TAPs: An Architecture and Protocols for a High-Performance Multi-hop Wireless Infrastructure
Ed Knightly
ECE/CS Departments
Rice University
http://www.ece.rice.edu/~knightly
Joint work with
V. Kanodia, A. Sabharwal, and B. Sadeghi
Ed Knightly
The Killer App is the Service
High bandwidth High availability
– Large-scale deployment– High reliability– Nomadicity
Economic viability
Why?– Broadband to the
home and public places
– Enable new applications
Ed Knightly
WiFi Hot Spots?
Why? poor economics– High costs of wired infrastructure ($10k + $500/month)– Pricing: U.S. $3 for 15 minutes; CH: 0.90 CHF/minute– Dismal coverage averaging 0.6 km2 per 50 metro areas
projected by 2005
11 Mb/sec, free spectrum, inexpensive APs/NICs
Carrier’s Backbone/Internet
T1
Medium bandwidth (wire), sparse, and expensive
Ed Knightly
Cellular?
Cellular towers are indeed ubiquitous– Coverage, mobility, …
High bandwidth is elusive– Aggregate bandwidths in Mb/sec range, per-user
bandwidths at dial-up speeds– Expensive: spectral fees and high infrastructure costs
High availability, but slow and expensive
Ed Knightly
Ad Hoc Networks?
Availability– Problems: intermediate nodes can move, power off,
fade, DoS attack routes break, packets are dropped, TCP collapses, …
Low bandwidth– Poor capacity scaling– Unlike cellular, users consume wireless resources at
remote locations
“Free” but low availability and low bandwidth
Ed Knightly
TAPs: Multihop Wireless Infrastructure
Transit Access Points (TAPs) are APs with – beam forming antennas – multiple air interfaces– enhanced MAC/scheduling/routing
protocols Form wireless backbone with limited
wired gateways
Ed Knightly
Multihop Wireless Infrastructure
Transit Access Points (TAPs) are APs with – beam forming antennas – multiple air interfaces– enhanced MAC/scheduling/routing
protocols Form wireless backbone with limited
wired gateways
High bandwidth – High spatial reuse and capacity scaling – Opportunistic protocols
High availability– Redundant paths and non-mobile infrastructure– Deployability
Good economics– Unlicensed spectrum, few wires, exploit WiFi components
Ed Knightly
Prototype and Testbed Deployment
FPGA implementation of enhanced opportunistic, beamforming, multi-channel, QoS MAC
Build prototypes and deploy on Rice campus and nearby neighborhoods
Measurement study from channel conditions to traffic patterns
Ed Knightly
Outline
TAP architecture
OAR: an opportunistic auto-rate MAC
MOAR: multi-channel OAR
Open problems
Ed Knightly
0
5
10
15
20
25
30
35
40
126 51 76
101
126
151
176
201
226
251
276
301
326
351
376
401
426
451
476
Time (sec)
SNR
Received signal: superposition of different reflections, with different delays and attenuations
Motivation
Wireless channel is variable
Coherence time chan
nel
gai
n
time
Ed Knightly
Opportunistic MAC Goal
Constraint: distributed random access protocol
Exploit the variations inherent in wireless channel to increase throughput Maintain fair temporal shares for different flows
chan
nel
gai
n
time
user 1
user 2
user 1
user 2
Ed Knightly
IEEE 802.11 Multi-rate
Support of higher transmission rates in better channel condition– 802.11b available rates: 2, 5.5, 11 Mbps
– 802.11a available rates: up to 54 Mbps
Auto Rate Fallback (ARF)– [Monteban et al. 97]– Use history of previous transmissions to
adaptively select future rates
Ed Knightly
Temporal vs. Throughput Fairness
Equivalent in single-rate networks Throughput fairness results in significant inefficiency
in multi-rate networks
Example
user 1
user 2
access point
user 3
Ed Knightly
Temporal vs. Throughput Fairness
Equivalent in single-rate networks Throughput fairness results in significant inefficiency
in multi-rate networks
Exampleuser 1
user 2access point
user 3
Throughput Fair
user 1
user 2 DATA
DATA
user 3 DATA
Even 1 user with low transmission rate results in
a very low network throughput
