1. Towards a More Fair and Robust Internet Backbone Year 1 Status Report Rene Cruz, Tara Javidi, Bill Lin Center for Networked Systems University of California, San Diego July 20, 2006

2. CNS Research Review, July 19-20, 20062 Internet Routing Today  OSPF: packets are routed from source to sink via minimum cost (shortest) paths

3. CNS Research Review, July 19-20, 20063Problems  Node failures  Can no longer route along OSPF path  Difficult to predict future Internet traffic patterns  Actual traffic down the road may be different than what was expected when network expansion took place

4. CNS Research Review, July 19-20, 20064 Basic Approach  Load-balance across multiple paths to mitigate both node failures and unpredictable traffic changes

5. CNS Research Review, July 19-20, 20065 Two Problem Formulations  Optimize for a given traffic profile  Use state-of-the-art network measurement techniques to measure traffic conditions  Find at run-time optimal path load-distribution for measured traffic profile  No need to change previous path load-distribution if it can handle current traffic condition – only change when necessary  Optimize for worst-case, assuming all “admissible” traffic profiles are possible  Determine optimal path load-distribution under this worst-case model once offline

6. CNS Research Review, July 19-20, 20066 Optimizing for Specific Traffic Profile  Use a centralized control plane model – e.g. AT&T’s RCP proposal  Periodic traffic measurements reported back to central controller  For given traffic matrix T, find optimal path load- distribution as a Maximum Concurrent Flow (MCF) problem  We’ve implemented a very fast approximate MCF solver that can quickly compute optimal path load-balancing

7. CNS Research Review, July 19-20, 20067 Maximum Concurrent Flow maximize s.t.  P:e in P x(P) <= u(e) forall e in E  P in P(i, j) x(P) >= T(i, j) forall (i, j) x(P) >= 0 forall P

8. CNS Research Review, July 19-20, 20068 Initial Results  Robustness  Different experiments measured over an hour each at 5 min intervals  Compare possible with MCF vs. OSPF  The for MCF is up to 50% better than OSPF  Link utilization  Over the same hour periods, compare number of links that are more than 50% loaded  OSPF has about 5x more links with greater than 50% load vs. MCF …  Which indicates MCF is effective in spreading load

9. CNS Research Review, July 19-20, 20069 Two Problem Formulations  Optimize for a given traffic profile  Optimize for worst-case, assuming all “admissible” traffic profiles are possible

10. CNS Research Review, July 19-20, 200610 “Admissible” Traffic Model  R i = max ingress traffic at node i, C i = max egress traffic at node i  Traffic matrix T is admissible iff   j t ij <= R i,  j t ji <= C i, for all i  Set of all such traffic matrices denoted by  ( R, C ) CiCi RiRi

11. CNS Research Review, July 19-20, 200611 “Admissible” Traffic Model  Maximizing for the set of all admissible traffic matrices  ( R, C ) can be formulated as an LP problem  We are currently implementing this approach

12. CNS Research Review, July 19-20, 200612 Continuing into Year 2 …  Continue to explore the optimization problem for the “admissible” traffic model  Investigate MCF and LP framework to restrict the paths considered. e.g.  Consider only shortest length paths  Consider only paths that are with some % of the shortest length paths  Consider notions of path delays other than number of hops  Investigate into failure models and robustness of multipath routing  More generally, investigate more into what optimization criteria to use

13. Questions?