TCP for Mobile and Wireless Hosts

Improving Performance of
Wireless Networks
Nitin Vaidya
Joint work with Fan Wu, Tae Hyun Kim, Jian Ni,
Vijay Raman, R. Srikant
November 4, 2010
1
What Makes Wireless Networks
Interesting?
Many forms of diversity
• Time
• Route
• Antenna
• Spatial
• Channel
2
Multi-Channel Environments
Available spectrum
Spectrum divided into channels
1
2
3
4
…
c
3
Multi-Channel Wireless Networks
Benefits of channelization

Channel diversity

Channel access efficiency gain
• Gain variations
• Interference mitigation
4
Recent Contributions on
Multi-Channel Networks

Incorporating opportunism in multi-channel
networks

Improving channel utilization

Game theoretic approach for channel management
5
Opportunistic Routing
Opportunism


Traditional routing: S  R  D
But D may sometimes overheard S  R
transmission
S

R
D
No need to forward such packets on R  D
7
Opportunism using MORE

Source sends linear combinations of packets in batches

Forwarders keep all heard packets in a buffer

Nodes transmit linear combinations of buffered packets

Destination decodes once it receives enough combinations
S
R
D
P1
2,1,3
2,1,3
P1
P2
0,2,1
7,4,9
P2
P3
3,0,2
1,6,6
P3
2
0 P1 ++2
102b P2
2c+ 2
== a,b,c
-+ 1 3,0,2 = 7,4,9
1 P30,2,1
a3
2,1,3++13
3,0,2
2,1,3
0,2,1
1,6,6
P2
Opportunism versus Concurrency


For opportunistic scheme to work,
nodes must be on the same channel
Reduces concurrency
S
R
D
9
Trade-Off
Advantages
Exploits broadcast nature
Opportunism Reduces average # hops
Fewer transmissions
Concurrency
Multichannel Lower contention
Disadvantages
Higher contention
No multiple channel support
No opportunistic overhearing
Potentially longer routes
10
Example
Traditional Channel
Assignment
C1 C2
A
C1
S
Loss probability
C2
0.9
0.5

C3
D
C3
B
C3
End-to-end throughput = 0.5
“Opportunism-Aware” Channel Assignment
C1 C2
A
S
C1
0.9
0.5
C1
C2
D
C2
B
C1 C2
End-to-end throughput = 0.6475
Our Contribution


Take into account both opportunistic gains
obtained by assigning identical channels to the
nodes, as well as concurrency gains by assigning
different channels
Extended MORE to a multi-radio multi-channel
(MRMC) environment
13
Summary

Opportunistic schemes can benefit in multi-channel
environments

Channel assignment needs to be opportunism-aware

Proposed such an assignment scheme
14
Packet Size-Dependent Channel Selection
15
Channel Width

Typically, channels are assumed identical width
1

2
3
4
…
c
May benefit by varying channel widths
16
Motivation
Rate-independent MAC overhead
L1 /R
DIFS
Header
DIFS
Header
L1 bits
L2 bits
L2 /R
T
T
Overhead 
T  ( Li / R)
17
MAC Overhead vs Packet Size
T = 50μs; R = 54 Mbps
T
Overhead 
T  ( Li / R)
Packet size Li
18
Current Approach

Frame Aggregation (used in IEEE 802.11n)
 Aggregate and send multiple packets in a single
transmission opportunity
DIFS
Header
overhead
L1 bits
L2 bits
L3 bits
Multiple packets to amortize overhead
19
Packet Size-Dependent Channel Widths

Partition a channel into narrow and wide sub-channels

Use narrow sub-channel for short packets

Use wide sub-channel for long packets
20
Proof-of-Concept

Consider a node (A) communicating with
multiple other nodes
A
21
Proposed Approach
2
Node A estimates aggregate short packet
load
Node A determines
3 partition {BW , BW }
S
L
Clients use BWS for short
4
& BWL for long packets
Clients estimate own
1 short packet load,
and inform node A
22
Summary

Channel width selection based on packet size
distribution

Can perform better than frame aggregation

Ideas can be extended to arbitrary networks
23
CSMA with Imperfect Carrier Sensing
Carrier Sensing (CS)



Not perfect
With narrower channels and mobility,
fading can be an issue
What happens to network performance when
CS is imperfect ?
25
Throughput-Optimal Schedulers


A scheduler is throughput-optimal if
it can serve all schedulable traffic
Throughput-optimal scheduler by
Tassiulas-Ephremides’92
• Schedule =
• Computationally complex and centralized solution
Related Work


Continuous-time CSMA-like algorithm by
Jiang-Walrand’08
Discrete-time CSMA by Ni-Srikant’09
Our Contribution:
Preemptive CSMA




Discrete-time medium access
Per-packet scheduling decision
Data packet collisions modeled
Non-zero carrier sense time
Analysis for


