1 Optimal Oblivious Routing in Hole-Free Networks Costas Busch Louisiana State University Malik Magdon-Ismail Rensselaer Polytechnic Institute.

1

Optimal Oblivious Routing in Hole-Free Networks

Costas BuschLouisiana State University

Malik Magdon-IsmailRensselaer Polytechnic Institute

2

1u

1v

2u2v

3u

3v

Routing: choose paths from sources to destinations

3

Edge congestion

edgeC

maximum number of paths that use any edge

Node congestion

nodeC

maximum number of paths that use any node

4

Length of chosen pathLength of shortest path

uv

Stretch=

5.18

12stretch

shortest path

chosen path

5

Oblivious RoutingEach packet path choice is independent of other packet path choices

6

1q

2q

3q

Path choices:

4q

4q

5q

kqq ,,1

Probability of choosing a path: ]Pr[ iq

1]Pr[1

k

iiq

7

Benefits of oblivious routing:

•Appropriate for dynamic packet arrivals

•Distributed

•Needs no global coordination

8

Hole-free network

9

Our contribution in this work:Oblivious routing in hole-free networks

Constant stretch

Small congestion

)log( * nCOC nodenode )1(stretch O

10

Holes

11

Related Work

Valiant [SICOMP’82]:First oblivious routing algorithmsfor permutations on butterfly and hypercube

butterfly butterfly (reversed)

12

d-dimensional Grid:

nCdOC edgeedge log*

d

nCC edgeedge

log*Lower bound for oblivious routing:

Maggs, Meyer auf der Heide, Voecking, Westermann [FOCS’97]:

13

Azar et al. [STOC03]Harrelson et al. [SPAA03]Bienkowski et al. [SPAA03]

Arbitrary Graphs (existential result): nCOC edgeedge

3* log

Constructive Results:

Racke [FOCS’02]:

nCOC edgeedge log* Racke [STOC’08]:

14

Hierarchical clusteringGeneral Approach:

15

Hierarchical clusteringGeneral Approach:

16

At the lowest level every node is a cluster

17

source destination

18

Pick random node

19

Pick random node

20

Pick random node

21

Pick random node

22

Pick random node

23

Pick random node

24

Pick random node

25

26

Adjacent nodes may follow long paths

Big stretchProblem:

27

An Impossibility Result

Stretch and congestion cannot be minimized simultaneously in arbitrary graphs

28

)( nEach path has length

n paths

Length 1

Source of packetsn

Destinationof all packets

Example graph:

nodesn

29

n packets in one path

Stretch =

Edge congestion =

1

n

30

1 packet per path

n

1

Stretch =

Edge congestion =

31

nCdOC edgeedge log*

)(stretch 2dO

Result for Grids:

Busch, Magdon-Ismail, Xi [TC’08]

For d=2, a similar result given by C. Scheideler

32

Special graphs embedded in the 2-dimensional plane:

Constant stretch

Small congestion

)log( * nCOC nodenode

)log( * nCOC edgeedge

degree

Busch, Magdon-Ismail, Xi [SPAA 2005]:

)1(stretch O

33

Embeddings in wide, closed-curved areas

34

Graph models appropriate for various wireless network topologies

Transmission radius

35

Basic Idea

source destination

36

Pick a random intermediate node

37

Construct path through intermediate node

38

However, algorithm does not extend to arbitrary closed shapes

39

Our contribution in this work:Oblivious routing in hole-free networks

40

Approach: route within square areas

)1(stretch O )log( * nCOC nodenode

41

nn grid

42

simple area in grid (hole-free area)

43

Hole-free network

44

Canonical square decomposition

45

Canonical square decomposition

46

Canonical square decomposition

47

Canonical square decomposition

48

49

50

u

v

Shortest path

51

u

v

Canonical square sequence

52

u

v

A random path in canonical squares

53

u

v

Path has constant stretch

54

Random 2-bend pathsor 1-bend paths in square sequence