Title: QCRA: Quasistatic Centralized Rate Allocation for Sensor Networks
1QCRA Quasi-static Centralized Rate Allocation
for Sensor Networks
- Fang Bian
- Sumit Rangwala
- Ramesh Govindan
- Computer Science Department
- University of Southern California
Presented by Abhishek Sharma
2Motivation
- Capacity efficiency is important in WSN
- Applications such as structural health monitoring
need to transmit high-volume data in near real
time - Solution Space
- Centralized Rate Allocation
- Distributed Rate Control
Motivation
Design
Evaluation
Conclusions
3Problem Description
- Given a sensor network CSMA MAC layer,
tree-routing, and all nodes have a backlog of
data to send to some base-station. - Centrally find the fair and efficient rate for
each source node.
Having had experience with sensor network
deployments, we didnt think centralized rate
allocation would work
Motivation
Design
Evaluation
Conclusions
4Our Focus and Contributions
- Is near-optimal centralized rate allocation even
feasible in real sensor networks? - Why study feasibility ?
- Resource constraints in sensor networks
- Optimality and Simplicity of centralized rate
control - Our contribution
- Design of a practical centralized rate allocation
heuristic - Extensive experiments on a real wireless test bed
- Comparison with alternative approaches
Motivation
Design
Evaluation
Conclusions
5Design Assumptions
- CSMA MAC layer
- Limited number of link-layer retransmissions
- Tree-routing protocol
- Standard sensor network protocols can be used
- Nodes are continuously sensing and transmitting
data to a base-station - Network conditions dont change significantly on
a time-scale of tens of minutes - Will discuss this assumption more later
Motivation
Design
Evaluation
Conclusions
6Two Parts
- Rate Allocation
- Uses topology and link loss-rate information
- Rate Adaptation
- Adjusts the assigned rates based on network
status.
Similar to prior theoretical work on centralized
rate allocation
Adaptation to CSMA radios
Motivation
Design
Evaluation
Conclusions
7Rate Allocation
- One-time measurement of the raw channel capacity
B - For each epoch (tens of minutes)
- Each node measures the link quality, records
routing tree and neighbor list - Base station collects the measurements
- Centrally calculates the fair and efficient rate
for next epoch
Sink
Neighbor
Child/Parent
2
1
3
4
5
6
7
9
8
11
10
Motivation
Design
Evaluation
Conclusions
8Rate Allocation
- Rate Calculation
- Estimate traffic on each link
- Get a contention list for each node
- Arbitrarily order the potential contention nodes,
cluster them into independent sets - Calculate the aggregate traffic for contention at
each node - Find the most congested node, get the minimum
supported goodput - Assign the rate to each node as goodput over its
effective path quality
Neighbor
Sink
Child/Parent
9
2
2
17
1
8
5
1
3
3
4
5
1, 5 2 3 4 6 7
3 9 1 5 3 1 22
16
8
3
1
1
1
7
9
8
6
10
4
4
6
1
1
11
10
g B / 22
4
4
Motivation
Design
Evaluation
Conclusions
9Rate Adaptation
- In the rate allocation step, QCRA uses link
quality to estimate the expected number of
retransmissions - Two problems
- Over 10s of minutes, link quality may change
- Link quality measurement is done in quiescent
conditions, and does not take collisions into
account - Both can underestimate the actual number of
retransmissions on a link
Motivation
Design
Evaluation
Conclusions
10Rate Adaptation
- Our solution inflate the number of
retransmissions estimate by a factor C, learned
from network behavior during previous epoch - One C for all
- Per-link value C
Focus on simplicity
Measurement-based adaptation approximately
captures link quality variations and collisions
Motivation
Design
Evaluation
Conclusions
11Extensions
- Weighted fairness
- Treat as if each node has been duplicated
according to its weight - Adjust the calculation of the forwarding traffic
accordingly - Multiple base-stations
- Calculate traffic for each tree
- Efficient rate allocation
- Let the nodes get more goodput as long as the
max-min fairness is achieved. - Iterative application of basic rate allocation
Motivation
Design
Evaluation
Conclusions
12Limitations
- Node Failures
- Must re-compute the rate when node failures are
detected - When QCRA may not work
- Link qualities dramatically changes on the order
of minutes - Node fails frequently in short time
Motivation
Design
Evaluation
Conclusions
13Performance Questions
- Are the assigned rates achievable in a real
wireless network? - Is the achieved goodput fair?
