L11: Algebraic Path Problems with applications to Internet Routing Lecture 9

Transcription

1 L11: Algebraic Path Problems with applications to Internet Routing Lecture 9 Timothy G. Griffin timothy.griffin@cl.cam.ac.uk Computer Laboratory University of Cambridge, UK Michaelmas Term, 2017 tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routingc 2017 Lecture 91 / 34 Widest shortest-paths Metric of the form pd, bq, where d is distance pmin, `q and b is capacity pmax, minq. Metrics are compared lexicographically, with distance considered first. Such things are found in the vast literature on Quality-of-Service (QoS) metrics for Internet routing. wsp sp ˆ bw tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routingc 2017 Lecture 92 / 34

2 Widest shortest-paths 4 (1,100) (1,100) 0 (1,10) 1 (2,90) 2 (1,5) (1,100) 3 tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routingc 2017 Lecture 93 / 34 Weights are globally optimal (we have a semiring) Widest shortest-path weights computed by Dijkstra and Bellman-Ford R » 0 p0, Jq p1, 10q p3, 10q p2, 5q p2, 10q 1 p1, 10q p0, Jq p2, 100q p1, 5q p1, 100q 2 p3, 10q p2, 100q p0, Jq p1, 100q p1, 100q 3 p2, 5q p1, 5q p1, 100q p0, Jq p2, 100q 4 p2, 10q p1, 100q p1, 100q p2, 100q p0, Jq fi ffi ffi ffi ffi fl tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routingc 2017 Lecture 94 / 34

3 But what about the paths themselves? Four optimal paths of weight p3, 10q. P optimal p0, 2q tp0, 1, 2q, p0, 1, 4, 2qu P optimal p2, 0q tp2, 1, 0q, p2, 4, 1, 0qu There are standard ways to extend Bellman-Ford and Dijkstra to compute paths (or the associated next hops). Do these extended algorithms find all optimal paths? tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routingc 2017 Lecture 95 / 34 Surprise! Four optimal paths of weight p3, 10q P optimal p0, 2q tp0, 1, 2q, p0, 1, 4, 2qu P optimal p2, 0q tp2, 1, 0q, p2, 4, 1, 0qu Paths computed by (extended) Dijkstra P Dijkstra p0, 2q tp0, 1, 2q, p0, 1, 4, 2qu P Dijkstra p2, 0q tp2, 4, 1, 0qu Notice that 0 s paths cannot both be implemented with next-hop forwarding since P Dijkstra p1, 2q tp1, 4, 2qu. Paths computed by (extended) distributed Bellman-Ford P Bellman p0, 2q tp0, 1, 4, 2qu P Bellman p2, 0q tp2, 1, 0q, p2, 4, 1, 0qu tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routingc 2017 Lecture 96 / 34

4 Optimal paths from 0 to 2. Computed by Dijkstra but not by Bellman-Ford 4 (1,100) (1,100) 0 (1,10) 1 (2,90) 2 (1,5) (1,100) tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routingc 2017 Lecture 97 / 34 3 Optimal paths from 2 to 1. Computed by Bellman-Ford but not by Dijkstra 4 (1,100) (1,100) 0 (1,10) 1 (2,90) 2 (1,5) (1,100) tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routingc 2017 Lecture 98 / 34 3

5 Observations For distributed Bellman-Ford For Dijkstra s algorithm next-hop-pathspaq computed-pathspaq Ď optimal-pathspaq next-hop-pathspaq Ď computed-pathspaq Ď optimal-pathspaq tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routingc 2017 Lecture 99 / 34 How can we understand this (algebaically)? The Algorithm to Algebra (A2A) method original metric ` complex algorithm Ñ modified metric ` matrix equations (generic algorithm) We can capture path computation with this algebra sp ˆ bw ˆ pathspeq But this algebra is not distributive! LCpsp ˆ bwq LKppathspEqq tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 910 / 34

6 Towards a non-classical theory of algebraic path finding We need theory that can accept algebras that violate distributivity. Global optimality A pi, jq à wppq, ppppi, jq Left local optimality (distributed Bellman-Ford) L pa b Lq I. Right local optimality (Dijkstra s Algorithm) R pr b Aq I. Embrace the fact that all three notions can be distinct. tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 911 / 34 Left-Local Optimality Say that L is a left locally-optimal solution when L pa b Lq I. That is, for i j we have Lpi, jq à qpv Api, qq b Lpq, jq Lpi, jq is the best possible value given the values Lpq, jq, for all out-neighbors q of source i. Rows Lpi, _q represents out-trees from i (think Bellman-Ford). Columns Lp_, iq represents in-trees to i. Works well with hop-by-hop forwarding from i. tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 912 / 34

