Wireless Mesh Networks A Bhorkar University Qualifying Examination University of California San Diego Motivation 2 Wireless Mesh Networks Characteristics Fixed unlimited energy virtually unlimited processing power ID: 246043
Download Presentation The PPT/PDF document "A Perspective on Routing in" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
A Perspective on Routing in Wireless Mesh Networks
A. BhorkarUniversity Qualifying ExaminationUniversity of California, San DiegoSlide2
Motivation
2
Wireless Mesh Networks
Characteristics
:
Fixed, unlimited energy, virtually unlimited processing power
Dynamism
– Link Quality
Optimize
–
Low latency, High throughputSlide3
3
Research Objectives
3
To design
a routing
policy that
exhibits good
delay performance
1
Congestion Diversity Protocol (CDP)
Propose a metric that reacts
to congestion
Experimental
testbed
in Calit2
Evaluate CDP
To design an opportunistic routing policy that enhances
throughput by
minimizing the expected number of
transmissions
2Slide4
Congestion Impact
Congestion
Increase
in traffic
Increase in network congestion
Increase in buffer packet drops
and delay
4
Need
to provision for
congestion in the networkSlide5
5
Time
invariant policies
AODV
Hop count
GeRaF
Geographical
distance
SRCR
ETX
metric
Why New
Routing Policy?
S
D
Susceptible
to
congestion
Do not account for queue congestion Slide6
Selects forwarder with the
smallest queue backlogSpreads the trafficIncreases contention and
delay
6
Time variant (congestion-aware)
Multipath Routing
(SMR-LB)
Create multiple (2) paths
Timeshare routes proportional to their average delay
Roots in minimum delay routing [Gallager84]
Slow to react
Backpressure routing
(BCP 10)
Why New
Routing Policy?
S
D
Need routing policy suitable for all traffic conditionsSlide7
OutlineCDP: A bird’s eye viewCDP components
CDP issuesEvaluation Discussion and ongoing work7Slide8
Congestion Diversity Protocol : Bird’s eye view
GoalDecentralized computation of the proposed congestion measureRoute efficiently to avoid congestion
S
0.5
0.8
0.8
0.8
0.9
D
CDP’s
core is congestion
a
ware
r
outing
Routing decisions are based on congestion in the network
It exploits
p
ath
diversity
when available
8Slide9
Congestion Measure
Nodes are ordered according to a congestion measure Vi (t)V
i (
t
)
is the expected delay to reach the destination
V
i
(
t
)
<
V
j
(t
)
for all j
i
has lowest
rank
9
Routing policy:
select the forwarder with the lowest
rank
If
V2(t)
<
V1(t)
, node S will select
2 as the next forwarder
S
2
1
0.5
0.8
0.8
0.8
0.9
DSlide10
Characterization of congestion measure10
Vi (t) are solutions to “
Bellman-Ford”
like equation
Local congestion
Expected
delay from neighbor
j
Effective draining time for node
i
Link success probability between
i
and
j
Queue length at node
iSlide11
Congestion Granularity
CDP uses instantaneous queue lengths to detect congestionCDP tracks the state of the networkChanges in congestion can be rapid Due to channel fluctuations and traffic variationsTackles the short term instantaneous congestionDifferent from “minimum delay routing” [Gallager84]
Uses long term congestion for routing
Different from congestion control at transport layer
Transport layer caters to statistical, long term congestion
11Slide12
Mac Layer
Compute V and broadcast control packet
Control packets from neighbors
Phy layer
RX
TX
Data Packets
Probe Packets
Data Packet
Probe Packets
CDP Design
Determine Link probability
Determine
V(t
)
200ms
Delayed feedback
12Slide13
Link determinationLink quality updates (estimation of pij’s)
Active probing + passive probingBroadcast-based active probing Passive probing Monitor received data packets
Detect missing data packets
using sequence numbers
Estimate the link
success probabilities
This reduces probing overhead
A
B
Capacity
overhead - dedicated probe packets
13Slide14
1
D
N
i
j
Scheduling
802.11 Mac
Node
i
selects one node
j
as a relay
Node
j
takes the responsibility of the packet
14
j
Control packets computing
V
are critical
Can get stuck due to congestion
Assign higher priority than data packets
Use driver’s priority queuing Slide15
Formation of loopsCurse of decentralized implementationTemporary build-up of queue at F causes packets to loop in C-B
Counting to infinity situationSplit horizon rule is used to prevent routing loopsNode C knows B’s next hop is CC sets VB = ∞, C does not learn from BA node advertises routes as unreachable to the node through which they were learned
Design Issues (1)
15
F
B
D
A
CSlide16
Design Issues (2) Delay vs. link probability curveAssumption: Non linear, difficult to characterizePacket are dropped after RETRY_MAX
Chosen a naïve solutionNeglect links with p < ThresholdChoose Threshold = 0.616
?
