CMAP: Harnessing Exposed
Terminalsin Wireless Networks
Mythili Vutukuru
Joint work with Kyle Jamieson and Hari Balakrishnan
2
The Problemu
v
x
y
Which transmissions concurrently? Increase throughput by maximizing
concurrency.
Xz
3
u
v yz
Exposed terminal problem.
Today’s Solution: CSMA
x
Energy > carrier sense threshold
4
Key Insight Existing solutions: rules to predict which
concurrent transmissions increase throughput.
Instead, watch and discover which concurrent transmissions increase throughput.
5
CMAP: Conflict Map Map of conflicting transmissions based on
empirical evidence. Built in distributed, online manner.
Exposed terminals – 2x gain on CSMA.
u
v yz
x u v & x y NO!
6
Roadmap CMAP Design
• Conflict maps
• ACK & Backoff Policy
Implementation
Evaluation
7
What is a “conflict”?u
v y
x
Conflict – throughput lower when concurrent. Loss rate threshold to decide conflicts.
No Conflict.Conflict.
50% loss
8
Discovering Conflictsu
v y
x
Loss rate of uv when x is concurrent
>50% then infer conflict at v. Conflict entries timed out periodically.
When u transmits to me, x causes interference.
pq
9
Populating the Conflict Map
u
v y
x
When u transmits to me, x causes interference.
Do not transmit to v when x anyone.
Do not transmit to anyone when u v.
Conflict map
z
10
Channel Access Decisions Nodes always overhear channel.
Consult conflict map before transmission.
Carrier sense always disabled!
11
Roadmap CMAP Design
• Conflict maps
• ACK & Backoff Policy
Implementation
Evaluation
12
Windowed ACKs
Sliding window of packets at sender.
u
yz
xX
13
Backoff Policy
Cannot defer when hidden terminals. Exponential backoff. When loss rate in ACKs > threshold.
u
v y
xDo not transmit to v when x anyone.
u must hear x.
14
Roadmap CMAP Design
• Conflict maps
• ACK & Backoff Policy
Implementation
Evaluation
15
Implementation Challenges1. At receiver: Identify colliding senders.
2. At sender: Identify ongoing transmissions.
PHYMAC
Trailer
16
Implementation Options Software radios – Partial Packet Recovery.
[Jamieson and Balakrishnan, SIGCOMM 2007]
Commodity hardware – separate header and trailer packets.
Header pkt Trailer pkt
17
Prototype Implementation
MadWifi DriverAtheros 802.11 card
Conflict Maps ACKs & Backoff
CSMA, ACKs & Backoff disabled.
Click Kernel Module
PHYMAC
18
Roadmap CMAP Design
• Conflict maps
• ACK & Backoff Policy
Implementation
Evaluation
19
Evaluation
50-node 802.11a indoor testbed. Does CMAP improve throughput by
increasing concurrency?
20
Two Senders In Range Senders in range. 1400-byte UDP @ 6 Mbps. 50 unique sets of four nodes. CMAP, CSMA, no CS no acks.
Exposed terminals. Interfering transmissions.
21
Two Senders In Range
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10
Aggregate Throughput (Mbits/s)
CD
F
CSMA No CS, No Acks
CSMA better.
No CS better.
22
Two Senders In Range
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10
Aggregate Throughput (Mbits/s)
CD
F
CSMA No CS, No Acks Ideal
Ideal is max of CSMA & No CS
23
Two Senders In Range
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10
Aggregate Throughput (Mbits/s)
CD
F
Ideal CMAP
CMAP traces ideal curve.
24
Multiple Concurrent Senders
AP-client networks. Tree-based mesh networks.
25
Multiple Concurrent Senders
0
2
4
6
8
10
12
14
16
18
3 4 5 6
Number of concurrent senders
Agg
rega
te th
roug
hput
(Mbi
ts/s
)
CSMACMAP
AP-client – 20-47% better over CSMA.
Mesh – 52% better over CSMA.
26
Hidden terminals – backoff ensures CMAP similar to CSMA.
CMAP without windowed ACKs gets only half the gains → windowed ACKs useful.
CMAP’s gains hold across multiple bit-rates.
More results in the paper
27
Related Work
RTS but no CTS → exposed. [Karn, Shukla et al.]
Offline training to identify exposed terminals. [Mittal and Belding]
u
yz
xRTS
CTS
X
28
Limitations Losses until conflict map entries populated.
Unequal pkt sizes → longer to detect
conflicts.
Cannot detect conflicts when interfering
node’s headers cannot be decoded.
29
Contributions MAC to improve throughput by increasing
concurrency. Key idea: watch and discover conflicts. Experiments show increased throughput.
• 2x improvement over CSMA with exposed terminals.
• ~50% improvement in AP and mesh networks.