Perfect carrier sensing
Imperfect carrier sensing
Model

Link-centric model
Transmission rate is normalized to 1
One-hop traffic

Conflict graph to model interference
Medium Access Model
Last α-duration of each time slot for carrier sense
Preemptive CSMA
Carrier sense
u(t): preemption
x(t): transmission schedule
Ci: set of conflict links of i
ACK reception

Two access probabilities: ai and pi
Performance Analysis

Schedule evolution: discrete-time reversible
Markov chain
Stationary distribution
Cu : set of conflicting links of links in u
When p
i
1
=1exp{wi(qi)}
=exp{wi(qi)} -1
exp{wi(qi)}
Throughput-Optimality

Preemptive CSMA is throughput-optimal
When access probabilities are
• 0 < a ≤ ai ≤ a < 1
• p = 1 - 1/exp{w (q )}
LB
i
UB
i
i
where wi is a strict concave function with wi(0) = 0
Proof relies on time-scale separation
• At each time slot, the Markov chain in the steady
state
Carrier Sense Failure

i.i.d. failure events over time slots and links

Two types of carrier sense failures
• False positive
– No activity, but busy state sensed
– False positive with probability η
• False negative:
– Activity, but idle state sensed
– False negative with probability γ
Carrier Sense Failure:
Main Result

By choosing small enough access probability,
possible to stabilize arbitrarily large fraction of
capacity region
Proof complexity:
Markov chain is no longer reversible
Use perturbation theory for Markov chains
Summary
Preemptive CSMA
Good
performance achievable despite imperfect
carrier sensing
Small
access probability overcomes the effect of
carrier sensing failures
Where are we now ?
37
What Makes Wireless Networks
Interesting?
Many forms of diversity
• Time
• Route
• Antenna
• Spatial
• Channel
38
Wireless Diversity


This project has furthered our understanding of
approaches to wireless diversity using suitable
protocols
We now have a better understanding of
cross-layer protocol design
39
What Remains?

Physical layer community has also been making
significant progress
– Interference alignment
– Cooperation
– Security

Need to incorporate these ideas into
the protocol stack
Natural continuation
40
of DAWN MURI
What Remains?
Distributed
Applications
Unicast
Multicast
Higher
Layers
Physical
Layer
41
What Remains?
Much attention to

Moving bits between
nodes in the network
• throughput
• delay, jitter
• packet loss

Distributed
Applications
Unicast
Multicast
Higher
Layers
Physical
Layer
Cross layer ~ Layers 1-2-3
42
What Remains?


Not as much attention to
semantics of
distributed applications
How to exploit
application-awareness ?
Distributed
Applications
Unicast
Multicast
Higher
Layers
Physical
Layer
43
Wireless Network-Aware
Distributed Primitives
Distributed
Applications
Distributed
Primitives
Unicast
Multicast
Higher
Layers
Physical
Layer
44
Wireless Network-Aware
Distributed Primitives
Example primitives:
Ordered group communication
Consensus
Aggregation
Synchronization
Coordination
Distributed
Applications
Distributed
Primitives
Unicast
Multicast
Higher
Layers
Physical
Layer
45
Wireless Network-Aware
Distributed Primitives
Example primitives:
Ordered group communication
Consensus
Aggregation
Synchronization
Coordination
Network-awareness
Wireless capacity region
Diversity
Broadcast capability
Energy constraints
Distributed
Applications
Distributed
Primitives
Unicast
Multicast
Higher
Layers
Physical
Layer
46
Past Work on Middleware


Similar motivation
But optimized for wired networks
with high capacity
and more benign characteristics
47
Wireless Network-Aware
Distributed Primitives


Wired algorithms not efficient
Do not exploit wireless capabilities
Many (new) fundamental problems open
48
Distributed Algorithms & Networking

Overlapping scope

But cultures differ
Communications /
Networking
Distributed
Algorithms
49
Communications /
Networking
Emphasis on “exact”
performance metrics
Constants matter
Distributed
Algorithms
Black box networks
Emphasis on
order complexity
50
Communications /
Networking
Emphasis on “exact”
performance metrics
Constants matter
Distributed
Algorithms
Black box networks
Emphasis on
order complexity
Information transfer
(typically “raw” info)
51
Communications /
Networking
Emphasis on “exact”
performance metrics
Constants matter
Information transfer
(typically “raw” info)
Distributed
Algorithms
Black box networks
Emphasis on
order complexity
Computation
affects communication
52
Wireless Network-Aware
Distributed Primitives
Beneficial to bring together researchers in
wireless networking & distributed algorithms
Picture from Wikipedia
Thanks!
54
Thanks!
55
Thanks!
56
Thanks!
57
Scheduling Example
Acces
s by aA
Acces
s by aB
Conflict
graph for
links A, B, C
Acces
s by pB
DATA
DATA
DATA
Acces
s by aB
Sensed
busy by
Link A & C
DATA
ACK
C
PROB
E
ACK
PROB
E
PROBE
ACK
PROB
E
A
B
Preempted
by Link B
PROB
E
Sensed
idle by
Link A & C
Preempted
by Link A &
C