1 Multiconstrained QoS Routing: Simple Approximations to Hard Problems Guoliang (Larry) Xue Arizona State University Research Supported by ARO and NSF Collaborators: W. Zhang, J. Tang, A. Sen, and K. Thulasiraman
2 Outline/Progress of the Talk Problem Definitions Related Works Simple K-Approximation Algorithms Faster Approximation Schemes Conclusions
3 Multi-Constrained QoS Routing Given a network where each link e has a cost c(e) and a delay d(e), we are interested in finding a source- destination path whose cost is within a given cost tolerance C and whose delay is within a given delay tolerance D. This problem is NP-hard. There are many heuristic algorithms which have no performance guarantees, and sophisticated approximation schemes which are too complicated for protocol implementation. We have designed the fastest approximation schemes, as well as very simple hop-by-hop routing algorithms that have good performance guarantees.
4 Multi-Constrained QoS Routing We study the general problem where there are K QoS parameters, for any constant K≥2. We are given an undirected graph G(V, E) where each edge e E is associated with K nonnegative weights 1 (e), 2 (e), …, K (e). We are also given a source s and destination t, and K positive constants W 1, …, W K. The multi-constrained QoS routing problem asks for an s—t path p such that k (p) ≤ W k, for k=1, 2, …, K. For simplicity, we assume K=2 for the most part of this talk. In this case, we will talk about cost and delay.
5 Illustration of the Problem (C=W 1, D=W 2 ) s x y z (2, 5) (12, 5) (12, 20) (2, 2) (10, 0) K = 2 W 1 = 16, W 2 = 8 The shortest path with regard to the 1 st edge weight is (s, z) (14, 1) The shortest path with regard to the 2 nd edge weight is (s, y, z) Neither of them is a feasible solution ! Path (s, x, y, z) is a feasible path.
6 Outline/Progress of the Talk Problem Definitions Related Works Simple K-Approximation Algorithms Faster Approximation Schemes Conclusions
7 Related Works J.M. Jaffe, Algorithms for finding paths with multiple constraints, Networks, S. Chen and K. Nahrstedt, On finding multi-constrained paths, IEEE International Conference on Communications, X. Yuan, Heuristic algorithms for multiconstrained quality of service routing, IEEE/ACM Transactions on Networking, R. Hassin, Approximation schemes for the restricted shortest path problems, Mathematics of Operations Research, D.H. Lorenz and D. Raz, A simple efficient approximation scheme for the restricted shortest path problem, Operations Research Letters, G. Xue, A. Sen, W. Zhang, J. Tang, K. Thulasiraman; Finding a path subject to many additive QoS constraints; IEEE/ACM Transactions on Networking, G. Xue, W. Zhang, J. Tang, K. Thulasiraman; Polynomial time approximation algorithms for multi-constrained QoS routing; IEEE/ACM Transactions on Networking, 2008.
8 Related Works G. Xue; Minimum cost QoS multicast and unicast routing in communication networks; IPCCC’2000/IEEE Transactions on Communications, A. Junttner et al., Lagrange relaxation based method for the QoS routing problems, IEEE INFOCOM, A. Goel et al., Efficient computation of delay-sensitive routes from one source to all destinations, IEEE INFOCOM, T. Korkmaz and M. Krunz, A randomized algorithm for finding a path subject to multiple QoS requirements, Computer Networks, P. Van Mieghem et al., Concepts of exact QoS routing algorithms, IEEE/ACM Transactions on Networking, F.A. Kuipers et al., A comparison of exact and eps-approximation algorithms for constrained routing, IFIP NETWORKING, A. Orda and A. Sprintson., Efficient algorithms for computing disjoint QoS paths, IEEE INFOCOM, 2004.
9 Outline/Progress of the Talk Problem Definitions Related Works Simple K-Approximation Algorithms Faster Approximation Schemes Conclusions
10 A Simple Idea The decision problem is to find a path p such that c(p)≤C and d(p)≤D. The optimization problem is to find a path p such that max {c(p)/C, d(p)/D} is minimized. Define l(p) = max {c(p)/C, d(p)/D} as a new path length. The original problem has a feasible solution if and only if there is a path p such that l(p)≤1. The optimization problem is NP-hard as well. The Idea: For each link e, define a new link weight w(e) = max{c(e)/C, d(e)/D}. The shortest path with respect to w(e) can be computed easily, and is within a factor of 2 from the optimal solution.
