ALLOCATION ALGORITHMS FOR NETWORKS WITH SCARCE RESOURCES Kanthi - - PowerPoint PPT Presentation

allocation algorithms for networks with scarce resources
SMART_READER_LITE
LIVE PREVIEW

ALLOCATION ALGORITHMS FOR NETWORKS WITH SCARCE RESOURCES Kanthi - - PowerPoint PPT Presentation

ALLOCATION ALGORITHMS FOR NETWORKS WITH SCARCE RESOURCES Kanthi Kiran Sarpatwar Dissertation Defense: Feb 13, 2015 Committee Samir Khuller (Advisor) MohammadTaghi Hajiaghayi Peter Keleher Mark A. Shayman Aravind Srinivasan Kanthi Kiran


slide-1
SLIDE 1

ALLOCATION ALGORITHMS FOR NETWORKS WITH SCARCE RESOURCES

Kanthi Kiran Sarpatwar Dissertation Defense: Feb 13, 2015 Committee Samir Khuller (Advisor) MohammadTaghi Hajiaghayi Peter Keleher Mark A. Shayman Aravind Srinivasan

Kanthi Kiran Sarpatwar 1 / 82

slide-2
SLIDE 2

Motivation

Theme How do we effectively handle bottleneck resources in networks? Data Storage Resource Replication Problems. Computational Resources Container Selection Problem. Energy Connected Dominating Set Problem.

Kanthi Kiran Sarpatwar 2 / 82

slide-3
SLIDE 3

Motivation

Theme How do we effectively handle bottleneck resources in networks? Data Video content providers, such as Netflix, must replicate data to minimize client latency. Content Distribution Network

Kanthi Kiran Sarpatwar 3 / 82

slide-4
SLIDE 4

Motivation

Theme How do we effectively handle bottleneck resources in networks? Computational Resources Cross platform schedulers must fairly allocate cluster resources to various platforms. Cluster Network

Kanthi Kiran Sarpatwar 4 / 82

slide-5
SLIDE 5

Motivation

Theme How do we effectively handle bottleneck resources in networks? Energy Wireless ad hoc networks have nodes with limited battery life. Key issues here are routing, target monitoring and interference. Wireless Adhoc Network

Kanthi Kiran Sarpatwar 5 / 82

slide-6
SLIDE 6

Papers in this Talk

Improved Approximation Algorithms for Resource Replication Problems. APPROX 2012 Khuller, Saha, S. Container Selection with Applications to Cloud Computing. Nagarajan, S., Schieber, Shachnai, Wolf Analyzing the Optimal Neighborhood: Approximation Algorithms for Partial and Budgeted Connected Dominating Set. SODA 2014 Khuller, Purohit, S.

Kanthi Kiran Sarpatwar 6 / 82

slide-7
SLIDE 7

Other Papers

The X-Flex Cross-Platform Scheduler: Who’s the Fairest of Them All? Middleware 2014 Wolf, Nabi, Nagarajan, Saccone, Wagle, Hildrum, Pring, S. Approximation Algorithms for Covering Problems in Energy Constrained Wireless Networks. Khuller, Purohit, S. Improved Approximation Algorithms for Steiner tree and Cheapest Tour Oracles. Bhatia, Gupta, S.

Kanthi Kiran Sarpatwar 7 / 82

slide-8
SLIDE 8

Part I Computational Resources: Container Selection Problem

Kanthi Kiran Sarpatwar 8 / 82

slide-9
SLIDE 9

Cross platform scheduler

Kanthi Kiran Sarpatwar 9 / 82

slide-10
SLIDE 10

Examples of Cross Platform Schedulers

Dominant Resource Fairness (DRF) - NSDI 2011 Ghodsi, Zaharia, Hindman, Konwinski, Shenker, Stoica X-Flex Cross Platform Scheduler- Middleware 2014 Wolf, Nabi, Nagarajan, Saccone, Wagle, Hildrum, Pring, S.

Kanthi Kiran Sarpatwar 10 / 82

slide-11
SLIDE 11

Problem Definition

Example Input Given N jobs requiring two resources say CPU and memory. Number of dimensions d = 2. Goal Find a few representative points for all the input points.

Kanthi Kiran Sarpatwar 11 / 82

slide-12
SLIDE 12

Problem Definition

Example Container point A point (x,y) dominates another point

(x′,y′) if

x′ ≤ x and y′ ≤ y We call such a point (x,y) a container point.

Kanthi Kiran Sarpatwar 12 / 82

slide-13
SLIDE 13

Problem Definition

Example Container point A point (x,y) dominates another point

(x′,y′) if

x′ ≤ x and y′ ≤ y We call such a point (x,y) a container point.

Kanthi Kiran Sarpatwar 13 / 82

slide-14
SLIDE 14

Problem Definition

Example Cost of assignment Cost of assigning an input point to a container point (x,y) = x + y

Kanthi Kiran Sarpatwar 14 / 82

slide-15
SLIDE 15

Problem Definition

Example Objective Find k container points that minimize the total assignment cost of all input points.

