Download presentation

Presentation is loading. Please wait.

Published byDominick Haddon Modified about 1 year ago

1
Dynamic Traffic Engineering Techniques for the Internet Federico Larroca Supervisor: Jean-Louis Rougier December 18th 2009

2
page 1 Introduction Traffic Engineering has regained the research community interest. Why? Increasing complexity and dynamicity of traffic -(new) Services convergence -External routing modifications -Large-volume network attacks -Flash crowds -Equipment failures Blind overprovisioning still a viable solution? -Ever increasing access rates (e.g. FTTH) -New architectures with intrinsically scarce resources (e.g. Wireless) -Greedy and irresponsible waste of resources with non-negligible environmental impact Dynamic Traffic Engineering Federico Larroca

3
page 2 Introduction Network Operators in need of TE techniques: Robust with respect to changes in traffic demands Tolerant to node/link failures Efficient in the use of resources Automatic so as to ease network management Dynamic Load-Balancing Answer: Dynamic Load-Balancing Each Origin-Destination (OD) pair connected by several pre-established paths How to distribute traffic among paths so as to optimize a certain objective function? On-line adaptation: function is “always” optimized Dynamic Traffic Engineering Federico Larroca

4
page 3 Introduction Example: Minimizing the maximum link utilization Dynamic Traffic Engineering Federico Larroca Intranet

5
page 4 Contributions of the Thesis Resulting performance clearly depends on the objective function. How to choose it? Which performs better out of the previously proposed ones? Are they the only possibilities? How may the optimum be attained? To avoid oscillations, previously proposed algorithms are conservatively slow Is there a way to adapt convergence speed on-line? Alternatives to Dynamic Load-Balancing? Which one is better? In which case? Dynamic Traffic Engineering Federico Larroca

6
page 5 Agenda Introduction Objective Function Attaining the Optimum Evaluation Conclusions and Future Work Dynamic Traffic Engineering Federico Larroca

7
page 6 (Brief) Notation l=1,..,L : links (c l is its capacity) s=1,..,S : OD pairs (or commodities) d s : traffic demand (amount from origin to destination) P si : one path of OD pair s (i=1,..,n s ) d P si : amount of traffic sent along P si ∑ d P si = d s d P si ≥ 0 d : the vector of d P si ’s ρ l : load on link l Dynamic Traffic Engineering Federico Larroca

8
page 7 Objective Function I Link Utilization (u l ) A Link Utilization u l close to 1.0 means a link operating near its capacity To support sudden increases in traffic and link/node failures: Keep it as low as possible! Dynamic Traffic Engineering Federico Larroca

9
page 8 Objective Function II Path Available Bandwidth (ABW P ) Twofold path performance indicator: Rough estimation of TCP rate on path P (current conditions) How much extra traffic supports path P (prudence) How should ABW P be distributed among paths? Idea: draw on Congestion Control ideas Dynamic Traffic Engineering Federico Larroca

10
page 9 Objective Function III Link Congestion (C l ( l )) C l ( l ) (non-decreasing and continuous) measures the congestion on link l Use what as C l ( l )? The mean queuing delay (D l ( l )) Simplicity: total path queuing delay D P = ∑ D l ( l ) Versatility -Streaming traffic: bigger queuing delays mean more delay and jitter -Elastic traffic: bigger queuing delays mean bottlenecked flows Dynamic Traffic Engineering Federico Larroca

11
page 10 Classic Objective Function III Anyway, use what as the mean queue size f l ( l )? Typical answer: assume a queuing model (e.g. M/M/1) Unrealistic and arbitrary choice! Idea: why not learn it from measurements? Complications Robustness of the estimation Estimated f l ( l ) still convex continuous and non-decreasing Solution: Convex Nonparametric Weighted Least Squares (CNWLS) Dynamic Traffic Engineering Federico Larroca

12
page 11 Agenda Introduction Objective Function Attaining the Optimum Evaluation Conclusions and Future Work Dynamic Traffic Engineering Federico Larroca

13
page 12 Converging to the Optimum Necessary (and sufficient) condition of the optimum d* Equilibrium of greedy OD pairs minimizing a certain path cost function P (Wardrop Equilibrium) E.g. Minimum-Congestion New Question Given P, how may we convergence to the Wardrop Equilibrium (WE) ? no-regret Possible answer: all OD pairs use a no-regret algorithm Intuitively: The mean extra cost incurred by not having used ONLY the cheapest path (regret) is bounded by zero Dynamic Traffic Engineering Federico Larroca

