Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
Network Design withCoverage Costs
Siddharth Barman 1 Shuchi Chawla 2 Seeun William Umboh 2
1Caltech
2University of Wisconsin-Madison
APPROX-RANDOM 2014
Motivation
Physical Flow vs Data Flow
vs.
Commodity Network Internet
Unlike physical commodities, data can be easily duplicated,compressed, and combined.
Physical Flow vs Data Flow
Flow = 1
Flow = 1
Flow = 2
P = 10010
P = 10010
P = 10010vs.
Unlike physical commodities, data can be easily duplicated,compressed, and combined.
Why Coverage Costs?
• Want a cost structure that captures bandwidth savings obtainedby eliminating redundancy in data.
▶ Packet deduplication capability exists in networks today.
Problem Statement
Load Function
We focus on redundant-data elimination:
Load on an edge ℓe = # distinct data packets on it.
P1 : 11100P1 : 11100P2 : 00011
Load = 2
Input:• Graph with edge costs, g terminal pairs (s1, t1), . . . ,(sg, tg)
• A global set of packets Π.• A demand set Dj ⊂ Π for each j.
Goal: Find (sj, tj) path on which to route Dj for each j minimizing
s1 t1
s2 t2
Input:• Graph with edge costs, g terminal pairs (s1, t1), . . . ,(sg, tg)• A global set of packets Π.
• A demand set Dj ⊂ Π for each j.
Goal: Find (sj, tj) path on which to route Dj for each j minimizing
�s1 t1
s2 t2
Input:• Graph with edge costs, g terminal pairs (s1, t1), . . . ,(sg, tg)• A global set of packets Π.• A demand set Dj ⊂ Π for each j.
Goal: Find (sj, tj) path on which to route Dj for each j minimizing
�
D1D2
s1 t1
s2 t2
Input:• Graph with edge costs, g terminal pairs (s1, t1), . . . ,(sg, tg)• A global set of packets Π.• A demand set Dj ⊂ Π for each j.
Goal: Find (sj, tj) path on which to route Dj for each j minimizing
∑edges
Cost of edge×Load on edge
�
D1D2
s1 t1
s2 t2
Input:• Graph with edge costs, g terminal pairs (s1, t1), . . . ,(sg, tg)• A global set of packets Π.• A demand set Dj ⊂ Π for each j.
Goal: Find (sj, tj) path on which to route Dj for each j minimizing
∑edges
Cost of edge×# distinct packets on edge
�
D1D2
s1 t1
s2 t2
Input:• Graph with edge costs, g terminal pairs (s1, t1), . . . ,(sg, tg)• A global set of packets Π.• A demand set Dj ⊂ Π for each j.
Goal: Find (sj, tj) path on which to route Dj for each j minimizing
∑edges
Cost of edge×# distinct packets on edge
�
D1D2
s1 t1
s2 t2
Input:• Graph with edge costs, g terminal pairs (s1, t1), . . . ,(sg, tg)• A global set of packets Π.• A demand set Dj ⊂ Π for each j.
Goal: Find (sj, tj) path on which to route Dj for each j minimizing
∑edges
Cost of edge×# distinct packets on edge
�
D1D2
s1 t1
s2 t2
|D1|
|D2|
|D1|
|D2||D1 U D2|
More generally, we consider terminal groups.Input:
• Graph with edge costs, g terminal groups X1, . . . ,Xg ⊂ V.• A global set of packets Π.• A demand set Dj ⊂ Π for each j.
Goal: Find a Steiner tree for Xj on which to broadcast Dj for each jminimizing
∑edges
Cost of edge×# distinct packets on edge
Special cases (for terminal pairs):
• Pairwise disjoint demands =⇒ shortest-paths
s1 t1
s2 t2
|D1|
|D2|
�
D1 D2
• Identical demands =⇒ Steiner forests1 t1
s2 t2
|D|
|D|
|D|
|D||D|
�
D1, D2 = D
Special cases (for terminal pairs):
• Pairwise disjoint demands =⇒ shortest-paths
s1 t1
s2 t2
|D1|
|D2|
�
D1 D2
• Identical demands =⇒ Steiner forests1 t1
s2 t2
|D|
|D|
|D|
|D||D|
�
D1, D2 = D
Challenges
• Intuitively, difficulty depends on complexity of demand sets
�
• Desire an approximation in terms of g (number of groups), e.g.O(logg) (tree embeddings only yield O(logn)).
Challenges
• Intuitively, difficulty depends on complexity of demand sets
�
• Desire an approximation in terms of g (number of groups), e.g.O(logg) (tree embeddings only yield O(logn)).
Our results
Laminar DemandsThe family of demand sets is said to be laminar if ∀ i, j one of thefollowing holds: Di ∩Dj = ϕ or Di ⊂ Dj or Dj ⊂ Di.
