structure-free data aggregation
DESCRIPTION
Structure-free Data Aggregation. Kaiwei Fan, Sha Liu, and Prasun Sinha (speaker) The Ohio State University Dept of Computer Science and Engineering. Outline. Introduction Structure-free Data Aggregation Simulation Results Experiments on a testbed Conclusion. Introduction. Data Aggregation - PowerPoint PPT PresentationTRANSCRIPT
Structure-freeData Aggregation
Kaiwei Fan, Sha Liu, and Prasun Sinha (speaker)The Ohio State UniversityDept of Computer Science and Engineering
Outline
Introduction Structure-free Data Aggregation Simulation Results Experiments on a testbed Conclusion
Introduction
Data Aggregation In-network processing Reduces communication cost
Approaches Static Structure
[LEACH, TWC ’02] [PEGASIS, TPDS ’02]
Dynamic Structure [Directed Diffusion, Mobicom ‘00] [DCTC, Infocom ‘04]
Static Structure
Pros Low maintenance cost Good for unchanging
traffic pattern Cons
Unsuitable for event triggered network Long link-stretch Long delay sink
Static Structure
Pros Low maintenance cost Good for unchanging
traffic pattern Cons
Unsuitable for event triggered network Long link-stretch Long delay sink
Dynamic Structure
Pros Reduces communication
cost Cons
High maintenance overhead
sink
Structure-free Data Aggregation
Challenge Routing: who is the next hop? Waiting: who should wait for
whom? Approach
Spatial Convergence Temporal Convergence
Solution Data Aware Anycast Randomized Delay
Routing?
Waiting?
sink
Data Aware Anycast
Improve Spatial Convergence Anycast
One-to-Any forwarding scheme Anycast for Immediate Aggregation
To neighbor nodes having packets for aggregation
Keep Anycasting for Immediate Aggregation
sink
Data Aware Anycast
50 nodes in 200mx200m
sink
Data Aware Anycast
Forward to Sink To neighbor nodes closer to the sink Using Anycast for possible Immediate
Aggregation
sink
Data Aware Anycast
Forwarding and CTS replying priority Class A: Nodes for Immediate Aggregation Class B: Nodes closer to the sink Class C: Otherwise, do not reply
Class B
Canceled CTS
Canceled CTS
RTS
CTS
Sender
Class A Nbr
Class B Nbr
Class C Nbr
Class A Nbr
CTS slotmini-slot
Class A
Randomized Waiting
Improve Temporal Convergence Naive Waiting Approach
Use delay based on proximity to sink (closer to sink => higher delay)
Long delay for nodes close to the sink in case the event is near the sink
Our Approach: Random Delay at Sources
Analysis Y: Number of hops a packet is forwarded before being
aggregated Assumptions:
Each node has k choices for next hops closer to sink All n nodes have packets to send
E[Y] = x : random delay in [0,1] picked up by a node dh :random delay chosen by a node h hops away from sink
Total Number of Transmissions =
dxxdYE h )]|([1
0
… …Sink
h=n/k
kn
h
h
ik nn
k
n
k
nHnYEk
/
1
1
0
log)()1(][
Analysis vs. Simulation
Results matches up to 40 hops
Gap increases as network size increases
Reason: transmission delay is ignored in analysis
Simulation Results
Evaluated Protocols Opportunistic (OP) Optimum Aggregation
Tree (AT) Data Aware Anycast
(DAA) Randomized Waiting (RW) DAA+RW
Evaluated Metric Normalized Number of
Transmissions
Parameters Studied Maximum Delay Event Size Aggregation Function Network Size
nInformatioReceivedofUnits
onsTransmissiTotalofNumber
Simulation Results – Maximum delay
Configuration 33 x 33 grid network event moves at 10m/s event radius: 200m 140 nodes triggered by t
he event data rate: 0.2 pkt/s data payload: 50 bytes
AT-2: Aggregation tree approach with varying delay
DAA+RW improve OP by 70%
Simulation Results – Maximum delay
AT is sensitive to delay AT has best performance
with highest delay
Simulation Results – Event Size
Configuration event radius: 50m ~
300m 8 ~ 260 nodes
triggered by the event event radius: 200m
Key Observations DAA+RW is much
better than OP DAA+RW is close
to AT (optimal tree)
Simulation Results – Aggregation Ratio
Configuration Aggregation Ratio ρ:
0 ~ 1 Packet size:
max(50, 50* (1-ρ)* n) Max packet size:
400 bytes
Key Observation DAA+RW performs be
tter than AT Following the best tre
e is not optimum if the packet size is limited
Simulation Results – Network Size
event distance to the sink: 300m ~ 700m
event radius: 200m
Key Observation Improvement is higher
for events farther from the sink
Experiment – Randomized Waiting
Linear network with 5 sources and 1 sink
0.2 pkt/s data payload: 29 bytes
Key Observation Delay as low as 0.1 is suff
icient for optimizing performance
Conclusion
Data Aware Anycast for Spatial Convergence Randomized Waiting for Temporal Convergence Efficient Aggregation without a Structure
High Aggregation No maintenance overhead