Kanthi Kiran Sarpatwar 15 / 82

slide-16
SLIDE 16

Problem Definition

Example Objective Find k container points that minimize the total assignment cost of all input points.

Kanthi Kiran Sarpatwar 16 / 82

slide-17
SLIDE 17

Problem Definition

Example Total cost computation 2(2+ 4)+ 2(4+ 3)+ 2(5+ 5) = 46

Kanthi Kiran Sarpatwar 17 / 82

slide-18
SLIDE 18

Main Result for the Continuous Setting

Definition (continuous container selection) In an instance of the problem, we are given a set of input points C in Rd and a budget k. The goal is to find a subset S of k container points in Rd, such that the following cost is minimized:

p∈C

min

c∈S p≺c

c

Theorem For any fixed dimension d, there is a polynomial time approximation scheme for the container selection problem.

Kanthi Kiran Sarpatwar 18 / 82

slide-19
SLIDE 19

The Main Idea

Kanthi Kiran Sarpatwar 19 / 82

slide-20
SLIDE 20

Potential Container Points

The Sets X and Y X be the set of all input x-coordinates and Y be the set of all input y-coordinates. Observation Any container point chosen by an optimal solution must be in X × Y.

Kanthi Kiran Sarpatwar 20 / 82

slide-21
SLIDE 21

Reducing General Case to Restricted Case

Transformation Input: Set of N points and a budget k on the number of containers.

Kanthi Kiran Sarpatwar 21 / 82

slide-22
SLIDE 22

Reducing General Case to Restricted Case

Transformation Compute: The set of potential container points.

Kanthi Kiran Sarpatwar 22 / 82

slide-23
SLIDE 23

Reducing General Case to Restricted Case

Transformation Constant number of lines: For a given ε, we construct equiangular rays separated by

θ ≈ ε/2.

Kanthi Kiran Sarpatwar 23 / 82

slide-24
SLIDE 24

Reducing General Case to Restricted Case

Transformation Shifting: Shift potential container points onto these rays. Magnified View:

Kanthi Kiran Sarpatwar 24 / 82

slide-25
SLIDE 25

Reducing General Case to Restricted Case

Transformation Shifting: Using basic trigonometry, we obtain min(∆x,∆y) ≤ (x + y)2θ ≤ (x + y)ε Magnified View: Theorem There is a poly-time algorithm for the restricted container selection problem.

Kanthi Kiran Sarpatwar 25 / 82

slide-26
SLIDE 26

PTAS in Two and Higher Dimensions

Higher Dimensions The transformation can be extended to higher fixed dimensions! Restricted Problem There is a poly-time algorithm for the restricted container selection problem in any fixed dimension d. Theorem (PTAS) This implies a PTAS for the continuous container selection problem in any fixed dimension. Theorem (NP-hard) We further show that the problem is NP-hard in any fixed dimension d ≥ 3.

Kanthi Kiran Sarpatwar 26 / 82

slide-27
SLIDE 27

PTAS in Two and Higher Dimensions

Higher Dimensions The transformation can be extended to higher fixed dimensions! Restricted Problem There is a poly-time algorithm for the restricted container selection problem in any fixed dimension d. Theorem (PTAS) This implies a PTAS for the continuous container selection problem in any fixed dimension. Theorem (NP-hard) We further show that the problem is NP-hard in any fixed dimension d ≥ 3.

Kanthi Kiran Sarpatwar 26 / 82

slide-28
SLIDE 28

PTAS in Two and Higher Dimensions

Higher Dimensions The transformation can be extended to higher fixed dimensions! Restricted Problem There is a poly-time algorithm for the restricted container selection problem in any fixed dimension d. Theorem (PTAS) This implies a PTAS for the continuous container selection problem in any fixed dimension. Theorem (NP-hard) We further show that the problem is NP-hard in any fixed dimension d ≥ 3.

Kanthi Kiran Sarpatwar 26 / 82

slide-29
SLIDE 29

PTAS in Two and Higher Dimensions

Higher Dimensions The transformation can be extended to higher fixed dimensions! Restricted Problem There is a poly-time algorithm for the restricted container selection problem in any fixed dimension d. Theorem (PTAS) This implies a PTAS for the continuous container selection problem in any fixed dimension. Theorem (NP-hard) We further show that the problem is NP-hard in any fixed dimension d ≥ 3.

Kanthi Kiran Sarpatwar 26 / 82

slide-30
SLIDE 30

PTAS in Two and Higher Dimensions

Higher Dimensions The transformation can be extended to higher fixed dimensions! Restricted Problem There is a poly-time algorithm for the restricted container selection problem in any fixed dimension d. Theorem (PTAS) This implies a PTAS for the continuous container selection problem in any fixed dimension. Theorem (NP-hard) We further show that the problem is NP-hard in any fixed dimension d ≥ 3.

Kanthi Kiran Sarpatwar 26 / 82

slide-31
SLIDE 31

Discrete Variant

Discrete vs Continuous Containers must be chosen from a given set of potential container points

F .

The previous transformation fails as we are not allowed to “move” them! Much harder than continuous version! In fact, we show that: Theorem (Hardness of Approximation) For any dimension d ≥ 3, the discrete container selection problem is NP-hard to obtain any approximation.