- How efficient is the rate allocation?
Motivation
Design
Evaluation
Conclusions
14Evaluation Setup
- Setup
- 40 nodes
- Power 6 Frequency 25
- Method
- Six epochs
- Each epoch is 15 min
- 2 days, 2 nights
- 2 routing trees
Motivation
Design
Evaluation
Conclusions
15Routing Trees
Routing Tree A
Routing Tree B
Motivation
Design
Evaluation
Conclusions
16Performance
Epoch 1
Epoch 1
Night-time Experiment
Day-time Experiment
Epoch 3 to 6 are nearly identical, which shows
that QCRA is able to find a stable network
goodput within two epochs
Motivation
Design
Evaluation
Conclusions
17Fairness
Nodes are assigned rates that results in fair
achieved goodput
Motivation
Design
Evaluation
Conclusions
18QCRA vs. IFRC
- IFRC
- Rangwala et. al, SIGCOMM 2006
- Distributed rate control for sensor networks
- Uses AIMD to adapt the rate with congestion
detection based on queue size - Uses a novel congestion sharing mechanism
- Shared with all potential interferers
- Each node piggybacks on every transmitted packet
Motivation
Design
Evaluation
Conclusions
19QCRA vs. IFRC
- QCRA achieved 50 higher goodput than IFRC
- QCRA is nearly optimal, within 10 of the
upper-bound
Motivation
Design
Evaluation
Conclusions
20Weighted Fairness
Epoch-3
Epoch-4
QCRA can achieve weighted fairness.
Motivation
Design
Evaluation
Conclusions
21QCRA with Dynamic Routing
Day-time Experiment
Night-time Experiment
Surprisingly QCRA performs well even with
dynamic routing.
QCRA Evaluation
22QCRA Conclusion and Open Questions
- Conclusion
- A quasi-static centralized rate allocation is
surprisingly promising in wireless sensor
networks - Open Questions
- Accurate link quality measurement
- Light-weight/low-cost network status measurement
and collection - Low-cost rate distribution
- Fast adaptation in face of node failures.
Motivation
Design
Evaluation
Conclusions
23Thank you very much!
24QCRA Related Work
- Centralized rate control
- Joint Scheduling and Routing
- Jain et.al. (Mobicom 2003), Kodialam et. al.
(Mobicom 2003) - ESRT
- Sankarasubramaniam et. al. (ACM Trans. of
Networks, 2005) - Fair, yet not efficient
- Distributed rate control
- Theoretic max-min fairness rate allocation
- Huang et. al. (MobiHoc 2001)
- AIMD based distributed rate control
- Woo et. al, (Mobicom 2001), IFRC (SIGCOMM 2005)
- Less optimal, require careful parameter tuning
Motivation
Design
Evaluation
Conclusions
25Our Focus
- The feasibility of a centralized rate allocation
- Are previous theoretical studies valid in
assuming that centralized rate allocation makes
sense? - Motivated by
- Resource constraints in sensor networks
- Optimality and Simplicity of centralized rate
control
Motivation
Design
Evaluation
Conclusions
26QCRA Related Work
Network Stack
Application
Transport
Rangwala 06
Woo 01
Routing
Kodialam 03
Jain 03
Efficiency Requirements
Scalability
Energy
Capacity
QCRA Related Work
27QCRA Rate Adaptation
- Problem inaccurate estimation of traffic
- Link quality dynamics
- Inaccuracy in link quality measurement
- Our solution inflate the estimation of
retransmission by some value learned from network
behavior during the previous epoch
QCRA Design
28QCRA vs. Lower Bound Jain 03
- Average achieved goodput over assigned goodput
- Ratio 1.05537 vs. 0.785836
- QCRA more accurately estimates the network
capacity
QCRA Evaluation
29QCRA Per-link C value
Per-Link C Value achieves higher goodput with
slightly worse fairness.
Motivation
Design
Evaluation
Conclusions
30Efficiency In Multi-sink case
SINK-B
SINK-A
QCRA is efficient and fair Achieving higher rate
for other nodes does not affect the nodes with
smaller assigned rates.
Motivation
Design
Evaluation
Conclusions