1. ECE 256: Wireless Networking and Mobile Computing
Diagnosing Wireless Packet Losses in : Separating Collision from Weak Signal
Presented By:
Jacob H. Cox Jr
For
ECE 256: Wireless Networking and Mobile Computing
February 10, 2009

2. Acknowledgments
Authors ~ Shravan Rayanchu, Arunesh Mishra, Dheeraj Agrawal, Sharad Saha, Suman Banerjee
Kuo-Chung Wang (Slide Presentation)

3. Presentation Outline Packet Loss Problem Current Rate Adaption Schemes
COLLIE Overview
COLLIE Metrics
COLLIE Analysis
Conclusion

4. Motivation Packet Loss 802.11 Solution Inadequate Question:
2 Causes: Weak Signal and Collision
Solution Inadequate
defaults to BEB for a substantial number of packet losses
Question:
Does the type of packet loss matter?
What if we could determine its cause?

5. Problem Defined Collision or Weak Signal, why does knowing matter?
Beamforming?
Depending on the reason for packet loss, the data-rate/power
adaptation may decide to backoff, adjust its rate, or make some other modification.

6. Fixing packet loss Appropriate actions For collision BEB CW Max
Binary-Exponential Backoff (BEB)
Retries
REF:

7. Rate Adaptation
a/b/g standards allow for the use of multiple transmission rates
802.11a, 8 rate options (6,9,12,18,24,36,48,54 Mbps)
802.11b, 4 rate options (1,2,5.5,11Mbps)
802.11g, 12 rate options (11a set + 11b set)
Some papers report that rate adaptation is important yet unspecified in standards
I assume that unspecified means that RA still has some flexibility with how it can be implemented.
Reference: Robust Rate Adaptation in Networks Presentation by Kuo-Chung Wang

8. Rate Adaptation Example
Rate adaptation affects throughput performance and should be adjusted by channel condition
54Mbps
Signal is good
12Mbps
Receiver
Signal becomes weaker
Sender
Rate Too High
Rate Too Low
Increases Loss Ratio
Capacity Under-Utilized
Decreased Throughput
Most designs follow a few conceptually intuitive and seemingly effective guidelines
Decrease rate upon severe loss
Use deterministic success/loss patterns
Use probe packets
Use PHY-layer metrics
Use long-term statistics
Reference: Robust Rate Adaptation in Networks Presentation by Kuo-Chung Wang

9. Related Work Rate Adaptation Algorithms ARF ~ Auto-rate Fallback
–Differentiate between loss behaviors
–Adapt to realistic scenarios
–Handle hidden stations
ARF ~ Auto-rate Fallback
CARA ~ Collision-Aware Rate Adaptation
MRD ~ Multi-Radio Diversity
RBAR ~ Receiver Based Auto Rate
RRAA ~ Robust Rate Adaptation Algorithm
Most designs follow a few conceptually intuitive and seemingly effective guidelines
Decrease rate upon severe loss
Use deterministic success/loss patterns
Use probe packets
Use PHY-layer metrics
Use long-term statistics

10. RAA Problem
Sender
12Mbps
Signal is still good
54Mbps
Signal is good
Sender
12 Mbps
Sender
12Mbps
Signal is still good
Receiver
In the presence of hidden stations, a receiver may experience significant packet losses.
This subsequently triggers rate adaptation at the sender to decrease its rate according to the stated guideline.
However, the sender should not decrease its transmission rate because reducing the rate prolongs the transmission
time for each packet, which worsens channel collisions and further reduces the rate.
The fundamental problem is that rate adaptation may ex-
perience much richer set of packet loss scenarios in practice,
which are well beyond the simplistic one of only fading/path
loss envisioned by the original designs. The guideline of de-
creasing rate upon severe packet loss does not apply in other
lossy scenarios. The rate adaptation solution has to differ-
entiate various losses and react accordingly.
Sender
54Mbps
Signal is good
With hidden terminals, reducing the rate prolongs transmission time for each packet and results in more collisions
Sender
54Mbps
Signal is good

11. Introduction to COLLIE
802.11, CARA, and RRAA use multiple attempts to deduce cause of packet loss
COLLIE uses a direct approach
Error packet kickback
Client analysis

12. COLLIE Collision Inferencing Engine Utilizes receiver feedback
Analyzes:
Bit and symbol level error patterns
Received signal strength
Design:
Signal analysis algorithms
Link layer protocol which adjusts link layer parameters
Claim to obtain “significant throughput and capacity improvements for high mobility usage scenarios.”