14
page 13 The Algorithm: iAWM Our choice: Incrementally Adaptive Weighted Majority It works with. In particular we used Dynamic Traffic Engineering Federico Larroca The path’s regret (measures how bad it performed) Controls the speed. Self-regulated based on the best regret!! Simple: best paths route more traffic Regret: accumulated difference with outcome y t s

15
page 14 A Flow-Level Simulation Real Network with real TMs (one every 5 min.) We apply iAWM every 1 min. Dynamic Traffic Engineering Federico Larroca

16
page 15 A Simulation Results for the Minimum-Congestion Load-Balancing Dynamic Traffic Engineering Federico Larroca Excellent! Very Bad! What’s wrong? One of the paths has a very bad history (very big regret L t P ) When in an anomalous situation history should be ignored: iAWM-R iAWM-R When a suspicious situation occurs for a certain number of consecutive times: Reset!

17
page 16 Agenda Introduction Objective Function Attaining the Optimum Evaluation The three objective functions The cost of an arbitrary choice A comparison with Robust Routing Conclusions and Future Work Dynamic Traffic Engineering Federico Larroca

18
page 17 The Three Objective Functions : MinMaxU (Minimum Maximum Utilization) : MaxU (Maximum Utility) : MinQ (Minimum Queue) How do they perform? Dynamic Traffic Engineering Federico Larroca Link Utilization (u l ) Path Available Bandwidth (ABW P ) Total Mean Congestion (∑f l (ρ l ))

19
page 18 Results in Abilene For each Objective Function and for each TM, repeatedly apply iAWM-R for the corresponding P until convergence Dynamic Traffic Engineering Federico Larroca Link Utilization (u l ) MinMaxU - MinQMinMaxU - MaxU Conclusions: MinMaxU performs only slightly better MinQ and MaxU perform very similarly

20
page 19 Results in Abilene Dynamic Traffic Engineering Federico Larroca Path Available Bandwidth (ABW P ) MaxU / MinMaxUMaxU / MinQ Conclusions: MaxU outperforms MinMaxU MinQ and MaxU perform very similarly

21
page 20 Results in Abilene Dynamic Traffic Engineering Federico Larroca Total Mean Congestion ( ∑f l (ρ l ) ) (MaxU or MinMaxU) / MinQ Conclusion: MinQ outperforms MinMaxU and MaxU

22
page 21 Agenda Introduction Objective Function Attaining the Optimum Evaluation The three objective functions The cost of an arbitrary choice A comparison with Robust Routing Conclusions and Future Work Dynamic Traffic Engineering Federico Larroca

23
page 22 Results in Abilene Dynamic Traffic Engineering Federico Larroca Link Utilization (u l ) Conclusion: The precise choice of f l ( l ) is not significant MinMaxU - MinQ(CNWLS)MinMaxU - MinQ(M/M/1) What happens when we use the M/M/1 instead of our regression? What is the impact? For each Objective Function and for each TM, repeatedly apply iAWM-R for the corresponding “estimation” of f l ( l ) (CNWLS or M/M/1) until convergence

24
page 23 Results in Abilene Dynamic Traffic Engineering Federico Larroca Total Mean Congestion ( ∑f l (ρ l ) ) Conclusion: M/M/1 obtains very poor results MinQ(M/M/1) / MinQ(CNWLS)

25
page 24 Agenda Introduction Objective Function Attaining the Optimum Evaluation The three objective functions The cost of an arbitrary choice A comparison with Robust Routing Conclusions and Future Work Dynamic Traffic Engineering Federico Larroca

26
Robust Routing all uncertainty set Unique routing configuration for all possible traffic matrices in some uncertainty set X Uncertainty set: -Largest values of links load previously seen -A set of previously observed TMs Objective: Optimize worst-case performance page 25Dynamic Traffic Engineering Federico Larroca

27
page 26 DLB vs RR Dynamic Traffic Engineering Federico Larroca Case 1 – Normal operation Conclusions: DLB and RR perform similarly Under normal operation RR is enough Maximum Link UtilizationTotal Mean Congestion

28
page 27 DLB vs RR Dynamic Traffic Engineering Federico Larroca Case 2 – TM outside X Conclusion: If the traffic demand is outside the uncertainty set the obtained performance in RR is very poor DLB obtains very good results Maximum Link UtilizationTotal Mean Congestion

