1. S. Suri, M, Waldvogel, P. Warkhede CS University of Washington Profile-Based Routing: A New Framework for MPLS Traffic Engineering

2. Overview Dynamic routing of bandwidth guaranteed flows Online Goal –minimize number of rejections –maximize network utilization

3. Assumptions ingress-egress nodes are known traffic profile between pairs of ingress- egress nodes are known –aggregate expected traffic between ie pairs –inferred from SLA or measured –ex: avg bw requirement over a certain time period –profile is a good predictor MPLS networks

4. Routing Requirements No splitting: A flow should be routed on a single path Online routing: No knowledge of future requests Must be fast and scalable Should be able to handle additional policy constraints Traffic Profile

5. Motivation and Review Current routing schemes Shortest path: –simple –may create bottlenecks –may lead low network utilization Widest Shortest Path: choose shortest path with largest residual capacity

6. Minimum Interference Routing (MIRA) (by Kodialam & Lakshman): Idea: avoid routing a flow along paths that can reduce max-flow value between some other ie pair –no true admission control –may cause high # of rejections, low utilization –computationally very expensive

7. Minimum Inference Routing (MIRA) INPUT: G(N,L) with residual capacities Request ((a,b),D) OUTPUT: A route between (a,b) ALGORITHM: –for each ie pair\(a,b) compute maxflow, critical links and weights –eliminate links with residual capacity<D –use Dijkstra to compute shortest path –update residual capacities along the path –route the demand

8. Example I Online LSP requests arrives in order (S0,D0), (S1,D1), (S2,D2),..,(Sn,Dn) with bandwidth requirement of 1.

9. Example II Online LSP requests arrives in order (S0,D), (S1,D), (S2,D),..,(Sn,D) with bandwidth requirement of n,1,1,..,1.

10. Example III Online LSP requests arrives in order (S0,D),..,(S0,D), (n of them, with bw=1) & (S1,D), (S2,D),..,(Sn,D) with bw=1

11. Profile-Based Routing Given a set of LSP requests, what is the max. number of requests that can be routed? NP-Complete. Problem: Unsplittability Two phases: Offline (Preprocessing) phase: use multi- commodity flow framework on traffic profiles

12. Offline phase: (cont.) Goal: route as much commodity as possible Linear Programming: update G to have feasible solution always

13. Offline phase: (cont.) –x i (e) amount of commodity i routed through edge e –Solve for G’ with appropriate constraints. Output: x i (e) To maximize network utilization, e’s capacity is preallocated for each class.

14. Online phase: –(on each edge, residual capacities for each traffic class is kept) –route each LSP request as they arrive –update appropriate residuals

16. RESULTS Worst-Case Results

17. RESULTS (cont.) Simulation bandwidth [1,..,4]

18. RESULTS (cont.) Effect of increasing maximum bandwidth requested [1,..,48]

19. RESULTS (cont.) Bandwidth Fragmentation –causes deviation from upper bound –how bad is it?

20. RESULTS (cont.) Amount of bw wasted increases with larger requests, but small (4% of link capacity at worst) What if expected flows aren’t requested? To measure, look at the snapshots: –what fraction of incoming requests accepted –if PBR is aggressively rejecting at the beginning, performance will be lower at the beginning

21. RESULTS (cont.)

22. Conclusion and Extensions accepts more flows computationally more efficient preprocessing phase can be extended by using different cost functions to provide –minimum service level –fairness

23. Conclusion and Extensions what if profiles are not accurate, how to track it if a request does not arrive for a long time, can we make resources available to others in bw guaranteed environment