1. ISP and Egress Path Selection for Multihomed Networks Amogh Dhamdhere, Constantine Dovrolis Networking and Telecommunications Group Georgia Institute of Technology Presented by Karl Deng

2. Provision the multihoming configuration of a source network S. Problem Definition Inputs: S D = {D i } (i = 1.. M) R = {r i } K

3. Phase I - ISP Selection Select K ISPs that S will subscribe to. Objectives: Optimize monetary cost and availability. Phase II - Egress Path Selection Determine the ISP that should be used to reach each of the M major destinations. Objectives: Select congestion-free paths and minimize cost. (Avoid long-term congestion.) Two Phases Phase-I can be repeated in long time scales, from weeks to months, while Phase-II can be repeated whenever there is a major change in the egress traffic distribution.

4. Phase I - ISP Selection - the set of possible ISPs to which S can subscribe Select K ISPs out of by taking into account of the following three factors:  Monetary cost  AS-level path length  Path diversity possible selections  Exhaustive search For example, if = 15 and K = 4, = 1365

5. Exhaustive search - the set of all possible combinations of K ISPs from the set - total cost by taking into account of all three factors - optimal combination

6. -- normalization factors - monetary cost - cost associated with the AS-level path length - cost associated with path diversity - total cost Calculate the cost for each combination

7. Monetary cost - pricing function of ISP j G - a mapping between ISP and destination e.g., j = G(i), map destination i to ISP j T j depends on G NP-hard and K M possible ways of mapping  Heuristic (FFD-like algorithm) Constraint: T j < A

8. Algorithm-1: a FFD-like algorithm FFD - First Fit Decreasing Basic idea: Start with the largest destination, in terms of traffic rate, and route it through the lowest-cost ISP. It is possible that Algorithm 1 will fail to find a feasible mapping (due to the capacity constraint).

9. Cost associated with the path length p j (i) - AS-level path length to reach a destination i through ISP j. Constraint: T j < A Similar to the monetary cost problem, also use Algorithm-1.

10. Cost associated with the path diversity - number of K-shared links to destination i through the ISPs in - a path diversity metric; indicates the resiliency to single inter-AS link failures

11. LGSs are routers inside an ISP that report AS-level paths to given destination networks. Most ISPs maintain public Looking Glass Servers Looking Glass Servers (LGS) We assume that each ISP in has a LGS from which S can determine the AS-level paths to destinations in D.

12. Phase II - Egress Path Selection Given the K ISPs selected by Phase I, determine an optimal destination-ISP mapping. Constraint: None of the paths to the destinations in D is congested. Difficulty: Available bandwidth of the upstream network paths is generally unknown.  We cannot know a priori whether a given mapping will be congestion-free or not  Iterative routing approach

13. S routes its egress traffic based on a certain mapping for some time while measuring the loss rate in the corresponding paths. Iterative Routing Approach We allow a certain cost increase while trying to keep the amount of rerouted and dropped traffic as low as possible. If any of these paths is congested the traffic is rerouted based on a different mapping. A two-step algorithm.

14. 1. Initial mapping  Assume that bottlenecks of all paths locate at the K access links.  Calculate the minimum-cost mapping  Algorithm-1 2. Stochastic search  Find a congestion-free mapping in the vicinity of the minimum-cost mapping  Simulated annealing A Two-step Algorithm

15. Algorithm-2: Simulated Annealing Starts with an initial mapping G and an initial temperature T. Route traffic as in mapping G. ccurr = cost(G) repeat if ccurr = 0 then return G {congestion-free solution} else Generate new mapping Gnew Route traffic as in mapping Gnew cnew = cost(Gnew) if cnew ≤ ccurr then Accept Gnew (i.e., G= Gnew, curr=cnew) else Accept Gnew with probability e −(cnew−ccurr)/T end if T = ρT {cooling rate} end if until T ≈ 0 Two additional termination Conditions: If monetary cost is too large. If the congestion cost has not decreased significantly over a number of iterations. Reroute a single congested flow at a time Allocate the “max-loss” destination to the ISP that will result in the minimum cost increase

16. Summary Phase I - ISP Selection  Select K ISPs.  To minimize: Monetary cost - Algorithm 1 (FFD-like heuristic) Cost associated with AS-level path length - Algorithm 1 Cost associated with Path diversity Phase II - Egress Path Selection  Determine the destination-ISP mapping.  Two-step Algorithm: Initial mapping - Algorithm 1 Stochastic search - Algorithm 2 (Simulated Annealing)