Distributed Spectrum Assignment for Home WLANs Julien Herzen (EPFL) - - PowerPoint PPT Presentation

Distributed Spectrum Assignment for Home WLANs

Julien Herzen (EPFL) Ruben Merz (Swisscom) Patrick Thiran (EPFL) April 17th, 2013

1 / 14

Context

Interfering neighboring wi-fi home/office networks

www.wigle.net

• Several possible channels (center frequencies)
• Variable bandwidth (5 → 20 → 40 → 160 MHz), limited spectrum
• Non-heterogeneous density
• No central control

2 / 14

Goal

Joint allocation of channel center frequencies and bandwidths Conflicting goals:

• Bandwidthր

⇒ Capacityր

• Bandwidthր

⇒ Interference likelihoodր

3 / 14

Goal

Joint allocation of channel center frequencies and bandwidths Conflicting goals:

• Bandwidthր

⇒ Capacityր

• Bandwidthր

⇒ Interference likelihoodր f f1 f f1 Capacityր

3 / 14

Goal

Joint allocation of channel center frequencies and bandwidths Conflicting goals:

• Bandwidthր

⇒ Capacityր

• Bandwidthր

⇒ Interference likelihoodր f f1 f f1 Capacityր f f1 f2 f f1 f2 Capacity ?

3 / 14

Design Goals

• Decentralized algorithm
• Global convergence guarantees
• Online for adaptivity to time-varying conditions
• Transparent to user traffic
• Practical for implementation on off-the-shelf 802.11 hardware

Main contribution The first decentralized algorithm for joint center frequency and bandwidth adaptation with global convergence guarantees

4 / 14

Interference Model

l k → frequency time power fl fk Interference produced by k on neighbor l: Il(k) = airtime(k) · overlap(k, l)

5 / 14

Interference Model

l k → frequency time power fl fk Interference produced by k on neighbor l: Il(k) = airtime(k) · overlap(k, l) For two BSSs A and B: IA(B) =

• l∈A
• k∈B

Il(k)

5 / 14

Optimization Objective

Explicit interference vs. bandwidth trade-off: minimize E :=

• A
• B∈NA

IA(B)

• Total interference

+

• A

costA(bA)

• Sum of bandwidth ”costs”
• costA(bA) is the cost that BSS A attributes to using bandwidth bA
• E.g., costA(bA) ∝ 1/bA

6 / 14

Algorithm at BSS A

Initialization: Pick a random configuration (fA, bA) After random (exp. distributed) time intervals: Pick a random configuration (fnew, bnew) Measure e1 :=

B∈NA (IA(B) + IB(A)) + costA(bA) if A uses (fA, bA)

Measure e2 :=

B∈NA (IA(B) + IB(A)) + costA(bnew) if A uses (fnew, bnew)

Compute βT =

• 1

if e2 < e1 exp e1−e2

T

else Set (fA, bA) = (fnew, bnew) with probability βT

7 / 14

Convergence

Metropolis sampling for center frequency and bandwidth Theorem Denote Xn the global state of the network after the n-th iteration. Consider a network where all the BSSs run our algorithm using a given parameter T. Then Xn is a Markov chain, and it converges in distribution to π(X) ∝ e−E(X)/T, where X is a global state.

• State gets arbitrarily close to optimal for T small enough
• T encodes a trade-off between likelihood of local optima and

asymptotic efficiency

8 / 14

Implementation

• 802.11g with 5, 10 and 20 MHz channel widths
• Interference measured by spending ≤ 50 ms. out-of-band
• Optional client collaboration for interference measurement
• C++ implementation using Click in userspace
• costA(bA) = 1/bA

40 m 65 m

9 / 14

Performance Evaluation

UDP traffic, client-agnostic: UDP traffic, client-aware:

1000 2000 3000 4000 time [s] 20 30 40 50 60 70 80 total throughput [Mbps] Bench 1000 2000 3000 4000 time [s] 20 30 40 50 60 70 80 total throughput [Mbps] Bench

TCP traffic, client-agnostic: TCP traffic, client-aware:

1000 2000 3000 4000 time [s] 20 25 30 35 40 45 50 55 60 65 total throughput [Mbps] Bench 1000 2000 3000 4000 time [s] 20 25 30 35 40 45 50 55 60 65 total throughput [Mbps] Bench

”Bench” line: centralized graph-coloring for fixed-width channels

10 / 14

Simulation

• Random distribution of BSSs on the plane
• Capacity of link l = bl · log(1 + SINR)
• costA(bA) = c/bA, optimization objective becomes:

minimize

• A
• B∈NA

IA(B) + c ·

• A

1/bA

• c = 0: minimize interference
• c → ∞: use largest bandwidth, irrespective of interference