11 Illustration of the Concepts (C=W 1, D=W 2 ) s x y z (2, 5) (12, 5) (12, 20) (2, 2) (10, 0) K = 2 W 1 = 16, W 2 = 8 The shortest path with regard to the 1 st edge weight is (s, z), l(p)=20/8. (14, 1) The shortest path with regard to the 2 nd edge weight is (s, y, z), l(p)=11/8. Neither of them is a feasible/optimal solution ! The optimal path is (s, x, y, z), l(p)=7/8
12 A Simple 2-Approximation Algorithm s x y z (2, 5) (12, 5) (12, 20) (2, 2) (10, 0) K = 2 W 1 = 16, W 2 = 8 (14, 1) The shortest path with regard to the new edge weight is (s, y, z) whose path length is 11/8. This path has a length that is guaranteed to be within a factor of 2 from the optimal value. In this case, we have 11/8 ≤ 2×7/8. (2/16, 5/8)5/8 14/16 20/8 12/16 2/8 10/16
13 A Better Greedy 2-Approximation Algorithm s x y z (2, 5) (12, 5) (12, 20) (2, 2) (10, 0) K = 2 W 1 = 16, W 2 = 8 (14, 1) The path found by Greedy is (s, x, z) with path length 1 [0,0] [2/16, 5/8] A path from s to x with path weights [2/16, 5/8] is stored at node x. The path length is 5/8 [12/16, 20/8][12/16, 5/8] The path at node x is chosen because it has the minimum path length [4/16, 7/8][16/16, 6/8] The path at node y is chosen because it has the minimum path length among the unmarked nodes [22/16, 5/8] The optimal solution is (s, x, y, z) with path length 7/8
14 Proof of Correctness K-Approx: The central idea used in the proof of K-Approx relies on the following simple fact. Let x be a point in the K-dimensional Euclidean plane. Then ||x|| ≤||x|| 1 ≤K ||x|| Greedy: Greedy never violates the upper-bound on path length used in the proof of K-Approx.
15 Numerical Results Algorithms compared Greedy Previously best known K-approximation algorithm FPTAS for the OMCP problem K = 3, W = W 1 = W 2 = W 3 Networks well-known Internet topologies ArpaNet (20 nodes and 32 edges) and ItalianNET (33 nodes, 67 edges) randomly generated topologies BRITE [BRITE] Waxman model [WaxJSAC88], and have the default parameters set by BRITE the edge weights were uniformly generated in a given range (we used the range [1,10]). Three scenarios Infeasible W = 5 Tight W = 10 Loose W = 20 [BRITE] BRITE; [WaxJSAC88] B.M. Waxman; Routing of multipoint connections; IEEE Journal on Selected Areas in Communications; Vol. (1988). (ε = 0.1)
16 On ArpaNet Topology The number of better paths: path p1 is better than path p2 if l(p1) < l(p2) For any path p, its relative error is calculated as (l(p) - l(p OMCP ))/ l(p OMCP ), where p OMCP is the path found by OMCP for the source-destination pair.
17 On Large Random Network Topologies Path quality, eps = 0.1, 100 nodes, 390 links. Scalability of the algorithms, eps= x314, 210x474, 140x560, 160x634.
18 Outline/Progress of the Talk Problem Definitions Related Works Simple K-Approximation Algorithm Faster Approximation Schemes Conclusions
19 Approximation Scheme for SMCP We are given an undirected graph G(V, E) where each edge e E is associated with K nonnegative weights 1 (e), 2 (e), …, K (e). We are also given a source s and destination t, and K positive constants W 1, …, W K. We want to find an s-t path p s.t max{ k (p)/ W k, 1≤k≤K} is minimized. In a paper published in TON’2007, we designed an algorithm that can find a (1+ )-approximation in O(m(n/ ) K-1 ) time. This is the first FPTAS for the general SMCP problem (K 2).