Ed Knightly
Temporal vs. Throughput Fairness
Equivalent in single-rate networks Throughput fairness results in significant inefficiency
in multi-rate networks
Exampleuser 1
user 2access point
user 3
Temporal Fair
user 1
user 2 DATA
DATA
user 3 DATA
DATA DATA DATA
DATA DATA DATA
DATA DATA
DATA DATA
Same time-shares of the channel for different flows, also higher throughput
Ed Knightly
Opportunistic MAC
Goal Exploit short-time-scale variations inherent in
wireless channel to increase throughput in wireless ad hoc networks
Issue Maintaining temporal share of each node
Challenge Channel info available only upon transmission
Ed Knightly
Opportunistic Auto Rate (OAR)
Observation: coherence time on order of multiple packet transmission times– measure channel quality on RTS/CTS handshake– hold good channels for multiple transmissions
Ensure fairness by scaling number of packets transmitted to channel quality– # packets = Current rate / Base rate– with random access, all flows equally likely to
access channel
OAR: High throughput, while maintaining temporal fairness properties of single rate IEEE 802.11
Ed Knightly
RTS
RBAR Protocol
source
destination
ACKCTS
Receiver Based AutoRate (RBAR)
DATA
Receiver controls the sender’s transmission rate
Control messages sent at Base Rate
Reservation Sub-Header
Ed Knightly
DATA DATARTSACKCTSACK
OAR Protocol
source
destination
Pkts Rate Pkts Rate Pkts Rate
802.11 1 2 1 2 1 2
802.11b 1 2 1 5.5 1 11
OAR 1 2 3 5.5 5 11
Protocol
Channel Condition
BAD MEDIUM GOOD
OAR
Once access granted, it is possible to send multiple packets if the channel is good
Reservation Sub-Header
Ed Knightly
Observation I Time spent in contention per packet is the same for
RBAR and single-rate IEEE802.11Transmitter
Receiver
OAR
Performance Comparison
IEEE 802.11
R
C A
D1Transmitter
Receiver
R
C A
D1 R
C A
D1Transmitter
Receiver
RBARObservation II OAR contends for the same total time as singe-rate
IEEE 802.11 but transmits more data
R
C A
D1
A
D2
A
D3
Ed Knightly
Transmitter
Receiver
OAR
Performance Comparison
IEEE 802.11
R
C A
D1Transmitter
Receiver
R
C A
D1 R
C A
D1Transmitter
Receiver
RBAR
Observation III OAR holds high-quality channels for multiple
transmissions
R
C A
D1
A
D2
A
D3
Ed Knightly
Analytical Model
Challenge: MAC and channel are random processes with memory
Model relates physical-layer characteristics to MAC throughput:– Time spent in contention
Markov model of modified IEEE 802.11– Average transmission rate
Due to channel distribution
Comparative model b/t multi-rate OAR and tractable systems– TIME: OAR contends as often as single-rate IEEE 802.11 with
increased data per contention– PACKETS: OAR reduces packet transmission time via per-
contention rate adaptation
Ed Knightly
Simulation Results Under Ricean Fading
OAR has 42% to 56% gain over RBAR Increase in gain as number of flows increases Model predicts OAR & RBAR throughput to within 7% accuracy
Nodes
ProtocolPackets Transmitted
with 11 Mbps (%)
OAR 80%
RBAR 65%
Ed Knightly
Outline
TAP architecture
OAR: an opportunistic auto-rate MAC
MOAR: multi-channel OAR
Open problems
Ed Knightly
– Example: at 2.4 GHz WiFi, 5 vs. 1-3 MHz
– Figure for Ricean, K=4
Multi-Channel Problem Formulation
Observe: for two MUs, quality of different channels can have low correlation if
channel separation >> coherence bandwidth
Ed Knightly
Challenge
Ideal protocol is simple: select the best channel at the instant of transmission
In practice, channel qualities are unknown a priori– Must first transmit and measure
Cost of measuring channels must be balanced with benefits of finding good ones
Ed Knightly
MOAR Protocol Sketch
Measure channel SNR at RTS/CTS handshake
If channel quality is high (above an SNR threshold), transmit via OAR
If channel quality is poor, skip to a new channel– next channel piggybacked in CTS
