1. SIGCOMM 2003 Making Intra-Domain Routing Robust to Changing and Uncertain Traffic Demands: Understanding Fundamental Tradeoffs David Applegate Edith Cohen

2. SIGCOMM 2003 Some ISP Challenges Utilize network capacity efficiently QoS Intra-Domain Traffic Engineering is increasingly deployed. Components: Understanding traffic demands Configuring routing protocols so that traffic is routed efficiently

3. SIGCOMM 2003 Financial reports (and traffic demands) Past results are not a guarantee of future performance. Past results are not even a guarantee of past performance.

4. SIGCOMM 2003 Traffic Demands Measurement of traffic data is inexact. –Inference from link loads  estimation errors –Sampled flows  sampling errors –Missing data Traffic demands are dynamic and change on multiple time scales.

5. SIGCOMM 2003 Routing configuration Knowing exact demands values allows for very efficient routings. But..., we don’t have accurate values. Moreover, even if we did… Demands are dynamic. But..., modifications to the routings cause disruptions and reduce QoS. Possible solution: Robust routings

6. SIGCOMM 2003 Robust routings A fixed routing configuration that works well (as well as possible) for a wide range (or all) traffic matrices (TMs). Built-in robustness to changing/unknown conditions is a natural objective of good engineering.

7. SIGCOMM 2003 Challenges Modeling: How to measure robustness? Algorithmic: Given no or some constraints on TMs, how to efficiently compute an optimal robust routing ? Understanding the tradeoff: Quantify the “generality cost”: A fixed routing that is optimized for many TMs may be suboptimal for a particular TM. What to expect?

8. SIGCOMM 2003 Modeling and Metrics: Competitive Analysis Framework Relative rather than absolute metric:  Compare yourself only to the best possible. That is,  For any applicable TM, compare your routing configuration performance to the best possible for that TM.

9. SIGCOMM 2003 Metrics... details Given a routing configuration f and a TM D, we look at the Maximum Link Utilization (MLU) when routing D using f. Performance ratio of f on D: ratio of MLU of f on D to the MLU of the optimal routing configuration for D. Performance ratio of f on a set of TMs is the max performance ratio over TMs in the set.

10. SIGCOMM 2003 Challenges Modeling: How to measure robustness? Algorithmic: Given no or some constraints on TMs, how to efficiently compute an optimal robust routing ? Understanding the tradeoff: Quantify the “generality cost”: A fixed routing that is optimized for many TMs may be suboptimal for a particular TM. What to expect?

11. SIGCOMM 2003 Algorithms for optimal robust (“demand oblivious”) routing Known: [ACFKR:STOC 03] Polynomial time algorithm through an exponential LP formulation using the Ellipsoid algorithm (separation) Our contribution (theoretical and practical): Compact polynomial-size LP formulation. Efficient implementation. Extensions to demand ranges constraints.

12. SIGCOMM 2003 Challenges Modeling: How to measure robustness? Algorithmic: Given no or some constraints on TMs, how to efficiently compute an optimal robust routing ? Understanding the tradeoff: Quantify the “generality cost”: A fixed routing that is optimized for many TMs may be suboptimal for a particular TM. What to expect?

13. SIGCOMM 2003 Understanding the Tradeoffs How well can we do with no knowledge of demands (what is the optimal “oblivious” performance ratio) ? What if we have some knowledge on applicable demands, say, using a “base” TM within some error margins ?

14. SIGCOMM 2003 Data Topologies: Six PoP to PoP ISP topologies from Rocketfuel, aggregated to cities; one topology from [MTSBD 02] 14—57 nodes ; 25—88 links Capacities: heuristic TMs: heuristic, bimodal and gravity TM-sets: All TMs; base bimodal/gravity TM with margins (error bars)

15. SIGCOMM 2003 Routing Configurations Optimal Robust routing for the applicable set of TMs (MPLS-style) (computed using our algorithms) OSPF routing (derived) (supplied with Rocketfuel data) For demand margins: optimal routing for the base TM (MPLS-style) (computed via a mcf LP)

16. SIGCOMM 2003 “Oblivious” Performance Ratio of Routing Configurations ASNPoPslinksOptimalOSPF 1221 Telstra 57591.43 4.2 1755 Ebone 23381.78 16.6 6461 Abovenet 22421.91 13.4 3967 Exodus 22371.62 49.2 3257 Tiscali 50881.80 51.2 1239 Sprintlink 44831.90234.0 N-14 [MTSBD02] 14251.97 7.7

17. SIGCOMM 2003 Scalability [Räcke 02] poly-logarithmic upper bound for symmetric networks; [HHR 03] O(log^2 n log log n) We observe < 2 for ISP networks. Supported by analysis showing that cycles and cliques of any size have <2 ratio. 1.4-1.9 is surprisingly low but probably not good enough to be practical

18. SIGCOMM 2003 Demand Margins

19. SIGCOMM 2003 Demand Margins

20. SIGCOMM 2003 Conclusions from experimental evaluation Can do reasonably well with no knowledge of TM (for all TMs), link utilization +40%- +90% Can do even better for error margins (x4 bars with +25% utilization). Routing designed to be optimal for a somewhat-off TM estimate can be much worse than an optimal demand-oblivious routing.

21. SIGCOMM 2003 Summary of Contributions New analytical framework and algorithms for computing and evaluating robust routing configurations. Experiments showing that optimal robust routings perform well on (Rocketfuel) ISP topologies, and significantly outperform naïve methods (optimize without margins, naïve OSPF)

22. SIGCOMM 2003 Future Robust restoration routing Optimal OSPF-style rather that MPLS- style robust routings Robust routing under varying demand constraints (link load data) More efficient computation Better measure (relative metric places too much emphasis on “easy” TMs)

23. SIGCOMM 2003 Thank you! Non sequitur, Wednesday, August 27, 2003 The media quickly responds to SIGCOMM