes our analyses technically different from the previous works as described in [1] and [4-6]. Depending on N_r[y], the number of packets successfully received at the (y + l)th RTT round after the first lost packet was detected, we divide the analytical model into two cases: N_r[y]≥ k (the number of duplicate acknowledgments) and N_r[y] < k. For both cases, we have respectively revealed that there are certain necessary conditions for the fast retransmission to occur. The revealed necessary conditions are represented as a number of upper bounds, beyond which the lost packets can not be recovered using TCP fast retransmission algorithm. In addition, from the analytical results, we have shown that TCP Reno certainly leads to timeout if the number of packet losses in a TCP window exceeds two packets.