20 Approximation Scheme for SMCP The idea follows that used by other researchers in this field. Find an initial pair of lower and upper bounds not too far away from each other. Use scaling/rounding/approximate testing to refine the bounds to within a constant factor Compute an (1+ )-approximation. The difference is that we got a pair of lower and upper bounds with a constant (K) factor in a single step, using our K-Approx. This leads to faster running time. O(mn(loglogn+1/ )) O(mn/ ) for K=2. However, the problem is slightly different from the DCLC problem. None of the constraints is enforced. Motivation for the second TON paper.
21 Faster Approximation Schemes for OMCP All previous approximation schemes for OMCP are based on Initial bounds Scaling and rounding, and approximate testing Final solution Hassin rounds to floor. Lorenz and Raz round to floor plus one, and showed its advantage over that of Hassin. A simple combination of the two techniques leads to an approximation scheme that is better than both.
22 Basic Definitions Again
23 Decision Version of the Basic Problem
24 A Restricted Decision Version (for comparison)
25 Optimization Version of the Problem
26 The DCLC Problem (used as a subproblem)
27 MCPN: non-negative integers (2 weights)
28 MCPP: positive integers (K weights)
29 Solving MCPP in O(mC K-1 ) Time
30 Solving MCPP in O(mC K-1 ) Time
31 Illustration of the idea (acyclic graph)
32 Solving MCPN in O((m+nlogn)C) Time
33 Solving MCPN in O((m+nlogn)C) Time
34 Solving MCPN in O((m+nlogn)C) Time
35 Non-negative rounding and approximate testing
36 Non-negative rounding and approximate testing
37 Positive rounding and approximate testing
38 Positive rounding and approximate testing
39 The Power of Approximate Testing Assume UB 2(1+ )LB. Set C =sqrt(LB UB/(1+ )). Run TEST(C, ). If TEST(C, ) YES, then DCLC <C(1+ ). Decrease UB to C(1+ ). If TEST(C, ) NO, then DCLC >C. Increase LB to C. In both cases, UB/LB is reduced to sqrt((1+ )(UB/LB)). We will have UB≤2(1+ )LB, after loglog(initial UB/LB ratio) iterations.
40 Faster Approximation Scheme for DCLC Use the technique of Lorenz and Raz to compute LB and UB of DCLC so that LB ≤ DCLC ≤UB≤n LB. This takes O((m+nlogn)logn)) time: logn shortest path computations. Set N to (logn) 2, and apply TEST N to refine LB and UB so that LB≤ DCLC ≤UB≤2(1+ (logn) 2 ) LB. This takes O(mn) time: loglog(n) TEST N, each requires O((m+nlogn)n/ N ) time. Set P to 1, and apply TEST P to refine LB and UB so that LB≤ DCLC ≤UB≤2(1+ 1 ) LB. This takes O(mnlogloglogn) time: loglog(logn) TEST P, each requires O(mn/ P ) time. Solve MCPP with scaling factor =(n-1)/(LB ). This takes O(mn/ ) time. O(mn(loglogn+1/ )) O(mn(logloglogn+1/ ))
41 Faster Approximation Scheme for DCLC
42 Faster Approximation Scheme for DCLC
43 Faster Approximation Scheme for DCLC O(mn(loglogn + 1/ )) time [Lorenz and Raz ORL’2001] O(mn(logloglogn + 1/ )) time. Conjecture: O(mn/ ) time both necessary and sufficient.
44 Dimension Reduction: OMCP DCLC
45 (1+ )(K-1)-Approx to OMCP, via DCLC
46 Faster Approximation Scheme for OMCP
47 Faster Approximation Scheme for OMCP This is essentially O(m(n/ ) K-1 ) time.
48 Faster Heuristic/Scheme for DMCP
49 Faster Heuristic/Scheme for DMCP O(mn(n/ ) K-1 ) time [Yuan TON’2002] O(m(H/ ) K-1 ) time.
50 Running Time
51 Running Time
52 Path Weight Ratios
53 Outline/Progress of the Talk Problem Definitions Related Works Simple K-Approximation Algorithm Faster Approximation Schemes Conclusions
54 Conclusions We know how to compute a shortest path. OSPF has been proposed by IETF as an RFC. We don’t know how to handle two or more QoS constraints with guaranteed performance. This is the first approach which is both simple and provably good. It is as simple as computing a shortest path. The computed path is within a factor of K from optimal. From the theoretical point of view, we have designed faster FPTAS for several versions of the problem.
55 THANK YOU!