Link probability
Media access
delay
0
1Slide17
Experimental Setup
17
~100
mt
.
Indoor Network of
12
Alix
nodes in Atkinson Hall (Calit2
) 500Mhz CPU and 1GB flash.
MAC:
Atheros
IEEE 802.11 chipset (5212) using the
MadWifi
-NG driver
.
Multiple
hops (5 hops maximum)
13
dBm
transmit power
4 neighbors average
Performance Measures
Mean end to end delay
Mean delivery
ratio
Comparative protocols:
SRCR
Back
Pressure (BP)
Enhanced Back
Pressure (E-BP)Slide18
Canonical Example
18
50x
100x
SRCR
CDP
10
BP
14
16
17Slide19
Examples(1)
19
10
14
17
16Slide20
Examples(2)
20
10
14
17
4
8Slide21
Examples(3)
21
Single-hop flows: No gain
Mean Delay
Delivery ratioSlide22
22
¥
Inter-flow interference
Increase in potential number of transmitters
Nothing for free
11
17Slide23
23
Intra-flow interference
(interference artificially increased by increasing power at interfering nodes)
Multiple paths create contention
Potential to degrade the performance
Nothing for free
¥
13
7
11
5Slide24
Rule of Thumb CDP shows improvement when Network provides diversity
Many multi-hop routes are availableMany short length flows congest long length flowInterflow and intraflow interference do not dominate the congestion24Slide25
Performance Results (Parameters)UDP trafficRandom source destination pairs Multiple concurrent flows (2 flows)
Choose UDP rates randomly 10 random topologies80 sample pointsPlot CDF of mean delay of CDP - mean delay of candidate protocol Ratio of delivery ratios 25Slide26
Performance Results
26
High Load
Low Load
30%Slide27
CDP: Summary
CDP exploits the congestion diversity in the networkCDP achieves high gain in the delay and delivery ratio for UDP trafficCDP shows up to 50x gain over SRCR and 100x over backpressure routing
Routing Policy
Delay Performance
Low Traffic
High Traffic
Shortest-Path
Good delay performance
Poor delay
Backpressure
Poor delay performance
Guaranteed bounded delay
Poor delay in practice
CDP
Good delay performance
Good delay in practice
27Slide28
28
15
13
5
11
7
4
8
6
17
16
10
14
0.1xsec
Next hop for CDP
Current Focus (Interaction with TCP)
TCP performance hurts
Self interference
Reordering of packets
Fast retransmit
Which factor is more
dominant
?Slide29
Investigate the impact of CDP on TCP and mixed trafficDetect the cause of performance loss for TCPTackle reordering issue in TCP
Include prioritized MAC in CDP Change backoff proportional to the congestion
Immediate Goals
29Slide30
Develop multi-rate CDP utilizing multiple PHY ratesDefine congestion measure for multi-rate CDP
Implement opportunistic version of CDP on testbed Modify MadWifi driver’s 802.11 Mac layer
Immediate Goals
30
S
1
2
D
3
0.8
0.8
0.8
0.8
0.2
D
SSlide31
Ongoing work
31Objective
Timeline
CDP performance
Reordering of TCP (July)
MIXED/TCP (August)
Summer
Infocom
,
ToN
Prioritized MAC
October
Fall
Multi rate CDP
December
Fall
ORCD (Opportunistic CDP)
Mac layer modification (August)
Analysis/Delay performance (Feb)
Spring 10
IMC/
MobicomSlide32
32
Research Objectives
To design an opportunistic routing policy that enhances
throughput by
minimizing the expected number of
transmissions
32
To design
routing
policy that
exhibits good
delay
performance
and
study in a
testbed
.
2
Distributed Adaptive Opportunistic Routing (D-AdaptOR)
Opportunistic routing
Optimality
1Slide33
33
With no link probability information, can we use only “DATA PACKETS” to efficiently route?Adaptive Opportunistic Routing Protocol
D-AdaptOR
S
D
?