Kanthi Kiran Sarpatwar 27 / 82

slide-32
SLIDE 32

What do we do?

Relax the problem a little We relax the restriction on the number of container points. What do we know? Bi-approximation Results Special case of non-metric k-median problem. There is a (1+ε,(1+ 1

ε )lnn) bi-approximation algorithm (Lin and Vitter,

STOC 1992). The Question Can we still use the geometric properties of our special problem to do better?

Kanthi Kiran Sarpatwar 28 / 82

slide-33
SLIDE 33

Our Results

2-Dimensions We obtain a (1+ε,3) bi-approximation algorithm. This approach does not work for higher dimensions. Higher Dimensions We obtain a (1+ε,O( d

ε logdk)) bi-approximation algorithm for

dimension d. Based on an LP relaxation.

Kanthi Kiran Sarpatwar 29 / 82

slide-34
SLIDE 34

Our Algorithm for 2-Dimensions

Cells

  • Decompose

Decompose the space in O(logn) cells. “Guess” which cells are touched by a fixed optimal solution. Call them good cells.

Kanthi Kiran Sarpatwar 30 / 82

slide-35
SLIDE 35

Our Algorithm for 2-Dimensions

Cells Representative points From the good cells choose two container points - one with maximum x-coordinate and the other with maximum y-coordinate.

Kanthi Kiran Sarpatwar 31 / 82

slide-36
SLIDE 36

Our Algorithm for 2-Dimensions

Cells

pmax pmin

Cells are approximately uniform Using a trigonometric argument, we can show that the costs of container points in any given cell are approximately the same, i.e., pmax/pmin ≤ (1+ε).

Kanthi Kiran Sarpatwar 32 / 82

slide-37
SLIDE 37

Our Algorithm for 2-Dimensions

Cells

i j r1 r2 C1 C2

Decoupling the cells Given input point i ∈ C1 and container point j ∈ C2 such that i ≺ j, then i ≺ r1

  • r i ≺ r2.

Kanthi Kiran Sarpatwar 33 / 82

slide-38
SLIDE 38

Our Algorithm for 2-Dimensions

Single Cell Problem For a given budget k1 to a cell, we try and satisfy input points only from that

  • cell. We can use a simple DP to solve this optimally.

Restricted problem The only “inter cell” allocations are to the representative container points. We use a dynamic program based scheme to solve the problem under this restrictions.

Kanthi Kiran Sarpatwar 34 / 82

slide-39
SLIDE 39

Results Summary

Continuous CSP PTAS for any d ≥ 2 NP-hard for any d ≥ 3 Discrete CSP For two dimensions, a

(1+ε,3)-bi-approximation

algorithm For any fixed dimension d, a

(1+ε,O( 1

ε logdk))

bi-approximation algorithm. NP-hard to approximate, for any d ≥ 3

Kanthi Kiran Sarpatwar 35 / 82

slide-40
SLIDE 40

Part II Data: Resource Replication Problems

Kanthi Kiran Sarpatwar 36 / 82

slide-41
SLIDE 41

Resource Replication Problems

Example Framework Clients and servers are embedded into a metric space. Clients need a subset of data

  • bjects {A,B,C,D,E}.

Servers have limited capacities to store data

  • bjects.

Kanthi Kiran Sarpatwar 37 / 82

slide-42
SLIDE 42

Resource Replication Problems

Example Framework Goal: Place data items on different servers to meet the demands of all clients. Minimize the distance a client has to travel to go to get a required data item.

Kanthi Kiran Sarpatwar 38 / 82

slide-43
SLIDE 43

Resource Replication Problems

Example Framework Goal: Place data items on different servers to meet the demands of all clients. Minimize the distance a client has to travel to go to get a required data item.

Kanthi Kiran Sarpatwar 39 / 82

slide-44
SLIDE 44

Resource Replication Problems

Example Objectives: Min Sum Minimize the aggregate distance travelled by all clients to obtain all

  • f their required data objects.

For this example, total cost = 10+ 3+(20+ 7)+ 15+(10+ 10)+(10+ 5+ 5) = 95

Kanthi Kiran Sarpatwar 40 / 82

slide-45
SLIDE 45

Resource Replication Problems

Example Objectives: Min Sum Minimize the aggregate distance travelled by all clients to obtain all

  • f their required data objects.

For this example, total cost = 10+ 3+(20+ 7)+ 15+(10+ 10)+(10+ 5+ 5) = 95

Kanthi Kiran Sarpatwar 40 / 82

slide-46
SLIDE 46

Resource Replication Problems

Example Objectives: Min Max Minimize the maximum distance travelled by all clients to obtain all

  • f their required data objects.

For this example, cost = 20

Kanthi Kiran Sarpatwar 41 / 82

slide-47
SLIDE 47

Resource Replication Problems

Example Objectives: Min Max Minimize the maximum distance travelled by all clients to obtain all

  • f their required data objects.

