Routing Games for Traffic Engineering F. Larroca and J.L. Rougier IEEE International Conference on Communications (ICC 2009) Dresden, Germany, June 14-18 2009 Introduction Current traffic is highly dynamic and unpredictable How may we define a routing scheme that performs well under these demanding conditions? Possible answer: Dynamic Load-Balancing • We connect each Origin-Destination (OD) pair with several pre-established paths • Traffic distribution depends on current TM and network condition Greedy algorithms on path cost function fP: • Minimum coordination • Ideal case study for game theory: Routing Game page 1 F. Larroca and J.L. Rougier IEEE ICC 2009 Introduction First Contribution: • New routing game designed for elastic traffic • Basic Idea: use load-balancing to further maximize the utility obtained by TCP flows Second Contribution: • Performance comparison of three routing games • Considered games: - Congestion Game - Bottleneck Game - Our proposition page 2 F. Larroca and J.L. Rougier IEEE ICC 2009 Agenda Introduction Basic New Definitions and Results Routing Game Evaluation Conclusions page 3 F. Larroca and J.L. Rougier IEEE ICC 2009 Definitions 3 functions to define a Routing Game: • Link cost: fl(rl) • Path cost: fP=g ({fl(rl)}lϵP) • Social Cost: SC(d) Congestion Game: fl (r l ) 1 cl r l g ( A ) a SC ( d ) a A s rl • Equilibrium minimizes ( d ) l 0 d si f Psi i f l ( x ) dx l rl cl r l instead of SC(d) • To converge to the optimum we should use fˆl ( r l ) cl c l • Example: MPLS adaptive traffic engineering (MATE) [EJLW01] page 4 F. Larroca and J.L. Rougier IEEE ICC 2009 rl 2 Definitions 3 functions to define a Routing Game: • Link cost: fl(rl) • Path cost: fP=g ({fl(rl)}lϵP) • Social Cost: SC(d) Bottleneck Game: fl (r l ) rl cl g ( A ) max a a A SC ( d ) max f P max P l rl cl • Equilibrium and social optimum coincide! • Examples: TeXCP [KKDC05] and REPLEX [FKF06] page 5 F. Larroca and J.L. Rougier IEEE ICC 2009 Agenda Introduction Basic New Definitions and Results Routing Game Evaluation Conclusions page 6 F. Larroca and J.L. Rougier IEEE ICC 2009 New Routing Game: Intuition Assume each OD pair s has exactly Ns TCP flows Congestion Control Problem (x = TCP rate): maximize x N si U si ( x si ) OD pairs ( s ) Paths ( i ) s.t. N s si x si c l i :l i Nsi (flows per path) are given. Why not optimize in both x and N? First idea: à la Multi-Path TCP (optimized by end-users) Our idea: keep the separation between end-to-end congestion control (maximization on x) and routing (maximization on N) page 7 F. Larroca and J.L. Rougier IEEE ICC 2009 New Routing Game: Definition First problem: Considered time-scale • Time-Scale(TCP) << Time-Scale(Routing) • Approximations of xsi and Nsi are necessary: x si ABW N si d si si min c l r l l si (number of flows amount of traffic) Second problem: Usi(x) is not known by routing • Use arbitrary U(x) Result: maximize Umax maximize f ( r l ) UN sicUl si (rxlsi ) g ( A ) max a SC ( d ) d dmin fcUl c lrl r l si l si l si P d l si x OD pairs ( s ) Paths ( i ) a A OD pairs ( s ) Paths ( i )s si i s.t.s.t. not d sid si 0 0 d sid si d sd s SC optimum are the same! i i s i :l i However, we provide an adaptation of fl(r l) Equilibrium s.t. N si x siand cl page 8 F. Larroca and J.L. Rougier IEEE ICC 2009 Agenda Introduction Basic New Definitions and Results Routing Game Evaluation Conclusions page 9 F. Larroca and J.L. Rougier IEEE ICC 2009 Evaluation: simple examples Example 1: Congestion Game is reluctant to use longer paths => bigger maximum link utilization page 10 F. Larroca and J.L. Rougier IEEE ICC 2009 Evaluation: simple examples Example 2: Path lengths relatively similar (even if link capacities are different) => UM and CG obtain similar results (plus: difference with BG not as important) page 11 F. Larroca and J.L. Rougier IEEE ICC 2009 Evaluation: simple examples Example 3: The only mechanism that enforce fairness at a path level is Utility Maximization page 12 F. Larroca and J.L. Rougier IEEE ICC 2009 Evaluation: Realistic Topologies ABWsi is always bigger in our proposal • Not very big over CG in mean (<5%) but significant in the minimum (>15%). Origin: fairness • More important with respect to BG Link utilization relatively similar among all games • CG obtains a bigger maximum (5-10%) page 13 F. Larroca and J.L. Rougier IEEE ICC 2009 Agenda Introduction Basic New Definitions and Results Routing Game Evaluation Conclusions page 14 F. Larroca and J.L. Rougier IEEE ICC 2009 Conclusions and Future Work The proposed game is the most balanced one: • It generally outperforms the rest • When it does not, the difference is not important However, it is more difficult to implement We are interested in the total mean delay • Answer: Congestion Routing Game • Heavily depends on the assumed model • Load-balancing mechanism that converges to the minimum-delay configuration without assuming any model? Yes! [LR09][LR09a] page 15 F. Larroca and J.L. Rougier IEEE ICC 2009 References • [EJLW01]: A. Elwalid; C. Jin; S. Low and I. Widjaja "MATE: MPLS adaptive traffic engineering" INFOCOM 2001. • [KKDC05]: S. Kandula; D. Katabi; B. Davie and A. Charny "Walking the tightrope: responsive yet stable traffic engineering" ACM SIGCOMM '05 • [FKF06]: S. Fischer; N. Kammenhuber and A. Feldmann "REPLEX: dynamic traffic engineering based on wardrop routing policies" CoNEXT '06 • [LR09]: F. Larroca and J.L. Rougier "Minimum-Delay Load-Balancing Through Non-Parametric Regression" IFIP/TC6 NETWORKING 2009 • [LR09a]: F. Larroca and J.L. Rougier "Robust Regression for Minimum-Delay Load-Balancing" 21st International Teletraffic Congress (ITC 21) Thank you Questions? page 16 F. Larroca and J.L. Rougier IEEE ICC 2009