29
page 28 DLB vs RR Dynamic Traffic Engineering Federico Larroca Case 3 – Big X Conclusion: If the uncertainty set is too big, the obtained performance for a particular demand may be poor Maximum Link UtilizationTotal Mean Congestion

30
page 29 Agenda Introduction Objective Function Attaining the Optimum Evaluation The three objective functions The cost of an arbitrary choice A comparison with Robust Routing Conclusions and Future Work Dynamic Traffic Engineering Federico Larroca

31
page 30 Conclusions Which is the best objective function? MinQ ( min ∑f l (ρ l ) ) How may the optimum be attained? Is there a way to adapt convergence speed on-line? Yes! No-Regret Algorithms Very fast and stable Alternatives to Dynamic Load-Balancing? RR is adequate if no anomalies Else, some form of dynamism is needed! and Traffic and Topology Uncertainty?? Dynamic Traffic Engineering Federico Larroca

32
page 31 Future Work Wireless Mediums What is the best objective function? No-Regret algorithms for on-line adaptation (e.g. power control) Extending DLB to a pure-IP architecture Multi-Topology routing? (problem: choosing the logical topologies) Future Internet: multi-homing and LISP (Locator/ID Separator Protocol) will generalize multi-path How should load-balancing be performed? Dynamic Traffic Engineering Federico Larroca

33
page 32 Publications Pedro Casas, Federico Larroca, Jean-Louis Rougier and Sandrine Vaton, “Comparative Study and New Directions of Traffic Engineering Techniques for Dynamic Traffic,” submitted for fast-tracking in Computer Communications (COMCOM) Journal (Elsevier). Federico Larroca and Jean-Louis Rougier, “Minimum delay load-balancing via nonparametric regression and no-regret algorithms,” submitted to IEEE/ACM Transactions on Networking. Pedro Casas, Federico Larroca, Jean-Louis Rougier and Sandrine Vaton, “Robust Routing vs Dynamic Load- Balancing: A Comprehensive Study and New Directions" in proceedings of the 7th International Workshop on the Design of Reliable Communication Networks (DRCN 2009). Washington D.C., USA, October Pedro Casas, Federico Larroca and Sandrine Vaton, “Robust Routing Mechanisms for Intradomain Traffic Engineering in Dynamic Networks" in proceedings of IEEE/IFIP 6th Latin American Network Operations and Management Symposium (LANOMS 2009). Punta del Este, Uruguay, October Federico Larroca and Jean-Louis Rougier, “Robust Regression for Minimum-Delay Load-Balancing" in proceedings of the 21st International Teletraffic Congress (ITC 21). Paris, France, September Federico Larroca and Jean-Louis Rougier, “Routing Games for Traffic Engineering" in proceedings of the IEEE International conference on Communications (ICC 2009). Dresden, Germany, June Federico Larroca and Jean-Louis Rougier, “Minimum-Delay Load-Balancing Through Non-Parametric Regression" in proceedings of the 8th International IFIP-TC 6 Networking Conference (NETWORKING ‘09). Aachen, Germany, May Federico Larroca and Jean-Louis Rougier, “A Fair and Dynamic Load-Balancing Mechanism" in proceedings of the International Workshop on Traffic Management and Traffic Engineering for the Future Internet (FITraMEn 08). Porto, Portugal, December Selected to appear on Traffic Management and Traffi Engineering for the Future Internet, First Euro-NF International Workshop, FITraMEn 2008, Porto, Portugal, December 11-12, 2008, Revised Selected Papers edited by LNCS, Springer. Andrés Ferragut, Daniel Kofman, Federico Larroca, and Sara Oueslati, “Design and analysis of flow aware load balancing mechanisms for multi-service networks" in proceedings of the 4th EURO-NGI Conference on Next Generation Internet Networks (NGI 2008). Krakow, Poland, April Andrés Ferragut, Daniel Kofman, Federico Larroca, and Sara Oueslati, “Design and analysis of flow aware load balancing mechanisms for multi-service networks - Extended Abstract" presented at the EuroFGI Workshop on IP QoS and Traffic Control. Lisbon, Portugal, December Ingénierie de Trafic Dynamique F. Larroca

34
Thank you! page 33Dynamic Traffic Engineering Federico Larroca

Similar presentations

© 2016 SlidePlayer.com Inc.

All rights reserved.

Ads by Google