Design optimal stopping rule for skipping– stop when throughput gain of skipping to a better
channel is outweighed by overhead
Ensure fairness
Ed Knightly
Optimal Stopping Rule Formulation
Let Xn denote the SNR of the nth measured channel
Let c denote the cost (in time) of measuring the channel
After observing Xn transmit or measure again?– cannot go back to previous channel (coherence time)
The reward for the nth selection is Xn-nc – after scaling SNR to rate and then to time
Objective: maximize the expected reward
In a class of stopping rule problems (without recall)
Ed Knightly
Optimal Stopping Time
Let V* denote the expected return from the optimal stopping rule
Suppose pay c and observe X1= x1
If continue, x1 is lost and c is paid– continuing, can obtain return V*, but not more– start afresh
Optimal rule is threshold based– If xn < V*, continue; if xn > V* stop– N* = min{n 1: Xn V*}
Ed Knightly
Calculating the Stopping Threshold V*
V* = E max(X1,V*) – c
– F(x) represents the SNR distribution
Compute V* – channel model and parameters (ex. K, d)– system’s rate-SNR thresholds (ex. 1, 2, 5.5, 11)
cxdFVxV
*)(*)(
Ed Knightly
MOAR Throughput Gains
Ricean parameter K = 0 is no line-of-sight signal
Gains of 40%-60% increasing with K and SNR variance
Ed Knightly
Effect of Node Distance
Greatest help when far away Non-monotonic due to rate-SNR thresholds
Ed Knightly
Random Topologies
Nodes are uniform-randomly placed in a 250m circle
“Optimal Skipping” cheats: looks at all channels (with no cost) and jumps to the best
Observe– MOAR extracts most
available gain– close-by nodes detract
from average gain
Ed Knightly
Outline
TAP architecture
OAR: an opportunistic auto-rate MAC
MOAR: multi-channel OAR
Open problems
Ed Knightly
DoS Resilience and Security
Old methodology– Design a network protocol– Optimize for performance– Discover DoS/Security holes
Ex. Route query floods
– Patch one-by-one
Challenge– DoS-resilience and security as the foundation of
network protocols– Recognize these issues are as important as
performance
Ed Knightly
TAP Media Access and Scheduling
Challenge: distributed scheduling– Others’ channel states, priority, & backlog condition unknown
Ex. TAP A’s best recv’r may be transmitting elsewhere Ex. Traffic to be recv’d may be higher priority than that to be sent
– Traffic and system dynamics preclude scheduled cycles– Modulate aggressiveness according to overheard information
Ed Knightly
Multi-Destination Routing/Scheduling
Most data sources or sinks at a wire Routing protocols for any wire abstraction Scheduling
– At fast time scales, which path is best (channels, contention, …) now?
– Can delay/throughput gains be realized despite TCP?
Ed Knightly
Distributed Traffic Control
Distributed resource management: how to throttle flows to their system-wide fair rate?– Throttle traffic “near-the-wire” to ensure fairness
and high spatial reuse– TCP cannot achieve it (too slow and RTT biased)– Incorporate channel conditions as well as traffic
demands
Ed Knightly
Capacity Driven Protocol DesignProtocol Driven Capacity Analysis
Traditional view of network capacity assumes zero protocol overhead (no routing overhead, contention, etc.)
Protocols themselves require capacity A new holistic system view: “the network is the channel”
– Incorporate overhead in discovering/measuring the resource– Explore capacity limits under real-world protocols
Ed Knightly
Problem: Multiple APs/TAPs/…within Radio Range
PHY Interference has disproportionate throughput degradation at MAC layer
Interference can lead to severe scaling limitations and starvation (worse than zero-sum game)
Ed Knightly
Summary
Transit Access Points– WiFi “footprint” is dismal– Removing wires is the key for economic viability
Opportunistic Scheduling (OAR/MOAR)– Exploit time and frequency diversity
Challenges– Multi-hop wireless architectures– Distributed control– Scalable protocols