7 Right-Local Optimality Say that R is a right locally-optimal solution when R pr b Aq I. That is, for i j we have Rpi, jq à qpv Rpi, qq b Apq, jq Rpi, jq is the best possible value given the values Rpq, jq, for all in-neighbors q of destination j. Rows Lpi, _q represents out-trees from i (think Dijkstra). Columns Lp_, iq represents in-trees to i. tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 913 / 34 With and Without Distributivity With distributivity For (bounded) semirings, the three optimality problems are essentially the same locally optimal solutions are globally optimal solutions. Without distributivity A L R It may be that A, L, and R exists but are all distinct. Back and Forth L pa b Lq I ðñ L T pl T b T A T q I where b T is matrix multiplication defined with a b T b b b a tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 914 / 34

8 Dijkstra s Algorithm Classical Dijkstra Given adjacency matrix A over a selective semiring and source vertex i P V, Dijkstra s algorithm will compute A pi, _q such that A pi, jq à w A ppq. Non-Classical Dijkstra ppppi,jq If we drop assumptions of distributivity, then given adjacency matrix A and source vertex i P V, Dijkstra s algorithm will compute Rpi, _q such P V : Rpi, jq Ipi, jq à qpv Rpi, qq b Apq, jq. Routing in Equilibrium, João Luís Sobrinho and Timothy G. Griffin, MTNS tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 915 / 34 Dijkstra s algorithm Input : adjacency matrix A and source vertex i P V, Output : the i-th row of R, Rpi, _q. begin S Ð tiu Rpi, iq Ð 1 for each q P V tiu : Rpi, qq Ð Api, qq while S V begin find q P V S such that Rpi, qq is ď L -minimal S Ð S Y tqu for each j P V S Rpi, jq Ð Rpi, jq prpi, qq b Apq, jqq end end tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 916 / 34

9 Classical proofs of Dijkstra s algorithm (for global optimality) assume Semiring Axioms ASp q : a pb cq pa bq c CMp q : a b b a IDp q : 0 a a ASpbq : a b pb b cq pa b bq b c IDLpbq : 1 b a a IDRpbq : a b 1 a ANLpbq : 0 b a 0 ANRpbq : a b 0 0 LD : a b pb cq pa b bq pa b cq RD : pa bq b c pa b cq pb b cq tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 917 / 34 Classical proofs of Dijkstra s algorithm assume Additional axioms SLp q : a b P ta, bu ANp q : 1 a 1 Note that we can derive right absorption, RA : a pa b bq a and this gives (right) b : a ď a b b. a pa b bq pa b 1q pa b bq a b p1 bq a b 1 a tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 918 / 34

10 What will we assume? Very little! Semiring /////////// Axioms ASp q : a pb cq pa bq c CMp q : a b b a IDp q : 0 a a ASpbq //////// : a////////////// b pb b cq // pa ////////////// b bq b c IDLpbq : 1 b a a IDRpbq ///////// : a////// b 1 // a/ ANLpbq ////////// : 0////// b a // 0/ ANRpbq ////////// : a////// b 0 // 0/ LD /// : a////////////// b pb cq // pa ///////////////////// b bq pa b cq RD //// : pa ////////////// bq b c // pa ///////////////////// b cq pb b cq tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 919 / 34 What will we assume? Additional axioms SLp q : a b P ta, bu ANLp q : 1 a 1 RA : a pa b bq a Note that we can no longer derive RA, so we must assume it. Again, RA says that a ď a b b. We don t use SL explicitly in the proofs, but it is implicit in the algorithm s definition of q k. We do not use ASp q and CMp q explicitly, but these assumptions are implicit in the use of the big- notation. tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 920 / 34

11 Under these weaker assumptions... Theorem (Sobrinho/Griffin) Given adjacency matrix A and source vertex i P V, Dijkstra s algorithm will compute Rpi, _q such P V : Rpi, jq Ipi, jq à qpv Rpi, qq b Apq, jq. That is, it computes one row of the solution for the right equation R RA I. tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 921 / 34 Dijkstra s algorithm, annotated version Subscripts make proofs by induction easier... begin S 1 Ð tiu R 1 pi, iq Ð 1 for each q P V S 1 : R 1 pi, qq Ð Api, qq for each k 2, 3,..., V begin find q k P V S k 1 such that R k 1 pi, q k q is ď L -minimal S k Ð S k 1 Y tq k u for each j P V S k R k pi, jq Ð R k 1 pi, jq pr k 1 pi, q k q b Apq k, jqq end end tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 922 / 34

12 Main Claim, : 1 ď k ď V P S k : R k pi, jq Ipi, jq à qps k R k pi, qq bapq, jq We will use Observation 1 (no backtracking) : 1 ď k ă V P S k`1 : R k`1 pi, jq R k pi, jq Observation 2 (Dijkstra is greedy : 1 ď k ď V P S k P V S k : R k pi, qq ď R k pi, wq Observation 3 (Accurate : 1 ď k ď V P V S k : R k pi, wq à R k pi, qqbapq, wq qps k tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 923 / 34 Observation : 1 ď k ă V P S k`1 : R k`1 pi, jq R k pi, jq Proof: This is easy to see by inspection of the algorithm. Once a node is put into S its weight never changes again. tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 924 / 34

