Of Mice and Models Marco Ajmone Marsan, Giovanna Carofiglio, Michele Garetto, Paolo Giaccone, Emilio Leonardi, Enrico Schiattarella, and Alessandro Tarello Dipartimento di Elettronica, Politecnico di Torino, C.so Duca degli Abruzzi 24, Torino, Italy ⋆ firstname.lastname@polito.it Abstract. Modeling mice in an effective and scalable manner is one of the main challenges in the performance evaluation of IP networks. Mice is the name that has become customary to identify short-lived TCP connections, that form the vast majority of packet flows over the Internet. On the contrary, long-lived TCP flows, that are far less numerous, but comprise many more packets, are often called elephants. Fluid models were recently proved to be a promising effective and scalable approach to investigate the dynamics of IP networks loaded by elephants. In this paper we extend fluid models in such a way that IP networks loaded by traffic mixes comprising both mice and elephants can be studied. We then show that the newly proposed class of fluid models is quite effective in the analysis of networks loaded by mice only, since this traffic is much more critical than a mix of mice and elephants. 1 Introduction The traffic on the Internet can be described either at the packet level, modeling the dynamics of the packet generation and transmission processes, or at the flow level, modeling the start times and durations of sequences of packet transfers that correspond to (portions of) a service requested by an end-user. Examples of flows can be either the train of packets corresponding to an Internet telephone call, or the one corresponding to the download of a web page. In the latter case, the flow can be mapped onto a TCP connection. This is the most common case today in the Internet, accounting for the vast majority of traffic. The number of packets in TCP connections is known to exhibit a heavy-tailed distribution, with a large number of very small instances (called TCP mice) and few very large ones (called elephants). Models for the performance analysis of the Internet have been traditionally based on a packet-level approach for the description of Internet traffic and of queuing dynamics at router buffers. Packet-level models allow a very precise description of the Internet operations, but suffer severe scalability problems, such that only small portions of real networks can be studied. Fluid models have been recently proposed as a scalable approach to describe the behavior of the Internet. Scalability is achieved by describing the network and traffic dynamics at a higher level of abstraction with respect to traditional discrete packet- level models. This implies that the short-term random effects typical of the packet-level network behavior are neglected, focusing instead on the longer-term deterministic flow- level traffic dynamics. ⋆ This work was supported by the Italian Ministry for Education, University and Research, through the FIRB project TANGO. M. Ajmone Marsan et al. (Eds.): QoS-IP 2005, LNCS 3375, pp. 15–32, 2005. c  Springer-Verlag Berlin Heidelberg 2005