Opportunistic setting
(no congestion consideration)
CDP : Explicit probe packets to determine topologySlide34
Opportunistic Routing
Routing decisions are made based on actual transmission outcomes via a three-way handshake.It exploits the broadcast nature of wireless transmissions.
Path diversity is available in many settings.
A
long hop
might be rare but it happens.
S
1
2
D
3
0.8
0.8
0.8
0.8
0.2
Routing Cycle
1
Transmission
(T)
2
Acknowledgment (A)
3
Relaying (R)
Slot n
(T)
(R)
(A)
34Slide35
Contribution
When topology known, there exists an algorithm with optimal performance
EXOR,
GeRaF
, SR
The
knowledge of channel statistics is not known or
erroneous
D-AdaptOR: Distributed Adaptive Opportunistic Routing protocol
D-AdaptOR has zero initial knowledge
Explores new opportunities
Decisions based on current information
Balances exploration with exploitation
Goal : Design
an
algorithm independent of channel knowledge while guaranteeing good performance
35Slide36
36Network Model
Multi-hop ad-hoc network of nodesNodes
connected by unreliable links
Transmission Model:
Node
i
broadcasts a packet
c
i
: fixed cost of transmission
Set of nodes S successfully receive with probability
P(S|i
)
Error-free
Ack
packets
Slotted time:
Transmission and relay selection in one time slot
Only
one
node is selected as forwarder
1
D
2
N
i
2
SSlide37
:
Sequence of relay nodes
:
Time of termination
r
:
If packet is dropped before reaching
destination
then
0
reward, else reward
R
0
R
Termination Reward-Total Cost
A measure of distance
The Setup
Expected Reward
1
2
o
D
37Slide38
Problem (P)
Maximize the expected average per packet reward when the link probabilities are not known
Based on received
acks
, the next decision is to
choose an
ack-ed
neighbor as the next relay,
retransmit,
drop the packet altogether
Reward for packet
m
`
38
Optimality CriterionSlide39
Main Result
The D-AdaptOR achieves optimalityWhen,
Channel statistics are ergodic
Control packets are reliable
Theorem:
D-AdaptOR
maximizes
the average per packet expected reward
D-AdaptOR
maximizes
39Slide40
Routing Cycle
1
Transmission (T)
2
Ack (A)
3
Relaying (R)
4
Adaptive Update (U)
Receiving nodes acknowledge
Receiving nodes feedback their estimated reward
Transmitter or when selecting the relay (randomized)
With high
prob
: Choose the node with best
estimate
With diminishing
prob
: Choose randomly among neighbors
Adaptive
Update
Transmitter updates estimated reward
explores
exploits
AdaptOR
Operation
i
1
2
D
40Slide41
AdaptOR: Summary
41
Developed distributed opportunistic routing algorithm
D-AdaptOR
D-AdaptOR
is optimal (under technical conditions)
D-AdaptOR
outperformed other existing protocols in realistic simulations
D-AdaptOR
achieved the exploration vs. exploitation tradeoff
D-AdaptOR
can be implemented in off the shelf hardware
Relevant Papers
A. Bhorkar, M.
Naghshvar
, T.
Javidi
and B.
Rao
"An Adaptive Opportunistic Routing Scheme for Wireless Ad-hoc Networks ," submitted to
IEEE Trans. on Networking.
A. Bhorkar, M.
Naghshvar
, T.
Javidi
and B.
Rao
”AdaptOR An Adaptive Opportunistic Routing Scheme for Wireless Ad-hoc Networks “, ISIT 09
A. Bhorkar, M.
Naghshvar
, T.
Javidi
and B.
Rao
”Exploration
vs
Exploitation in wireless Ad-hoc networks ,” CDC 09Slide42
Develop learning algorithms for time varying network dynamics
Optimize over exploration and increase convergence rateUse “Regret” as a measure of optimalityImplement D-AdaptOR on test-bed
Future
work
Relevant Papers
A. Bhorkar, T.
Javidi
, A.
Snoeren
, Achieving congestion diversity in wireless Ad-hoc networks, submitted to
Mobicom
10
A. Bhorkar, T.
Javidi
, A.
Snoeren
, Achieving congestion diversity in wireless Ad-hoc networks, in preparation
42
Objective
Timeline
No regret routing
Nearly complete
Asilomar
10