⇧
D1 D2
D3
D4D5
⇧
D2D1
D3 D5D4
TheoremNDCC with laminar demands admits a 2-approximation.
Previous work: O(log |Π|) [Barman-Chawla ’12].
Laminar DemandsThe family of demand sets is said to be laminar if ∀ i, j one of thefollowing holds: Di ∩Dj = ϕ or Di ⊂ Dj or Dj ⊂ Di.
⇧
D1 D2
D3
D4D5
⇧
D2D1
D3 D5D4
TheoremNDCC with laminar demands admits a 2-approximation.
Previous work: O(log |Π|) [Barman-Chawla ’12].
Sunflower DemandsThe family of demand sets is said to be a sunflower family if thereexists a core C ⊂ Π such that ∀ i, j, Di ∩Dj = C.
C
Di \ C
Dj \ C
No O(log1/4−γ g)-approximation for any γ > 0 (reduction fromsingle-cable buy-at-bulk [Andrews-Zhang ’02]).
TheoremNDCC with sunflower demands admits an O(log g) approximation overunweighted graphs with V =
∪j Xj.
Sunflower DemandsThe family of demand sets is said to be a sunflower family if thereexists a core C ⊂ Π such that ∀ i, j, Di ∩Dj = C.
C
Di \ C
Dj \ C
No O(log1/4−γ g)-approximation for any γ > 0 (reduction fromsingle-cable buy-at-bulk [Andrews-Zhang ’02]).
TheoremNDCC with sunflower demands admits an O(log g) approximation overunweighted graphs with V =
∪j Xj.
Sunflower DemandsThe family of demand sets is said to be a sunflower family if thereexists a core C ⊂ Π such that ∀ i, j, Di ∩Dj = C.
C
Di \ C
Dj \ C
No O(log1/4−γ g)-approximation for any γ > 0 (reduction fromsingle-cable buy-at-bulk [Andrews-Zhang ’02]).
TheoremNDCC with sunflower demands admits an O(log g) approximation overunweighted graphs with V =
∪j Xj.
Related WorkNetwork design with economies of scale.
∑edges
Cost of edge×Load on edge
• Submodular costs on edges [Hayrapetyan-Swamy-Tardos ’05]▶ O(logn) via tree embeddings.
• Buy-at-Bulk Network Design [Salman et al. ’97]▶ Single-source: O(1) [Guha et al. ’01, Talwar ’02,
Gupta-Kumar-Roughgarden ’03]▶ Multiple-source: Ω(log
14 n) hardness [Andrews ’04]
Related WorkNetwork design with economies of scale.
∑edges
Cost of edge×Load on edge
• Submodular costs on edges [Hayrapetyan-Swamy-Tardos ’05]▶ O(logn) via tree embeddings.
• Buy-at-Bulk Network Design [Salman et al. ’97]▶ Single-source: O(1) [Guha et al. ’01, Talwar ’02,
Gupta-Kumar-Roughgarden ’03]▶ Multiple-source: Ω(log
14 n) hardness [Andrews ’04]
Related WorkNetwork design with economies of scale.
∑edges
Cost of edge×Load on edge
• Submodular costs on edges [Hayrapetyan-Swamy-Tardos ’05]▶ O(logn) via tree embeddings.
• Buy-at-Bulk Network Design [Salman et al. ’97]▶ Single-source: O(1) [Guha et al. ’01, Talwar ’02,
Gupta-Kumar-Roughgarden ’03]▶ Multiple-source: Ω(log
14 n) hardness [Andrews ’04]
Approach: Laminar
Laminar⇧
D1 D2
D3
D4D5
⇧
D2D1
D3 D5D4
• Naive LP does not have much structure• Key idea: exploit laminarity to write a different LP• Run an extension of AKR-GW primal-dual algorithm to get
2-approx
Approach: Sunflower(single-source for this talk)
Sunflower Family of DemandsThe family of demand sets is said to be a sunflower family if thereexists C ⊂ Π such that ∀ i, j, Di ∩Dj = C.
C
Di \ C
Dj \ C
TheoremNetwork Design with Coverage Costs in the sunflower demands settingadmits an O(log g) approximation over unweighted graphs with V =
∪j Xj.