13. Link Adaptation Mechanism Enhancements
Auto Rate Fallback (ARF)
Used in conjunction w/COLLIE for this paper
Rate adaption mechanism enhanced with inferencing component
Using COLLIE, observed throughput gains of 20-60%
Gains based on channel conditions and level of contention

14. ? COLLIE Continued X Collision Inference Algorithm Client AP Data
Feedback
Received Signal Strength
Adjust
Data Rate/Power Or
Contention Window
Collision Inference Algorithm
Assumes Feedback is successfully received and that receiver is able to successfully abstract the Sender’s MAC Address.
Symbol error patterns
Bit error distribution and patterns
Note: assumes Feedback is successfully received and sender’s MAC address is decoded correctly by the AP

15. Metrics for Analysis Received Signal Strength (RSS) = S + I
S ~ Signal Strength
I ~ Interference
Bit Error Rate (BER) = total % incorrect bits
Symbol level errors: errors within transmission frame
Multiple tools used to analyze symbol-level errors
Symbol-level Network Coding forWireless Mesh Networks
Sachin Katti, Dina Katabi, Hari Balakrishnan, and Muriel Medard
Massachusetts Institute of Technology
A symbol is a small sequence of bits (typically a few bytes) that the code treats as a
single value.

16. Symbol-level Errors
Symbol Error Rate (SER)- % symbols received in error
Errors Per Symbol (EPS)- average # errors within each symbol
Symbol Error Score (S-score): , where Bi is a burst of n bits

17. S-Score
Collision 
S-Score =
Channel Fluctuation
We find that, for example, 98% of the packets in error due to weak signal have an S-Score of
500 or less, while 26% packets in error due to collision have an S-Score of 500 or less. Thus, by using a cutoff of 500, we would be able to detect 74% of collision cases. 
S-Score =

18. Experimental Design Three possibilities at R:
Packet received without error
Packet received in error
No packet received

19. Experimental Design Two transmitters, T1 and T2
Two receivers, R1 and R2
Receiver R hears all signals

20. Analysis of Results Metric Collision Weak Signal RSS
Higher (90% > -73dBm)
Lower (98% < -73dBm)
BER
Higher (24% =< 12% BER)
Lower (98% =< 12% BER)
SER
Unremarkable
EPS
Higher (45% =< 28% EPS)
Lower (98% =< 28% EPS)
S_Score
Higher (28% =< 500)
Lower (98% =< 500)
Plots are a cut and paste from the paper.
By using a cutoff of 500, Paper claims that it achieves the ability to detect 74% of collisions, with 2% being false positives.

21. Analysis of Results
While it is clear that using RSS in this case clearly distinguishes between the cases of collision and weak signal, using BER does not provide the same level of accuracy. In particular, we see that it becomes difficult to distinguish between cases (i) and (ii) using BER because a smaller colliding packet (200-byte in this case) would cause fewer bits in error. On the other hand, as shown in Figure 8, the joint distribution of SER and EPS is useful in distinguishing these cases.
The intuition follows from the observation that error packets in collision suffer higher symbol-error rates and correspondingly higher errors per symbol as a function of the symbol-error rates. From the scatter plot shown in Figure 8, we can observe that for higher values of SER, the values of EPS get streamlined into a high yet narrow range allowing for a more accurate prediction of collision versus signal as to the cause of a packet loss.

22. Begs the Question
Is it worth it? Successful almost 60%, false positive rate of 2.4%
Check out this accuracy?
Check out this accuracy?

23. Design Components Client Module
Optimation Logic place on Client Module and requires on minimal support from APs
Client Module

24. Multi-AP COLLIE Error packet sent to a central COLLIE server
Most important where the capture effect is dominant

25. Multi-AP Results Static situation averaged 30% gains in throughput
For multiple collision sources and high mobility, throughput gains reached 15-60%

26. Collision Analysis

27. Some Problems
Capture Effect
Packet size
Packet Kickback

28. Conclusions
COLLIE implementation achieves increased throughput (20-60%) while optimizing channel use
40% reduction in retransmission costs
Implementation can be done over standard , resulting in much lower startup costs than other protocols

29. Questions?