11 / 14

Simulation

• Random distribution of BSSs on the plane
• Capacity of link l = bl · log(1 + SINR)
• costA(bA) = c/bA, optimization objective becomes:

minimize

• A
• B∈NA

IA(B) + c ·

• A

1/bA

• c = 0: minimize interference
• c → ∞: use largest bandwidth, irrespective of interference

2 4 6 8 10 c 0.0 0.2 0.4 0.6 0.8 1.0 normalized value fairness capacity interference 11 / 14

Simulation

• Random distribution of BSSs on the plane
• Capacity of link l = bl · log(1 + SINR)
• costA(bA) = c/bA, optimization objective becomes:

minimize

• A
• B∈NA

IA(B) + c ·

• A

1/bA

• c = 0: minimize interference
• c → ∞: use largest bandwidth, irrespective of interference

2 4 6 8 10 c 0.0 0.2 0.4 0.6 0.8 1.0 normalized value fairness capacity interference 50 100 150 200 250 300 350 400 average spatial density [BSS/km2] 1000 2000 3000 4000 5000 6000 7000 8000 total capacity c = 0 c = 100 c = 1 11 / 14

Simulation

Improvement with respect to random allocations after 5 iterations: after 20 iterations:

20 40 60 80 100 percentage of BSSs running SAW 20 40 60 80 100 120 140 percentage improvement % interference decrease % capacity increase 20 40 60 80 100 percentage of BSSs running SAW 50 100 150 200 250 300 percentage improvement % interference decrease % capacity increase

12 / 14

Simulation

total spectrum: 45 MHz total spectrum: 70 MHz

10 20 30 40 50 60 average iterations per AP 0.0 0.2 0.4 0.6 0.8 1.0 normalized value Capacity, c+w Interference, c Capacity, c Interference, c+w 10 20 30 40 50 60 average iterations per AP 0.0 0.2 0.4 0.6 0.8 1.0 normalized value Capacity, c+w Capacity, c Interference, c Interference, c+w 10 20 30 40 50 60 average iterations per AP 0.3 0.4 0.5 0.6 0.7 0.8 fairness index Fairness, c+w Fairness, c 10 20 30 40 50 60 average iterations per AP 0.3 0.4 0.5 0.6 0.7 0.8 fairness index Fairness, c+w Fairness, c

13 / 14

Conclusion

• Distributed, joint allocation of center frequencies and bandwidths
• Bandwidth influences both capacity and interference; ideal spectrum

consumption should depend on network density

• Optimization of an explicit trade-off between interference mitigation

and use of advantageous bandwidths

• Simple optimization objectives yield best results irrespective of

network density

• Large capacity improvements, even when not all BSSs run the

algorithm

• Testbed implementation shows feasibility and improvements compared

to fixed-width graph coloring

14 / 14

1 / 5

Some Related Work

• Channel allocation / graph coloring, e.g., [Akella et al. 2005,

Kauffmann et al. 2007, Duffy et al. 2011, Leith et al. 2012]

◮ Main goal: minimize interference (no variable bandwidth)

• Variable bandwidth / white spaces, e.g., [Chandra et al. 2008, Bahl et
• al. 2009, Rayanchu et al. 2011]

◮ Heuristics, no focus on self-organization 2 / 5

Micro-sensing

compute list of bands to scan

• ptionally: inform clients

micro-sensing

• ptionally: receive

link stats from clients decision time time

• init. switch

block traffic

• init. switch back

tm-s tswitch tswitch tsensing tm-s

unblock traffic timer fires

20 40 60 80 100 120 time [s] 2 4 6 8 10 12 14 16 throughput [Mbps] Link 1 starts sensing Link 1 switches band

TCP link 1 TCP link 2

3 / 5

Channel widths

−20 20 −70 −60 −50 −40 −30

5 MHz dBm MHz

−20 20 −70 −60 −50 −40 −30

10 MHz MHz

−20 20 −70 −60 −50 −40 −30

20 MHz MHz

4 / 5

Performance Evaluation (uplink)

UDP traffic, client-agnostic: UDP traffic, client-aware:

1000 2000 3000 4000 time [s] 20 30 40 50 60 70 80 total throughput [Mbps] Bench 1000 2000 3000 4000 time [s] 20 30 40 50 60 70 80 total throughput [Mbps] Bench

TCP traffic, client-agnostic: TCP traffic, client-aware:

1000 2000 3000 4000 time [s] 20 25 30 35 40 45 50 55 60 65 total throughput [Mbps] Bench 1000 2000 3000 4000 time [s] 20 25 30 35 40 45 50 55 60 65 total throughput [Mbps] Bench

”Bench” line: centralized graph-coloring for fixed-width channels

5 / 5