Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
Rate Feasibility inWireless Networks
INFOCOM 2008Ramakrishna Gummadi, UIUC (speaker)
Kyomin Jung, MIT
Devavrat Shah, MIT
RS Sreenivas, UIUC
Ramakrishna Gummadi, UIUC – p. 1/35
IntroductionSimple abstraction of a wireless network:Vertices: wireless nodesEdges: possible communication links
Ramakrishna Gummadi, UIUC – p. 2/35
IntroductionSimple abstraction of a wireless network:Vertices: wireless nodesEdges: possible communication links
Assume each link has a unit capacity (withoutinterference)
Ramakrishna Gummadi, UIUC – p. 3/35
IntroductionSimple abstraction of a wireless network:Vertices: wireless nodesEdges: possible communication links
Assume each link has a unit capacity (withoutinterference)
But Interference ⇒ Not all links can besimultaneously active
Ramakrishna Gummadi, UIUC – p. 4/35
IntroductionSimple abstraction of a wireless network:Vertices: wireless nodesEdges: possible communication links
Assume each link has a unit capacity (withoutinterference)
But Interference ⇒ Not all links can besimultaneously active
A fundamental question: Can a given vectorof link demands be satisfied simultaneously?
Ramakrishna Gummadi, UIUC – p. 5/35
Rate Feasibility
Λ - set of all Rate Vectors for which thereexists a TDMA schedule.
r - the query Rate Vector
Problem: Does r ∈ Λ?
Ramakrishna Gummadi, UIUC – p. 6/35
Rate Feasibility
Λ - set of all Rate Vectors for which thereexists a TDMA schedule.
r - the query Rate Vector
Problem: Does r ∈ Λ?
T - the incidence matrix for the set of allsubsets of non conflicting links
z = min 1Tx
Tx = r,
Then r ∈ Λ iff z ≤ 1
Ramakrishna Gummadi, UIUC – p. 7/35
Is r “feasible”?
An answer to whether a given rate vector isfeasible depends on the following:
1. Graph structure (the specific probleminstance)
2. Interference constraints (the problemmodel - critical for the computationalcomplexity)
Ramakrishna Gummadi, UIUC – p. 8/35
InterferencePrimary Constraints : Links conflict when theyshare a node.
pi
(a) (b) (c)
pj pkpj
pj
pi pi
pkpk
Ramakrishna Gummadi, UIUC – p. 9/35
InterferenceSecondary constraints : Links can conflict evenwhile not sharing a common node.
pk
pl
pi
pj
Links (pi, pj) and (pk, pl) conflict
Ramakrishna Gummadi, UIUC – p. 10/35
Background . . .
Primary and Secondary constraints(appropriate for wireless networks): problemis NP-hard [Arikan,’84]
Ramakrishna Gummadi, UIUC – p. 11/35
Background . . .
Primary and Secondary constraints(appropriate for wireless networks): problemis NP-hard [Arikan,’84]
Primary alone: Polynomial schedulingalgorithms exist [Hajek, Sasaki ’88]
Ramakrishna Gummadi, UIUC – p. 12/35
Background . . .
Primary and Secondary constraints(appropriate for wireless networks): problemis NP-hard [Arikan,’84]
Primary alone: Polynomial schedulingalgorithms exist [Hajek, Sasaki ’88]
Question: Are there restricted subclasses ofwireless networks that are tractable?
Ramakrishna Gummadi, UIUC – p. 13/35
Our Results . . .Bounded density
A randomized approximation algorithmMore generally relevant to membership incomplex convex sets
Ramakrishna Gummadi, UIUC – p. 14/35
Our Results . . .Bounded density
A randomized approximation algorithmMore generally relevant to membership incomplex convex sets
Fixed width slabGeneralize results on fractional coloringunit disk graphs to a more general classDeterministic, Exact solution.
Ramakrishna Gummadi, UIUC – p. 15/35
Adjoint Graph
Adjoint Graph:Vertices - Link midpointsEdges - defined by link conflicts (bothprimary and secondary)
Communication (Interference) radius - rC(rI)for wireless nodes ⇒ Structure of its adjoint.
Link Rate-feasibility ≡ Node rate-feasibility inadjoint graph with independent sets.
Ramakrishna Gummadi, UIUC – p. 16/35
Bounded density
Let B(v,R) = |{u ∈ V : u 6= v, d(u, v) < R}|Then, G has bounded density D > 0, if for allv ∈ V
B(v,R)
R2≤ D
Bounded density of wireless graph ⇒Bounded density of adjoint
∃R > 0 such that two vertices farther than Rin the adjoint have no edge.
Ramakrishna Gummadi, UIUC – p. 17/35
Algorithm overview
Partition the graph into:
Regions, small enough to be efficiently solved
Boundaries, thick enough separate them
Ramakrishna Gummadi, UIUC – p. 18/35
Algorithm overview
Partition the graph appropriately.
Ramakrishna Gummadi, UIUC – p. 19/35
Algorithm overview
Partition the graph appropriately.