For this example, cost = 20

Kanthi Kiran Sarpatwar 41 / 82

slide-48
SLIDE 48

Related Work

Min Sum Objective : Baev, Rajaraman and Swamy - SIAM J. Compt. 2008 LP-based approximation algorithm. Has an approximation guarantee of 10. Min Max Objective : Ko and Rubenstein - ICNP 2003, ICNP 2004 Heuristic approach. 3-approximation algorithm for a basic version - but the algorithm does not necessarily terminate in poly time. No approximation guarantee for the general problem.

Kanthi Kiran Sarpatwar 42 / 82

slide-49
SLIDE 49

Related Work

Min Sum Objective : Baev, Rajaraman and Swamy - SIAM J. Compt. 2008 LP-based approximation algorithm. Has an approximation guarantee of 10. Min Max Objective : Ko and Rubenstein - ICNP 2003, ICNP 2004 Heuristic approach. 3-approximation algorithm for a basic version - but the algorithm does not necessarily terminate in poly time. No approximation guarantee for the general problem.

Kanthi Kiran Sarpatwar 42 / 82

slide-50
SLIDE 50

Related Work

Min Sum Objective : Baev, Rajaraman and Swamy - SIAM J. Compt. 2008 LP-based approximation algorithm. Has an approximation guarantee of 10. Min Max Objective : Ko and Rubenstein - ICNP 2003, ICNP 2004 Heuristic approach. 3-approximation algorithm for a basic version - but the algorithm does not necessarily terminate in poly time. No approximation guarantee for the general problem.

Kanthi Kiran Sarpatwar 42 / 82

slide-51
SLIDE 51

Basic Resource Replication

Definition Given a graph G = (V,E), a metric d : E → R+ ∪{0} and data types set C . Find a mapping φ : V → C to minimize the following quantity: max

v∈V,r∈C

min

u∋φ(u)=r d(u,v)

Every node needs all data items. Every node has a unit storage capacity. We give a simple 3-approximation algorithm.

Kanthi Kiran Sarpatwar 43 / 82

slide-52
SLIDE 52

Basic Resource Replication

Definition Given a graph G = (V,E), a metric d : E → R+ ∪{0} and data types set C . Find a mapping φ : V → C to minimize the following quantity: max

v∈V,r∈C

min

u∋φ(u)=r d(u,v)

Every node needs all data items. Every node has a unit storage capacity. We give a simple 3-approximation algorithm.

Kanthi Kiran Sarpatwar 43 / 82

slide-53
SLIDE 53

Basic Resource Replication

Definition Given a graph G = (V,E), a metric d : E → R+ ∪{0} and data types set C . Find a mapping φ : V → C to minimize the following quantity: max

v∈V,r∈C

min

u∋φ(u)=r d(u,v)

Every node needs all data items. Every node has a unit storage capacity. We give a simple 3-approximation algorithm.

Kanthi Kiran Sarpatwar 43 / 82

slide-54
SLIDE 54

Basic Resource Replication

Definition Given a graph G = (V,E), a metric d : E → R+ ∪{0} and data types set C . Find a mapping φ : V → C to minimize the following quantity: max

v∈V,r∈C

min

u∋φ(u)=r d(u,v)

Every node needs all data items. Every node has a unit storage capacity. We give a simple 3-approximation algorithm.

Kanthi Kiran Sarpatwar 43 / 82

slide-55
SLIDE 55

Basic Resource Replication

Definition Given a graph G = (V,E), a metric d : E → R+ ∪{0} and data types set C . Find a mapping φ : V → C to minimize the following quantity: max

v∈V,r∈C

min

u∋φ(u)=r d(u,v)

Every node needs all data items. Every node has a unit storage capacity. We give a simple 3-approximation algorithm.

Kanthi Kiran Sarpatwar 43 / 82

slide-56
SLIDE 56

Algorithm for Basic Resource Replication

Example Input Set of nodes and objects

{R,G,B}

Kanthi Kiran Sarpatwar 44 / 82

slide-57
SLIDE 57

Algorithm for Basic Resource Replication

Example Construct Threshold Graph: “Guess” optimal dist. δ. Add uv if d(u,v) ≤ δ.

Kanthi Kiran Sarpatwar 45 / 82

slide-58
SLIDE 58

Algorithm for Basic Resource Replication

Example Compute 2-hop MIS: Keep picking nodes and delete nodes within two hops.

Kanthi Kiran Sarpatwar 46 / 82

slide-59
SLIDE 59

Algorithm for Basic Resource Replication

Example Compute 2-hop MIS: Keep picking nodes and delete nodes within two hops.

Kanthi Kiran Sarpatwar 47 / 82

slide-60
SLIDE 60

Algorithm for Basic Resource Replication

Example Compute 2-hop MIS: Keep picking nodes and delete nodes within two hops.

Kanthi Kiran Sarpatwar 48 / 82

slide-61
SLIDE 61

Algorithm for Basic Resource Replication

Example Assign colors: For each vertex in MIS, we place k = 3 resources in its neighborhood in Gδ .