Structure of OPTDefine:
• E∗j = Steiner tree for Xj on which OPT broadcasts Dj ⊃ C• E∗0 =
∪j E
∗j
Note: E∗0 spans V because V =∪
j Xj and single-source (each Xjcontains s)
c(OPT) = |C| · c(E∗0)+∑j|Dj \C| · c(E∗j )
≥ |C| · c(MST)+∑j|Dj \C| · c(opt Steiner for Xj)
Structure of OPTDefine:
• E∗j = Steiner tree for Xj on which OPT broadcasts Dj ⊃ C• E∗0 =
∪j E
∗j
Note: E∗0 spans V because V =∪
j Xj and single-source (each Xjcontains s)
c(OPT) = |C| · c(E∗0)+∑j|Dj \C| · c(E∗j )
≥ |C| · c(MST)+∑j|Dj \C| · c(opt Steiner for Xj)
Structure of OPTDefine:
• E∗j = Steiner tree for Xj on which OPT broadcasts Dj ⊃ C• E∗0 =
∪j E
∗j
Note: E∗0 spans V because V =∪
j Xj and single-source (each Xjcontains s)
c(OPT) = |C| · c(E∗0)+∑j|Dj \C| · c(E∗j )
≥ |C| · c(MST)+∑j|Dj \C| · c(opt Steiner for Xj)
Group Spanners
DefinitionFor a graph G and g groups X1, . . . ,Xg ⊂ V, a subgraph H is a (α,β )group spanner if
• c(H)≤ αc(MST),• c(opt Steiner for Xj in H)≤ β c(opt Steiner for Xj in G) for all j.
Usual spanners: Xjs are all vertex pairs.
• c(shortest u− v path in H)≤ β c(shortest u− v path in G) for allu,v ∈ V.
Group Spanners
DefinitionFor a graph G and g groups X1, . . . ,Xg ⊂ V, a subgraph H is a (α,β )group spanner if
• c(H)≤ αc(MST),• c(opt Steiner for Xj in H)≤ β c(opt Steiner for Xj in G) for all j.
Usual spanners: Xjs are all vertex pairs.
• c(shortest u− v path in H)≤ β c(shortest u− v path in G) for allu,v ∈ V.
Using Group Spanners
Given (α,β ) group spanner H, route Dj along 2-approximate Steinertree for Xj in H.
Cost ≤ |C| · c(H)+∑j|Dj \C| ·2c(opt Steiner for Xj in H)
≤ |C| ·αc(MST)+∑j|Dj \C| ·2β c(opt Steiner for Xj)
≤ max{α,2β}OPT .
Group Spanners Result
LemmaGiven an unweighted graph G and g groups X1, . . . ,Xg ⊂ V with V =
∪j Xj,
we can construct in polynomial time a (O(1),O(log g)) group spanner.
TheoremNetwork Design with Coverage Costs in the sunflower demands settingadmits an O(log g) approximation over unweighted graphs with V =
∪j Xj.
Summary• Coverage cost model for designing information networks.
P1 : 11100P1 : 11100P2 : 00011
Load = 2
• Laminar demands: 2-approximation via primal-dual.⇧
D1 D2
D3
D4D5
⇧
D2D1
D3 D5D4
• Sunflower demands: O(log g)-approximation (unweighted graph,no Steiner vertices).
C
Di \ C
Dj \ C
• Group spanners: (O(1),O(log g)) group spanners (unweightedgraph, no Steiner vertices).
Summary• Coverage cost model for designing information networks.
P1 : 11100P1 : 11100P2 : 00011
Load = 2
• Laminar demands: 2-approximation via primal-dual.⇧
D1 D2
D3
D4D5
⇧
D2D1
D3 D5D4
• Sunflower demands: O(log g)-approximation (unweighted graph,no Steiner vertices).
C
Di \ C
Dj \ C
• Group spanners: (O(1),O(log g)) group spanners (unweightedgraph, no Steiner vertices).
Summary• Coverage cost model for designing information networks.
P1 : 11100P1 : 11100P2 : 00011
Load = 2
• Laminar demands: 2-approximation via primal-dual.⇧
D1 D2
D3
D4D5
⇧
D2D1
D3 D5D4
• Sunflower demands: O(log g)-approximation (unweighted graph,no Steiner vertices).
C
Di \ C
Dj \ C
• Group spanners: (O(1),O(log g)) group spanners (unweightedgraph, no Steiner vertices).
Open Problems
• Group spanners: (O(log g),O(log g)) in general?
• Sunflower demands: O(log g) approximation in general?• Terminal pairs with single-source: O(1)?
Thanks!
Open Problems
• Group spanners: (O(log g),O(log g)) in general?• Sunflower demands: O(log g) approximation in general?
• Terminal pairs with single-source: O(1)?
Thanks!
Open Problems
• Group spanners: (O(log g),O(log g)) in general?• Sunflower demands: O(log g) approximation in general?• Terminal pairs with single-source: O(1)?
Thanks!
Open Problems
• Group spanners: (O(log g),O(log g)) in general?• Sunflower demands: O(log g) approximation in general?• Terminal pairs with single-source: O(1)?
Thanks!