Using Partial Differential Equations to Modek TCP Mice and Elephantsin large IP Networks M. Ajmone Marsan, M. Garetto, P. Giaccone, E. Leonardi, E.Schiattarella, A. Tarello Politecnico di Torino - Italy Hong-Kong – March 7-11 , 2004 TANGO 1 Outline Dimensioning IP networks Queuing network models Fluid approaches Conclusions 2 Consideration Over 90 % of all Internet traffic is due to TCP connections TCP drives both the network behavior and the performance perceived by end-users Analytical models of TCP are a must for IP network design and planning 3 Consideration Accurate TCP models must consider: closed loop behavior short-lived flows multi-bottleneck topologies AQM schemes (or droptail) QoS approaches, two-way traffic, ... 4 Problem statement 1 2 finite flows (mice) F URLs/sec 2 3 greedy flows IP core URLs/sec finite flows 3 N greedy flows (elephants) F ... 4 N 4 5 Problem statement Input variables: only primitive network parameters: IP network: channel data rates, node distances, buffer sizes, AQM algorithms (or droptail), ... TCP: number of elephants, mice establishment rates and file length distribution, segment size, max window size, ... Output variables: IP network: link utilizations, queuing delays, packet loss probabilities, ... TCP: average elephant window size and throughput, average mice completion times, ... 6 Modeling approach Abandon a microscopic view of the IP network behavior, and model packet flows and other system parameters as fluids The system is described with differential equations Solutions are obtained numerically 7 Modeling approach A simple example: One bottleneck link RED buffer Elephants only (no slow start) 8 TCP model dWs(t)/dt = 1/Rs(t) – Ws(t) s(t) / 2 Where: • Ws(t) • Rs(t) • s(t) is the average window is the average round trip time is the congestion indication rate of TCP sources of class s at time t 9 IP network model dQk(t)/dt = Σs Ws(t) (1-P(t)) / Rs(t) – - C 1{Qk(t)>0} Where: • Qk(t) is the length of queue k at time t 10 IP network model Rs(t) = PDs + Qk(t)/C Where: • PDs is the propagation delay for source s 11 Problems Difficult to deal with mice since the start time of each mouse detemines the window dynamics over time. One equation shoud be written for each mouse Difficult to consider droptail buffers due to the intrinsic burstiness of the loss process experienced by sources 12 Problems Difficult to deal with mice since the start time of each mouse detemines the window dynamics over time. One equation shoud be written for each mouse 13 Our Approach Consider a population of TCP sources: P(w,t) is the number of TCP flows that at time t have window not greater than w. . P(w,t) w window Partial differential equations are obtained 14 Basic source model Where: 15 Mice Source Equations 16 Fluid models – extensions • Slow start (mice) • Finite window • Threshold • Fast recovery • Droptail buffers •Core network topologies 17 Fluid models – results 18 Fluid models – model results 19 Fluid models – NS results 20 Fluid models – model results 21 Fluid models – NS results 22 Fluid models – results 23 Fluid models – results 24 Fluid models – results 25 Fluid models – results 26 Fluid models – results 27 Fluid models – results We obtained results for the GARR network with over one million TCP flows, and link capacities up to 2.5 Gb/s. 28 Conclusions Fluid models today seem the most promising approach to study large IP networks Tools for the model development and solution are sought Efficient numerical techniques are a challenge 29 Conclusions The modeling paradigms to study the Internet behaviour are changing This is surely due to scaling needs, but probably also corresponds to a new phase of maturity in Internet modeling Reliable predictions of the behavior of significant portions of the Internet are within our reach 30 Thank You ! 31 Outline The Internet today Dimensioning IP networks Queuing network models Fluid approaches Conclusions 32 Source: Internet Software Consortium (http://www.isc.org/) 33 Source: Internet Traffic Report (http://www.internettrafficreport.com/) 34 Source: Internet Traffic Report (http://www.internettrafficreport.com/) 35 Source: Sprint ATL (http://ipmon.sprint.com/packstat) April 7th 2003, 2.5 Gbps link 36 Source: Sprint ATL (http://ipmon.sprint.com/packstat) April 7th 2003, 2.5 Gbps link 37 Source: Sprint ATL (http://ipmon.sprint.com/packstat) April 7th 2003, 2.5 Gbps link 38 Source: Sprint ATL (http://ipmon.sprint.com/packstat) April 7th 2003, 2.5 Gbps link 39 Source: Sprint ATL (http://ipmon.sprint.com/packstat) April 7th 2003, 2.5 Gbps link 40 And still growing ... Subject: [news] Internet still growing 70 to 150 per cent per year Date: Mon, 23 Jun 2003 09:55:45 -0400 (EDT) From: CAnet-NEWS@canarie.ca ... Andrew Odlyzko, director of the Digital Technology Center at the University of Minnesota, ... says Internet traffic is steadily growing about 70 percent to 150 percent per year. On a conference call yesterday to discuss the results, he said traffic growth slowed moderately over the last couple of years, but it had mostly remained constant for the past five years. ... 41 Literature V. Misra, W. Gong, D. Towsley, "Stochastic Differential Equation Modeling and Analysis of TCP Windowsize Behavior“, Performance'99 T. Bonald, "Comparison of TCP Reno and TCP Vegas via Fluid Approximation," INRIA report no. 3563, November 1998 V. Misra, W. Gong, D. Towsley, "A Fluid-based Analysis of a Network of AQM Routers Supporting TCP Flows with an Application to RED“, SIGCOMM 2000 42 Literature Y.Liu, F.Lo Presti, V.Misra, D.Towsley, Y.Gu, "Fluid Models and Solutions for Large-Scale IP Networks", SIGMETRICS 2003 F. Baccelli, D.Hong, "Interaction of TCP flows as Billiards“, Infocom 2003 F.Baccelli, D.Hong, "Flow Level Simulation of Large IP Networks“, Infocom 2003 43 Literature T. Lakshman and U. Madhow, "The performance of TCP/IP for networks with high bandwidth-delay products and random loss," IEEE/ACM Transactions on Networking, vol. 5, no. 3, 1997. M.Ajmone Marsan, E.de Souza e Silva, R.Lo Cigno, M.Meo, “An Approximate Markovian Model for TCP over ATM”, UKPEW '97 J. Padhye, V. Firoiu, D. Towsley, J. Kurose, "A Stochastic Model of TCP Reno Congestion Avoidance and Control“, UMASS CMPSCI Technical Report, Feb 1999. 44 Literature C.Casetti, M.Meo, “A New Approach to Model the Stationary Behavior of TCP Connections”, Infocom 2000 M.Garetto, R.Lo Cigno, M.Meo, E.Alessio, M.Ajmone Marsan, “Modeling Short-Lived TCP Connections with Open Multiclass Queueing Networks”, PfHSN 2002 A.Goel, M.Mitzenmacher, "Exact Sampling of TCP Window States", Infocom 2002 45 Consideration Developing accurate analytical models of the behavior of TCP is difficult. A number of approaches have been proposed, some based on sophisticated modeling tools. 46 Fluid models – results 47