Kanthi Kiran Sarpatwar 49 / 82

slide-62
SLIDE 62

Algorithm for the Basic Resource Replication

Analysis Every vertex has a degree at least k − 1 in Gδ . Therefore, our coloring is valid. By definition, every vertex is within a 2-hop distance of some vertex in MIS. Hence, every vertex has all the colors within a 3-hop distance. Theorem There is a 3-approximation algorithm for the basic resource replication problem.

Kanthi Kiran Sarpatwar 50 / 82

slide-63
SLIDE 63

Subset Resource Replication

Given: a graph G = (V,E) embedded into a metric d : E → R+ ∪{0} and a set

  • f data items C

each vertex has a storage capacity of sv each vertex needs a subset of data items Cv Goal: find a mapping φ : V → 2C that assigns at most sv data items to v and minimizes the following quantity: max

v∈V

min

r∈Cv u∋φ(u)=r

d(u,v)

Kanthi Kiran Sarpatwar 51 / 82

slide-64
SLIDE 64

Subset Resource Replication

Given: a graph G = (V,E) embedded into a metric d : E → R+ ∪{0} and a set

  • f data items C

each vertex has a storage capacity of sv each vertex needs a subset of data items Cv Goal: find a mapping φ : V → 2C that assigns at most sv data items to v and minimizes the following quantity: max

v∈V

min

r∈Cv u∋φ(u)=r

d(u,v)

Kanthi Kiran Sarpatwar 51 / 82

slide-65
SLIDE 65

Subset Resource Replication

Given: a graph G = (V,E) embedded into a metric d : E → R+ ∪{0} and a set

  • f data items C

each vertex has a storage capacity of sv each vertex needs a subset of data items Cv Goal: find a mapping φ : V → 2C that assigns at most sv data items to v and minimizes the following quantity: max

v∈V

min

r∈Cv u∋φ(u)=r

d(u,v)

Kanthi Kiran Sarpatwar 51 / 82

slide-66
SLIDE 66

Subset Resource Replication

Given: a graph G = (V,E) embedded into a metric d : E → R+ ∪{0} and a set

  • f data items C

each vertex has a storage capacity of sv each vertex needs a subset of data items Cv Goal: find a mapping φ : V → 2C that assigns at most sv data items to v and minimizes the following quantity: max

v∈V

min

r∈Cv u∋φ(u)=r

d(u,v)

Kanthi Kiran Sarpatwar 51 / 82

slide-67
SLIDE 67

Subset Resource Replication

Given: a graph G = (V,E) embedded into a metric d : E → R+ ∪{0} and a set

  • f data items C

each vertex has a storage capacity of sv each vertex needs a subset of data items Cv Goal: find a mapping φ : V → 2C that assigns at most sv data items to v and minimizes the following quantity: max

v∈V

min

r∈Cv u∋φ(u)=r

d(u,v)

Kanthi Kiran Sarpatwar 51 / 82

slide-68
SLIDE 68

Algorithm for Subset Resource Replication

Example Input: Objects set {R,G,B} Subsets of data items needed by vertices are next to them Storage capacities (for this example) are unit

Kanthi Kiran Sarpatwar 52 / 82

slide-69
SLIDE 69

Algorithm for Subset Resource Replication

Example Threshold Graph: Guess optimal δ Construct threshold Gδ Mark 2-hop edges, represented by dashed lines

Kanthi Kiran Sarpatwar 53 / 82

slide-70
SLIDE 70

Algorithm for Subset Resource Replication

Decompose: For each color r, construct a subgraph on nodes needing resource r. Example

Kanthi Kiran Sarpatwar 54 / 82

slide-71
SLIDE 71

Algorithm for Subset Resource Replication

Decompose: Compute 2-hop maximal independent set in each of these subgraphs. Example

Kanthi Kiran Sarpatwar 55 / 82

slide-72
SLIDE 72

Algorithm for Subset Resource Replication

Example Side A: Union of 2-hop maximal independent sets in each subgraph Side B: Vertices of the graph.

Kanthi Kiran Sarpatwar 56 / 82

slide-73
SLIDE 73

Algorithm for Subset Resource Replication

Example Compute: A b-matching with bounds sv on the vertex v on the right side 1 on the vertices on the left.

Kanthi Kiran Sarpatwar 57 / 82

slide-74
SLIDE 74

Algorithm for Subset Resource Replication

Example Color: Use the determined matching to color the nodes Theorem There is a 3-approximation algorithm for the subset resource replication problem.

Kanthi Kiran Sarpatwar 58 / 82

slide-75
SLIDE 75

Algorithm for Subset Resource Replication

Example Color: Use the determined matching to color the nodes Theorem There is a 3-approximation algorithm for the subset resource replication problem.

Kanthi Kiran Sarpatwar 58 / 82

slide-76
SLIDE 76

Results Summary

Problem

  • Approx. Guar.

Hardness Basic Resource Replication (BRR) 3 2−ε Subset Resource Replication (SRR) 3 3−ε Robust BRR 3 2−ε Robust K-BRR 5 2−ε Robust SRR

  • NP-hard to approx.

Capacitated BRR

(4,2)-bi-approx.

