1. Improving Loss Resilience with Multi-Radio Diversity in Wireless Networks
Allen Miu, Hari Balakrishnan
MIT Computer Science and Artificial Intelligence Laboratory
C. Emre Koksal
Computer and Communication Sciences, EPFL
Today, I'll be presenting a system that uses multi-radio diversity to improve loss resilience in wireless networks. This is joint work with my advisor, Hari Balakrishnan, and Emre Koksal from EPF.

2. The problem  Inconsistent and poor performance
New wireless applications demand high performance
But: wireless channels are loss-prone
Interference
Noise
Attenuation
Multi-path
Mobility
 Inconsistent and poor performance
We are motivated to design a wireless network to support new applications that demand high performance
But the main problem, as we all know very well, is that wireless communication suffers from time-varying losses, due to a number of well-known factors, leading to inconsistent and poor performance in wireless networks.

3. Main Idea of This Paper
Given that losses are often path-dependent, location-dependent, but statistically independent between different receiving radios, multiple radios that receive versions of the same frame may work together to correctly receover a frame.
One of the current solutions is to use coding. Because it’s hard to do well on highly variable channels, people use retransmission.
Unfortunately, retransmission can be useless and wasteful when the channel quality drops for a long time.
In case the channel becomes very poor, we can use a more robust modulation by adapting the bit-rate.
However, like coding, it is hard to do for highly varying channels
Also, it has the side effect of slowing down other clients as varying the bit-rate varies the channel time that is used in the shared medium.

4. Today’s wireless LAN (e.g., 802.11)
May use only one path
Uses only one communication path
AP1
We believe today’s solution for error recovery is limited by the system architecture. Today’s wireless LANs uses a cellular architecture that constrains each device to using only one communication path at a time. As a result, communication suffers due to the variations of that one path.
However, the wireless medium is a broadcast medium. A transmission has the ability to reach multiple receivers at the same time, and this ability comes for free.
Internet

5. Multi-Radio Diversity (MRD) – Uplink
Today’s wireless LAN (e.g., )
Multi-Radio Diversity (MRD) – Uplink
May use only one path
Allow multiple APs to simultaneously receive transmissions from a single transmitter
10%
MRDC
AP1
We took advantage of this property and developed a system, called MRD, which allows a sender to use multiple receivers simultaneously to increase the reliability of the wireless link.
This example shows how MRD works in the uplink direction.
In the uplink direction, we add access points to the infrastructure and tune them to listen in the same radio frequency as the one used by another nearby APs.
In this case, a transmission succeeds if either one of these APs receives it correctly.
The multiple paths help reduce loss in the network. As an example, assume that the loss rate of the two paths are 10% and 20% as shown. The loss rate effectively reduces from 10% (when we use only 1 path) to 2% when both paths are used, where 2% is just the product of the loss rates in each path assuming independent losses. Later, we’ll examine if losses are independent in real life.
There are times when more than one AP receives a correct copy of the transmission, producing a duplicate frame in the infrastructure. To filter out the duplicates, we add a centralized controller.
Although we describe transmission succeeds if at least one of the receivers get a correct copy of the transmission, later, we will show how we can use the centralized controller to produce good data frames, even when none of the receivers receive the transmission correctly.
SKIP to next slide
Note that AP2 needs only to listen and forward frames to the MRD controller, and is not required to perform any other function. It is really a passive radio so adding these additional APs doesn’t increase interference or reduce the capacity of the network.
In our design, client devices continues to associate with one AP at a time, just like a conventional wireless LAN. The only changes we make to the system architecture is adding passive radios for listening and a centralized controller in the infrastructure.
Scalability: sniff & suppress traffic on ethernet
20%
Internet
AP2
Loss independence 
simultaneous loss = 2%

6. Multi-Radio Diversity (MRD) – Downlink
Allow multiple client radios to simultaneously receive transmissions from a single transmitter
MRDC
AP1
MRD works in the downlink direction as well. To support downlink, we simply invert the uplink architecture by installing multiple radios, and by running the MRD Controller at the client.
Internet
AP2

7. Are losses independent among receivers?
Broadcast experiment at fixed bit-rate: 6 simultaneous receivers and 1 transmitter
Compute loss rates for the 15 receiver-pair (R1, R2) combinations
Frame loss rate FLR(R1), FLR(R2) vs. simultaneous frame loss rate FLR(R1 ∩ R2)
In order for MRD to work, it is crucial that losses are not correlated between receivers.
We ran an experiment to find out. In our set up, we used one transmitter and six simultaneous receivers.
We computed loss rates for 15 receiver-pair combinations and compare the loss rates observed at each individual receiver, and the rate at which both receivers get a corrupt frame at the same time for the same transmission.

