Trip-time-based Congestion Detection Julien Chaumond julien.chaumond@polytechnique.org July 29, 2006 Abstract To regulate network traffic and ensure that every user can get a reasonable quality of service whatever the load, networking protocols need to implement some sort of congestion control. To be able to control congestion, one must first know when it occurs and to which extent. This is the scope of congestion detection. Till now, con- gestion detection has only been based on packet losses. Protocols like TCP assume that a packet loss is a conse- quence and a signal of thenetwork’s overload. Anyway, this detection scheme has a number of flaws 1 . In this report, we discuss the possibility of detecting conges- tion from the packets’ trip times. We first analyse the characteristics and distribution of One-way Trip Time or OTT. We then design a scheme of detection based on OTT measurements, and validate this scheme by simu- lation over a large range of situations. 1 Distribution of the One-way Trip Time 1.1 Generalities The Queueing theory provides a useful framework for analyzing the distribution of trip times. Of course, re- sults obtained from the models we use will have to be corrected by real-world experiments. However our anal- ysis will provide some interesting insights on the charac- teristics and dynamics of trip times, which will be used to design algorithms. There is obviously no deterministic expression of the One-way Trip Time in the general case. OTT can be 1 First, it denies the fact that losses can occur intrinsically, without congestion. With the growing number of wireless links over the Internet, this assumption has to be put into question. Moreover, the TCP-like congestion detection scheme relies on fre- quent acknowledgements and thus does not apply well to multi- media traffic, which does not want to support the excessive cost of reliability. viewed as the sum of: • the physical transmission delay • a sum of buffering times in routers’ queues along the path • the processing times at the routers Of these three terms, only the first one may be de- terministic. The other two depend on the nature and intensity of the considered source’s traffic as well as of the cross-traffic. What we call transmission delay is the delay of the transmission of one byte of a packet. From now on, we assume that the processing time of a packet is a deter- ministic function of the packet’s size, namely α(|P |). 1.2 Multiple-hop models The multiple-hop model is a framework for modeling network routes. A network route is modeled as a se- quence of n hops. Each hop is a FIFO queue with infi- nite buffer representing the queue of a store-and-forward router, followed by the corresponding transmission link. Assume that the path between the sender and the receiver contains n intermediate routers. Then: OTT = n+1  j =1 D j + α(|P |)+ n  i=1 B i where D j is the physical transmission delay along the j th link and B i is the buffering time at the i th router. We assume that the physical route does not change in the middle of a connection. Therefore, ∑ n+1 j =1 D j is a constant value. We will also assume that the dependence of the pro- cessing time on the packet’s size is linear, that is to say: α(|P |)= α.|P |. We now have to find a model for the sucessive FIFO queues. The objective is to point out some characteris- tics of the buffering times B i . 1