  • Kanthi Kiran Sarpatwar

59 / 82

slide-77
SLIDE 77

Part III Energy: Partial and Budgeted Connected Dominating Set

Kanthi Kiran Sarpatwar 60 / 82

slide-78
SLIDE 78

Energy Issues in Wireless Adhoc Network

Wireless Adhoc Network Routing Communication backbones, i.e., virtual backbone Monitoring Target monitoring in sensor networks Interference Message propagation in radio networks

Kanthi Kiran Sarpatwar 61 / 82

slide-79
SLIDE 79

Problem Definitions

Dominating Set (DS) Connected Dominating Set (CDS) Definition (CDS) Given an undirected graph G = (V,E), find a connected subgraph with fewest number of nodes that dominates all the nodes.

Kanthi Kiran Sarpatwar 62 / 82

slide-80
SLIDE 80

Previous Work on CDS

Guha and Khuller, ESA 1996 ln∆+ 3 approximation algorithm in general graphs. Set cover hard. Dubhashi, Mei, Panconesi, Radhakrishnan, and Srinivasan, SODA 2003 Distributed O(ln∆) approximation algorithm in general graphs. Demaine and Hajiaghayi, SODA 2005 PTAS in planar graphs. Cheng, Huang, Li, Wu, and Du, Networks 2003 PTAS in unit disk graphs.

Kanthi Kiran Sarpatwar 63 / 82

slide-81
SLIDE 81

Previous Work on CDS

Guha and Khuller, ESA 1996 ln∆+ 3 approximation algorithm in general graphs. Set cover hard. Dubhashi, Mei, Panconesi, Radhakrishnan, and Srinivasan, SODA 2003 Distributed O(ln∆) approximation algorithm in general graphs. Demaine and Hajiaghayi, SODA 2005 PTAS in planar graphs. Cheng, Huang, Li, Wu, and Du, Networks 2003 PTAS in unit disk graphs.

Kanthi Kiran Sarpatwar 63 / 82

slide-82
SLIDE 82

Previous Work on CDS

Guha and Khuller, ESA 1996 ln∆+ 3 approximation algorithm in general graphs. Set cover hard. Dubhashi, Mei, Panconesi, Radhakrishnan, and Srinivasan, SODA 2003 Distributed O(ln∆) approximation algorithm in general graphs. Demaine and Hajiaghayi, SODA 2005 PTAS in planar graphs. Cheng, Huang, Li, Wu, and Du, Networks 2003 PTAS in unit disk graphs.

Kanthi Kiran Sarpatwar 63 / 82

slide-83
SLIDE 83

Previous Work on CDS

Guha and Khuller, ESA 1996 ln∆+ 3 approximation algorithm in general graphs. Set cover hard. Dubhashi, Mei, Panconesi, Radhakrishnan, and Srinivasan, SODA 2003 Distributed O(ln∆) approximation algorithm in general graphs. Demaine and Hajiaghayi, SODA 2005 PTAS in planar graphs. Cheng, Huang, Li, Wu, and Du, Networks 2003 PTAS in unit disk graphs.

Kanthi Kiran Sarpatwar 63 / 82

slide-84
SLIDE 84

Previous Work on CDS

Guha and Khuller, ESA 1996 ln∆+ 3 approximation algorithm in general graphs. Set cover hard. Dubhashi, Mei, Panconesi, Radhakrishnan, and Srinivasan, SODA 2003 Distributed O(ln∆) approximation algorithm in general graphs. Demaine and Hajiaghayi, SODA 2005 PTAS in planar graphs. Cheng, Huang, Li, Wu, and Du, Networks 2003 PTAS in unit disk graphs.

Kanthi Kiran Sarpatwar 63 / 82

slide-85
SLIDE 85

Motivation

CDS as Virtual Backbone A “small” CDS is a good model for a virtual backbone (Bhargavan and Das, ICC 1997) Outliers A few “far off” nodes might necessitate a large CDS - making it a bad model for a backbone. More Robust Model? Can we find a good backbone if we were to “serve” say only 90% of all nodes?

Kanthi Kiran Sarpatwar 64 / 82

slide-86
SLIDE 86

Motivation

CDS as Virtual Backbone A “small” CDS is a good model for a virtual backbone (Bhargavan and Das, ICC 1997) Outliers A few “far off” nodes might necessitate a large CDS - making it a bad model for a backbone. More Robust Model? Can we find a good backbone if we were to “serve” say only 90% of all nodes?

Kanthi Kiran Sarpatwar 64 / 82

slide-87
SLIDE 87

Motivation

CDS as Virtual Backbone A “small” CDS is a good model for a virtual backbone (Bhargavan and Das, ICC 1997) Outliers A few “far off” nodes might necessitate a large CDS - making it a bad model for a backbone. More Robust Model? Can we find a good backbone if we were to “serve” say only 90% of all nodes?

Kanthi Kiran Sarpatwar 64 / 82

slide-88
SLIDE 88

Motivation

CDS as Virtual Backbone A “small” CDS is a good model for a virtual backbone (Bhargavan and Das, ICC 1997) Outliers A few “far off” nodes might necessitate a large CDS - making it a bad model for a backbone. More Robust Model? Can we find a good backbone if we were to “serve” say only 90% of all nodes?