Solve for feasible schedules in each region
Ramakrishna Gummadi, UIUC – p. 20/35
Algorithm overview
Partition the graph appropriately.
Solve for feasible schedules in each region
Merge them to get a global schedule –(schedule satisfies everyone except nodes inboundary)
Ramakrishna Gummadi, UIUC – p. 21/35
Algorithm overview
Partition the graph appropriately.
Solve for feasible schedules in each region
Merge them to get a global schedule –(schedule satisfies everyone except nodes inboundary)
Randomize the boundary and compute anaverage schedule obtained over sufficientlylarge iterations. – (such that its unlikely for anode to fall in the boundary.)
Ramakrishna Gummadi, UIUC – p. 22/35
Main Properties
(w.h.p.) Given ǫ > 0:
(1) If r ∈ Λ, algorithm outputs a TDMA scheduleT̂ = (αk, Ik)k≤M (with M = poly(n)), such that:
(1 − ǫ)r ≤∑
k
αkIk,
(2) If (1 − ǫ)r /∈ Λ, it declares NOT FEASIBLE.
Complexity is O
n log n2O
(
R2D
ǫ2
)
ǫ
≈ O(n log n).
Ramakrishna Gummadi, UIUC – p. 23/35
Fixed width slabAssume wireless nodes are restricted in onecoordinate. (e.g. Y- dimension bounded,while X can range all over: IVHS)
X
Y
Ramakrishna Gummadi, UIUC – p. 24/35
Adjoint Structure
Represent the vertex corresponding to thelink between pi,pj as (pi,pj). Then:
1. There cannot be an edge between vertices(pi,pj) and (pk,pl) if‖(pi,pj) − (pk,pl)‖2 ≥ rc + ri
2. There will be an edge between vertices(pi,pj) and (pk,pl) if‖(pi,pj) − (pk,pl)‖2 < ri − rc
Ramakrishna Gummadi, UIUC – p. 25/35
Useful ResultsLet M be the incidence matrix for the set ofall independent sets of a graph G. Fractionalcoloring on G is to solve:
z = min 1Tx
Mx ≥ 1
x ≥ 0
Fractional coloring problem on a Unit DiskGraph has polynomial solution if nodes arewithin a fixed width slab. [Matsui 02]
Ramakrishna Gummadi, UIUC – p. 26/35
Contrast with RateFeasibility
Fractional coloring algorithms solve a Nodebased rate feasibility problem with equalrates.
Ramakrishna Gummadi, UIUC – p. 27/35
Contrast with RateFeasibility
Fractional coloring algorithms solve a Nodebased rate feasibility problem with equalrates.
Wireless adjoints do not have unit disk graphstructure.
D1
D2
Ramakrishna Gummadi, UIUC – p. 28/35
Key Observations
1. The results for unit disk graphs can beextended to a more general class called(dmin, dmax) graphs, which includes ourwireless adjoints.
Ramakrishna Gummadi, UIUC – p. 29/35
Key Observations
1. The results for unit disk graphs can beextended to a more general class called(dmin, dmax) graphs, which includes ourwireless adjoints.
2. The equal rates in fractional coloring can begeneralized to arbitrary rates.
Ramakrishna Gummadi, UIUC – p. 30/35
(dmin, dmax) - graphs
G(P ) = (P,E) induced by point setP ⊆ {(x, y) ∈ R2} satisfies the followingproperties:
1. Property P1:∀p1,p2 ∈ P, ‖p1 − p2‖ ≥ dmax ⇒ (p1,p2) /∈ E
2. Property P2: ‖p1 − p2‖ < dmin ⇒ (p1,p2) ∈ E
If dmin = dmax, we get Unit Disk Graphs.
Note: Between dmin and dmax we have norestriction.
Ramakrishna Gummadi, UIUC – p. 31/35
(dmin, dmax) - graphs
Theorem: Matsui’s algorithm can begeneralized to (dmin, dmax) graphs.
Proof idea: The properties P1 and P2, thoughimplicitly buried together in the specificationof the unit disk graphs, are actually onlyneeded separately.
Ramakrishna Gummadi, UIUC – p. 32/35
Rate Feasibility onFixed Width Slabs
Wireless adjoints are (rI − rC , rI + rC)graphs.
Apply the Fractional Coloring algorithm,generalized to arbitrary rates.
Works in polynomial time, due to ourgeneralization from unit disk graphs to(dmin, dmax) graphs.
Ramakrishna Gummadi, UIUC – p. 33/35
Conclusion andFuture interests
Have shown two relevant subclasses whererate feasibility is tractable for wirelessnetworks, even though the general problem isNP-hard.
Negatives: These are quite centralizeddescriptions of the algorithms.
Potential interest for practicality: Finding evenlower complexity distributed versions.
Integrate Rate Feasibility checkingseamlessly into cross layer design.
Ramakrishna Gummadi, UIUC – p. 34/35
Thank you for your attention!
. . . Questions?
Ramakrishna Gummadi, UIUC – p. 35/35