Sift: A MAC Protocol for Event-Driven Wireless Sensor
NetworksKyle Jamieson†, Hari Balakrishnan†, Y.C. Tay‡
† MIT Computer Science and Artificial Intelligence Laboratory‡ Dept. of Computer Science, National University of Singapore
Types of Traffic in Sensor Networks
• Periodic traffic– Animal habitat
monitoring– Indoor
environment• Temperature• Room occupancy
– Medical monitoring
• Patient vital signs
• Event-driven traffic– Failure of
mechanical structures
• Water pipes• Airplane wings
– Medical emergencies
– Vehicle tracking
Airplane Wing Example
For critical systems, low latency is important!
Sift
• Focus of our work– Designing MAC protocol to handle
event-driven workload
• Challenges– Low-latency– Good throughput– Good fairness
Problems for Traditional MAC
1. Spatially-correlated contention: correlation between geographical neighbors’ traffic.
2. Bursty traffic: the number of senders can quickly change.
3. Suppression (counter-intuitively)
Suppression: often, not all sensing nodes need to report an event.
The Status Quo: CSMA
Time
Busy Medium
MAC Goal: only one node transmit at a time
• Basis of existing sensornet MAC layers– B-MAC, S-MAC
• Timeslot: opportunity for a node to begin transmitting
• Process repeats after each packet
The Status Quo: CSMA
• Pick a timeslot chosen uniformly in [0, CW]• Listen up to chosen slot
– Transmit if nobody else started transmitting– Wait if somebody else started transmitting
Time
Example: A Successful Transmission
• A and B happened to choose different slots– Node A chooses slot 4, hears nothing, transmits– Node B chooses slot 8, hears Node A, waits
Success: exactly one node in first non-vacant slot
Node A:
Node B:
Slot choice (slot #4)
Slot choice (slot #8)
Time
Example: A Collision
• A and B happened to choose slot 4– Both listen and hear nothing– Both transmit simultaneously
Collision: ≥ 2 nodes in first non-vacant slot
Node A:
Node B:
Slot choice (slot #4)
Slot choice (slot #4)
Time
High Contention Causes Collisions in CSMA
Uniform distribution “fills up,” quickly
Numerical simulation
Unacceptable collision rate above ~15 transmitting sensors
Solving the Problem of Collisions in CSMA
1. Create more slots– Conventional approach– Called “binary exponential backoff”
(BEB)
2. Change the way we pick slots– Sift takes this approach
Create More Slots:Binary Exponential Backoff (BEB)
• The basis for Ethernet, B-MAC, S-MAC, 802.11, MACAW, many other MAC layers
Acknowledgement?
Reduce CWDouble CWand resend
Yes No
Problems with BEB
• Takes time for every node to increase CW– Especially if traffic is spatially-correlated
and bursty
• Waste backoff slots if collisions cause CW to increase– Especially with suppression
BEB causes performance to suffer
Our Proposal: Sift
• Sift is a MAC protocol for sensor networks– Event-driven traffic– Low-latency requirements
• Sift’s Properties– Extremely simple– Offers up to 7-fold lower latency– Maintains good channel utilization
(throughput)
Sift: Changing the Distribution
• Keep number of slots the same (simple)
• Use an increasing non-uniform slot selection probability distribution
– Make collisions unlikely for large range of N
1. Reduce the chance of collisions• Penalty: one packet- or RTS-time (ms)
2. Reduce wastage of backoff slots• Penalty: one slot time (μs)
Balls and Bins Analogy
• Bin represents a backoff slot in the contention window– Bin height represents probability of picking
that slot
• Ball represents a single node’s slot choice
A
Bins represent backoff slots →
Why an Increasing Slot-Selection Function?
Bins represent backoff slots →
Nod
es c
hoos
ing
each
slo
t →
Sift’s Slot Selection Distribution
rCW
CW
rp
1
)1(
Optimal Non-Persistent CSMA Performance
With knowledge of number of nodes (IEEE J-SAC ’04)
Numerical simulation
Sift Approaches Optimal
Sift needs no knowledge of the number of nodes
Numerical simulation
Sift keeps success rate above this unacceptable range
Experimental Setup
• Simulation-based results (ns-2)• Compare 802.11 (BEB), Sift,
and 802.11/copy– 802.11/copy: send CW in each
packet, copy overheard CW
Event-driven Traffic Pattern
• Event-based traffic pattern– Single-hop to one base
station– N nodes sense and report an
event– R ≤ N reports are required
• If a node hears ≥ R reports then it suppresses its own event report E.g. N=4, R=3
BaseStation
Sift Outperforms When N is Large
Experimental evaluation: R=1,16
R=16
R=1
Sift Outperforms as R Increases
Experimental evaluation: N=128
Exploring Sift’s Performance Space
Experimental evaluation
Hidden Terminal Experiment Setup
• Separate 128 sensors into mutually-hidden clusters– Nodes in one cluster cannot hear nodes in another
• All nodes send to the base station– Result: hidden
terminal collisions at the base station
Base Station
Sift Performs Well with Hidden Terminals
Experimental evaluation: N=128, R=1
Sift Resilient to Jitter in Event Time
Experimental evaluation: N=128, R=64
Sift Improves Fairness
Eight nodes 64 nodes
Experimental evaluation
Trace-Driven Experimental Setup
• Simulated vehicle tracking
• Captured live video from a street scene– Extract motion events
from image analysis
• Event trace drives ns-2 simulation– 128 sensors laid out in a
grid over the scene– Sensors nearby each
event send traffic in response to movement
Sift Outperforms When R is Large
Trace-driven experimental evaluation
Related Work
• TDMA suffers in terms of latency– PTD (Mowafi et al.), TSMA (Chlamtac et al.)
• BEB-based protocols waste time in backoff– MACAW (Bharghavan et al.), S-MAC (Ye et
al.), FAMA (Garcia-Luna-Aceves et al.)• The HIPERLAN standard for wireless LANs
uses noise bursts of exponentially-distributed length
• Periodic-sleeping and other MAC protocols can work with Sift– S-MAC (Ye et al.), B-MAC (Polastre)
Sift is a composable MAC primitive
Conclusion
• Sift is a latency- (and sometimes throughput-) enhancing MAC for event-driven sensor networks
• Sift can be used as a building block in many MAC protocols
http://nms.csail.mit.edu/projects/sift