On the Interaction between Dynamic Routing in the Native and Overlay Layers

1. On the Interaction between Dynamic Routing in theNative and Overlay Layers INFOCOM 2006 Srinivasan Seetharaman Mostafa Ammar College of Computing Georgia Institute of Technology

2. Inter-Layer Interaction Problem • Infrastructure overlay networks offer better services by deploying intelligent routing schemes. • Uncoordinated dynamic routing in the two layers lead to many problems. • We focus on the effect of native link failures, as they trigger each layer to reroute independently • Dual Rerouting

3. Temporal Dynamics Consider a native link failure in CE • Only one overlay link is affected. • The native path AE is rerouted over F(ACE → ACFDE) 2 G 3 A I 3 2 4 E OVERLAY NATIVE G B I A F C H + ∞ E D Cost Overlay recovery: 8 Original: 2 Native Rerouting: 2 Overlay rerouting: 4 Time Native Failure Native Recovery Native Repair

4. Downside to Dual Rerouting • Overlap of functionality between layers causing large number of route flaps (oscillations) • Unawareness of other layer’s decisions leading to • resource overloading, • multiple simultaneous failures • a low success rate in rerouting • sub-optimal paths after rerouting • Lack of flexibility and control

5. Problem Statement I • Assume the two ends of each link (native & overlay) use a keepAlive protocol for link verification. • 3 keepAlive messages lost  Failure • Understand the effects of different parameters on the rerouting performance. • KeepAlive-time: Time between two keepAlive messages • Hold-time: Time window to declare link as down • Overlay link cost scheme (Ex: Native hops, Overlay hops)

6. Performance Metrics • Hit-time: Time taken for traffic to be recovered. = Detection time + Convergence time + Device time (depends on timers) (protocol specific) (Negligible) • Success rate of recovery Success rate of a layer = Number of paths recovered Number of failed overlay paths • Number of route flaps Average route flaps = Number of route flaps Number of failed overlay paths • Peak & Stabilized inflation (before repair) Path cost inflation = Path cost after rerouting Path cost before failure

7. Temporal Dynamics Overlay path AE Overlay detects first 100% success rate 3 route flaps Peak inflation = 8/2 Stabilized inflation = 4/2 Hit time ∞ Cost Overlay recovery: 8 Original: 2 Native Rerouting: 2 Overlay rerouting: 4 Time Native Failure Native Recovery Native Repair

8. Performance Evaluation – ns2 • Using GT-ITM, we randomly generate: 25 topologies = (5 overlay network) x (5 native network) • Two scenarios • Inspect intra-domain failures in single-domain native network • Inspect inter-domain failures in multi-domain native network • In each scenario, tabulate failure recovery statistics of all overlay paths by breaking one native link at a time

9. Effect of Routing Parameters Observations: By varying the overlay keepAlive-time, hold-time and cost scheme, we observe: • hold-time  hit time (only until overlay hold-time < native hold-time) • hold-time  # route flaps  • hold-time  sub-optimality  • keepAlive-time hit-time hold-time

10. Conclusion I Dual rerouting can be made optimal by adopting the following recommendations: • Overlay hold-time very close to the native hold-time. • Overlay keepAlive-time that is half that of the hold-time as it leads to an earlier detection.

11. Problem Statement II • Main observation from previous simulations: • “Native-rerouting yields the optimal path, albeit a bit later” • Make the overlay layer aware of this observation and give higher precedence to native rerouting attempts • Improve overlay routing performance by adjusting the overlay layer functioning

12. Three Levels of Layer Awareness • No awareness • Dual rerouting • Awareness of native layer’s existence: • Probabilistically Suppressed Overlay Rerouting (PSOR): Suppress overlay rerouting attempt with probability ‘p’ • Deferred Overlay Rerouting (DOR): Delay overlay recovery by time ‘d’

13. Three Levels of Layer Awareness (contd.) • Awareness of native layer’s parameters: • Follow-on Suppressed Overlay Rerouting (FSOR) If follow-on time < threshold ‘f’, then suppress overlay rerouting Follow-on time Time Overlay layer detects failure Native layer detects failure Failure

14. Effect of Adjusting Overlay • All three schemes are simple and offer significant control over the tradeoffs between hit-time and the other metrics. • PSOR: • Least number of route flaps • Least peak inflation • DSOR and FSOR behave similarly (FSOR has slightly better hit-time): • Better success rate • Lower stabilized inflation

15. Conclusion II By appropriately tuning • keepAlive-time • hold-time • suppression probability • delay • follow-on threshold …we can improve results for: • Hit-time • # Route flaps • Path cost inflation • Stabilization time • Success rate

16. Problem Statement III • Main observation from previous simulations: • “It is not possible to improve all metrics simultaneously. Hence, performance is still bounded!” • As overlay applications proliferate, the native layer should gradually evolve to suit them • Improve overlay routing performance by adjusting the native layer functioning

17. Tuning the Native keepAlive-time • We adopt a non-invasive procedure to advance the native layer rerouting • Tuning of the native layer keepAlive-time • Constraints: • Tuning should not generate any extra overhead • Effective detection time should be same

18. Tuning the Native keepAlive-time (contd.) Consider the following scenarios for tuning. Scenario B is vanilla Dual rerouting Scenario A is the layer-aware overlay rerouting scheme Scenario C is the tuning we recommend here

19. Conclusions III • Native layer tuning we proposed achieves the best performance in all our metrics

20. Summary • We propose means to mitigate the problems associated in the inter-layer interaction • We explore two directions: • Adjusting the overlay layer functioning • Adjusting the native layer functioning