Kanthi Kiran Sarpatwar 64 / 82

slide-89
SLIDE 89

Problem Definitions

Partial Connected Dominating Set (PCDS)

Figure: PCDS on with quota Q = 19

Definition (PCDS) Given: undirected graph G = (V,E) a quota Q Find a connected subgraph with fewest number of nodes that dominates at least Q nodes.

Kanthi Kiran Sarpatwar 65 / 82

slide-90
SLIDE 90

Problem Definitions

Partial Connected Dominating Set (PCDS)

Figure: PCDS on with quota Q = 19

Definition (PCDS) Given: undirected graph G = (V,E) a quota Q Find a connected subgraph with fewest number of nodes that dominates at least Q nodes.

Kanthi Kiran Sarpatwar 65 / 82

slide-91
SLIDE 91

Problem Definitions

Partial Connected Dominating Set (PCDS)

Figure: PCDS on with quota Q = 19

Definition (PCDS) Given: undirected graph G = (V,E) a quota Q Find a connected subgraph with fewest number of nodes that dominates at least Q nodes.

Kanthi Kiran Sarpatwar 65 / 82

slide-92
SLIDE 92

Problem Definitions

Partial Connected Dominating Set (PCDS)

Figure: PCDS on with quota Q = 19

Definition (PCDS) Given: undirected graph G = (V,E) a quota Q Find a connected subgraph with fewest number of nodes that dominates at least Q nodes.

Kanthi Kiran Sarpatwar 65 / 82

slide-93
SLIDE 93

Problem Definitions

Budgeted Connected Dominating Set (BCDS)

Figure: BCDS on with budget k = 4

Definition (BCDS) Given: undirected graph G = (V,E) a budget k Find a connected subgraph on at most k nodes that dominates as many nodes as possible.

Kanthi Kiran Sarpatwar 66 / 82

slide-94
SLIDE 94

Our Results

Theorem A polynomial time algorithm with 4ln∆+ 2 approximation guarantee for the PCDS problem. Theorem A polynomial time algorithm with

1 13(1− 1 e) approximation guarantee for the

BCDS problem.

Kanthi Kiran Sarpatwar 67 / 82

slide-95
SLIDE 95

Challenges in Solving the PCDS problem.

Dominating Set (DS) Connected Dominating Set (CDS) Converting DS to CDS It can be shown that any dominating set of size D can be connected using at most 2D extra vertices. This yields a simple O(log∆) approximation.

Kanthi Kiran Sarpatwar 68 / 82

slide-96
SLIDE 96

Challenges in Solving the PCDS problem.

How about PCDS? Unfortunately, such an approach does not work for the PCDS problem. Greedy Approach? Greedily picking vertices until Q vertices are satisfied and then connecting them - bad idea. The components could be far away from each other. Conservative Greedy? Greedily picking vertices while maintaining connectivity. Fails! Favors “locally” productive areas over “globally” rich areas.

Kanthi Kiran Sarpatwar 69 / 82

slide-97
SLIDE 97

Challenges in Solving the PCDS problem.

How about PCDS? Unfortunately, such an approach does not work for the PCDS problem. Greedy Approach? Greedily picking vertices until Q vertices are satisfied and then connecting them - bad idea. The components could be far away from each other. Conservative Greedy? Greedily picking vertices while maintaining connectivity. Fails! Favors “locally” productive areas over “globally” rich areas.

Kanthi Kiran Sarpatwar 69 / 82

slide-98
SLIDE 98

Challenges in Solving the PCDS problem.

How about PCDS? Unfortunately, such an approach does not work for the PCDS problem. Greedy Approach? Greedily picking vertices until Q vertices are satisfied and then connecting them - bad idea. The components could be far away from each other. Conservative Greedy? Greedily picking vertices while maintaining connectivity. Fails! Favors “locally” productive areas over “globally” rich areas.

Kanthi Kiran Sarpatwar 69 / 82

slide-99
SLIDE 99

An Idea

An “easier” problem. Profit function in PCDS is a non-linear “coverage” function. What happens if the profit function is “linear”? We obtain (a simpler variant of) the well known quota Steiner tree. Definition Given an undirected graph G(V,E) and profit function p : V → Z+ ∪{0} and a quota Q. Find the tree T with least number of vertices with total profit ≥ Q Theorem (Johnson et al. [SODA 2000], Garg [STOC 2005]) Quota Steiner tree has a 2-approximation algorithm.

Kanthi Kiran Sarpatwar 70 / 82

slide-100
SLIDE 100

An Idea

An “easier” problem. Profit function in PCDS is a non-linear “coverage” function. What happens if the profit function is “linear”? We obtain (a simpler variant of) the well known quota Steiner tree. Definition Given an undirected graph G(V,E) and profit function p : V → Z+ ∪{0} and a quota Q. Find the tree T with least number of vertices with total profit ≥ Q Theorem (Johnson et al. [SODA 2000], Garg [STOC 2005]) Quota Steiner tree has a 2-approximation algorithm.