13 The algorithm is greedy Observation : 1 ď k ď V P S k P V S k : R k pi, qq ď R k pi, wq By induction. Base : Since S 1 tiu and R 1 pi, iq 1, we need to show that 1 ď Api, wq 1 1 Api, wq. This follows from ANLp q. Induction: P S k P V S k : R k pi, qq ď R k pi, wq and P S k`1 P V S k`1 : R k`1 pi, qq ď R k`1 pi, wq. Since S k`1 S k Y tq k`1 u, this means showing p1q P S k P V S k`1 : R k`1 pi, qq ď R k`1 pi, P V S k`1 : R k`1 pi, q k`1 q ď R k`1 pi, wq tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 925 / 34 By Observation 1, showing p1q is the same P S k P V S k`1 : R k pi, qq ď R k`1 pi, wq which expands to (by definition of R k`1 pi, P S k P V S k`1 : R k pi, qq ď R k pi, wq pr k pi, q k`1 qbapq k`1, wqq But R k pi, qq ď R k pi, wq by the induction hypothesis, and R k pi, qq ď pr k pi, q k`1 q b Apq k`1, wqq by the induction hypothesis and RA. Since a ď L b ^ a ď L c ùñ a ď L pb cq, we are done. tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 926 / 34

14 By Observation 1, showing p2q is the same as showing which expands P V S k`1 : R k pi, q k`1 q ď R k`1 pi, P V S k`1 : R k pi, q k`1 q ď R k pi, wq pr k pi, q k`1 q b Apq k`1, wqq But R k pi, q k`1 q ď R k pi, wq since q k`1 was chosen to be minimal, and R k pi, q k`1 q ď pr k pi, q k`1 q b Apq k`1, wqq by RA. Since a ď L b ^ a ď L c ùñ a ď L pb cq, we are done. tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 927 / 34 Observation 3 Observation : 1 ď k ď V P V S k : R k pi, wq à qps k R k pi, qqbapq, wq Proof: By induction: Base : easy, since à qps 1 R 1 pi, qq b Apq, wq 1 b Api, wq Api, wq R 1 pi, wq Induction step. P V S k : R k pi, wq à qps k R k pi, qq b Apq, wq and P V S k`1 : R k`1 pi, wq à R k`1 pi, qq b Apq, wq qps k`1 tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 928 / 34

15 By Observation 1, and a bit of rewriting, this means we must P V S k`1 : R k`1 pi, wq R k pi, q k`1 qbapq k`1, wq à qps k R k pi, qqbapq Using the induction hypothesis, this P V S k`1 : R k`1 pi, wq R k pi, q k`1 q b Apq k`1, wq R k pi, wq But this is exactly how R k`1 pi, wq is computed in the algorithm. tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 929 / 34 Proof of Main Claim Main : 1 ď k ď V P S k : R k pi, jq Ipi, jq à qps k R k pi, qq bapq, jq Proof : By induction on k. Base case: S 1 tiu and the claim is easy. Induction: Assume P S k : R k pi, jq Ipi, jq à qps k R k pi, qq b Apq, jq We must show P S k`1 : R k`1 pi, jq Ipi, jq à R k`1 pi, qq b Apq, jq qps k`1 tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 930 / 34

16 Since S k`1 S k Y tq k`1 u, this means we must show p1q P S k : R k`1 pi, jq Ipi, jq À qps k`1 R k`1 pi, qq b Apq, jq R k`1 pi, q k`1 q Ipi, q k`1 q À qps k`1 R k`1 pi, qq b Apq, q k`1 q By use Observation 1, showing p1q is the same as P S k : R k pi, jq Ipi, jq à R k pi, qq b Apq, jq, qps k`1 which is equivalent P S k : R k pi, jq Ipi, jq pr k pi, q k`1 qbapq k`1, jqq à R k pi, qqbapq, jq qps k By the induction hypothesis, this is equivalent P S k : R k pi, jq R k pi, jq pr k pi, q k`1 q b Apq k`1, jqq, tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 931 / 34 Put another P S k : R k pi, jq ď R k pi, q k`1 q b Apq k`1, jq By observation 2 we know R k pi, jq ď R k pi, q k`1 q, and so R k pi, jq ď R k pi, q k`1 q ď R k pi, q k`1 q b Apq k`1, jq by RA. tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 932 / 34

17 To show p2q, we use Observation 1 and Ipi, q k`1 q 0 to obtain R k pi, q k`1 q à R k pi, qq b Apq, q k`1 q qps k`1 which, since Apq k`1, q k`1 q 0, is the same as R k pi, q k`1 q à R k pi, qq b Apq, q k`1 q qps k This then follows directly from Observation 3. tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 933 / 34 Finding Left Local Solutions? L pa b Lq I ðñ L T pl T b T A T q I where R T pa T b T R T q I ðñ R pr b Aq I a b T b b b a Replace RA with LA, LA b : a ď b b a tgg22 (cl.cam.ac.uk) L11: Algebraic Path Problems with applications to Internet T.G.Griffin Routing c 2017 Lecture 934 / 34