Kanthi Kiran Sarpatwar 70 / 82

slide-101
SLIDE 101

An Idea

An “easier” problem. Profit function in PCDS is a non-linear “coverage” function. What happens if the profit function is “linear”? We obtain (a simpler variant of) the well known quota Steiner tree. Definition Given an undirected graph G(V,E) and profit function p : V → Z+ ∪{0} and a quota Q. Find the tree T with least number of vertices with total profit ≥ Q Theorem (Johnson et al. [SODA 2000], Garg [STOC 2005]) Quota Steiner tree has a 2-approximation algorithm.

Kanthi Kiran Sarpatwar 70 / 82

slide-102
SLIDE 102

Our Approach

Use QST

= ⇒

PCDS QST

Ideas Approximate the coverage function by a linear function? Can we do it with a logn loss? What are the candidates? How about degree function? bad idea! Somewhat surprisingly, a natural linear function defined by a greedy scheme to find the complete dominating set works!

Kanthi Kiran Sarpatwar 71 / 82

slide-103
SLIDE 103

Our Approach

Use QST

= ⇒

PCDS QST

Ideas Approximate the coverage function by a linear function? Can we do it with a logn loss? What are the candidates? How about degree function? bad idea! Somewhat surprisingly, a natural linear function defined by a greedy scheme to find the complete dominating set works!

Kanthi Kiran Sarpatwar 71 / 82

slide-104
SLIDE 104

Our Approach

Use QST

= ⇒

PCDS QST

Ideas Approximate the coverage function by a linear function? Can we do it with a logn loss? What are the candidates? How about degree function? bad idea! Somewhat surprisingly, a natural linear function defined by a greedy scheme to find the complete dominating set works!

Kanthi Kiran Sarpatwar 71 / 82

slide-105
SLIDE 105

Our Approach

Use QST

= ⇒

PCDS QST

Ideas Approximate the coverage function by a linear function? Can we do it with a logn loss? What are the candidates? How about degree function? bad idea! Somewhat surprisingly, a natural linear function defined by a greedy scheme to find the complete dominating set works!

Kanthi Kiran Sarpatwar 71 / 82

slide-106
SLIDE 106

Our Approach

Greedy Linear Function Description Use the natural greedy algorithm to define a linear function.

Kanthi Kiran Sarpatwar 72 / 82

slide-107
SLIDE 107

Our Approach

Greedy Linear Function

4 4 3 3 6 5 3 3 3 4 3

Description For each vertex, compute the number of uncovered neighbors.

Kanthi Kiran Sarpatwar 73 / 82

slide-108
SLIDE 108

Our Approach

Greedy Linear Function

1 1 1 1 6 2 2 2 3 3

Description Choose most profitable vertex and recompute for rest.

Kanthi Kiran Sarpatwar 74 / 82

slide-109
SLIDE 109

Our Approach

Greedy Linear Function

1 1 1 1 6 1 1 3 1

Description Tie breaking is arbitrary.

Kanthi Kiran Sarpatwar 75 / 82

slide-110
SLIDE 110

Our Approach

Greedy Linear Function

1 6 1 3

Description We may choose covered vertices if they qualify.

Kanthi Kiran Sarpatwar 76 / 82

slide-111
SLIDE 111

Our Approach

Greedy Linear Function

1 6 1 3

The Profit Function p(v) = # of newly covered neighbors by v.

Kanthi Kiran Sarpatwar 77 / 82

slide-112
SLIDE 112

The Algorithm for PCDS

Input Given an undirected graph G = (V,E) and a quota Q. STEP 1 Run the greedy dominating set algorithm and compute the linear profit function p : V → Z+ ∪{0}. STEP 2 Solve the quota Steiner tree on the instance (G,Q,p) and return it. Theorem Let OPT denote the optimal PCDS solution and T denote the optimal quota Steiner tree. Then T is a feasible solution for the PCDS instance and

|T| ≤ (2logn + 1)|OPT|.

Kanthi Kiran Sarpatwar 78 / 82

slide-113
SLIDE 113

Part IV Conclusion

Kanthi Kiran Sarpatwar 79 / 82

slide-114
SLIDE 114

Open Problems

Container Selection Problem Resolving the hardness of 2-dimensions problem What happens in the case of non-fixed dimensions - esp. the continuous variant Improving bounds for the discrete case. Resource Replication Problem Most results are almost tight, except the capacitated variant. Can we tighten the bounds further? Partial/Budgeted Connected Dominating Set Distributed setting? Planar graphs? Unit disk graphs? Tighten the bounds Capacitated PCDS/BCDS?

Kanthi Kiran Sarpatwar 80 / 82

slide-115
SLIDE 115

Co-authors

Samir Khuller Manish Purohit Barna Saha Viswanath Nagarajan Baruch Schieber Hadas Shachnai MohammadTaghi Hajiaghayi Joel Wolf Meghana Mande Prabhanjan Ananth Guy Kortsarz Randeep Bhatia Bhawna Gupta Robert Saccone Rohit Wagle Kirsten Hildrum Edward Pring Zubair Nabi

Kanthi Kiran Sarpatwar 81 / 82

slide-116
SLIDE 116

Thank You! Questions?

Kanthi Kiran Sarpatwar 82 / 82