8. Individual FLR > Simultaneous FLR
y = x
FLR R1 R2
R1*R2
In this graph, we plot the individual frame loss rate (on the y axis) against the simultaneous loss rate at both receivers (on the x axis) for the 15 receiver-pair combinations.
We observe that the loss rates at the individual receivers (lie above the y=x line), which means that they are higher than the simultaneous loss rate and suggest that losses are not totally correlated.
To test for independence, we plot the product of the two frame loss rates against the simultaneous loss rate.
We see that the product of the loss rates are more or less equal to the simultaneous loss rate, which suggests that losses were roughly independent in our experiments.
FLR(R1 ∩ R2)

9. Challenges in developing MRD
How to correct simultaneous frame errors?
Frame combining
How to handle retransmissions in MRD?
Request-for-acknowledgment protocol
How to adapt bit rates in MRD?
MRD-aware rate adaptation
We just saw some evidence that suggests MRD might work. So what are the challenges in developing this system?
Earlier, I mentioned how we could use the centralized MRD controller to combine multiple corrupt frames into a good one.
So one of the challenges is to develop a practical frame combining algorithm to correct frame errors from multiple corrupt frames.
There cases when frame combining fails and the sender needs to retransmit. So the next challenge is developing a way to handle retransmissions in a multi-radio scheme.
For this, we designed RFA to handle retransmissions.
Finally, there are times when both frame combining and retransmission fail to recover losses. In this case, we need to adapt the bit-rates.
For this, we introduce a MRD-aware scheme for rate adaptation.
In the upcoming slides, I will go over our solutions to the challenges in the order shown.

10. Bit-by-bit frame combining
TX:
Combine failure
2. Select bit combination at unmatched bit locations, check
CRC
Patterns
CRC Ok 1 R1 -- R2 X
1. Locate bits with
unmatched value X O
Corrected frame
MRD uses an algorithm based on bit by bit combining for correcting errors.
I am going to though an example to show how this works.
Suppose the transmitter just transmitted this bit pattern as shown, and suppose each receiver received a corrupt version of this transmission as shown.
The first step of combining is to locate the bits with disagreeing values by applying the XOR operation.
Then we initialize a test frame to the value of a corrupt frame R1, and for each unmatched bit position, which is shaded in blue, we assign a bit value from one of the corrupt frames in the corresponding unmatched position. For example, look at the second test frame and see how we assigned the bit value from frame R2 to the right-most unmatched bit position.
After that, we run CRC check on this test frame.
If it doesn’t pass, we repeat the process, assigning another bit combination at the unmatched positions until we find a correct frame.
It is easy to see how the algorithm fails when the two corrupt frames share a common corrupt bit because the error will be hidden from the first step of the algorithm.
This is pretty much the entire algorithm. Although it is very simple, it is not practical because the # of CRC checks is exponential to the # of unmatched bits. Since there can easily be more than 100 unmatched bits between two large data frames, this is a big problem.
Problem: Exponential # of CRC checks in # of unmatched bits.

11. Block-based frame combining
Observation: bit errors occur in bursts
Divide frame into NB blocks (e.g., NB = 6)
Attempt recombination with all possible block patterns until CRC passes
# of checks upper bounded by 2NB
Failure rate increases with NB
To make frame combining practical, we developed block-based combining.
Block based combining takes advantage of the observation that bit errors frequently occur in bursts in a frame. We have a number of graphs to show this is the case in our experiments.
This suggests we can save overhead by combining in blocks of bits at a time, rather than in one bit at a time.
In block-based combining, we divide frames into N blocks. In our implementation, N is 6.
And we attempt all block combinations until the test frame passes the CRC.
The upside of using this approach is that we limit the number of CRC checks to 2^N, where NB is a small fixed number.
But the downside is that the failure rate increases. Analogous to bit-by-bit combining, block-based frame combining fails to correct frame errors when bit errors occur in the same block between frames, and this failure rate increases with NB.
Comparing the upside and downside of this algorithm, we see a fundamental tradeoff between complexity and the effectiveness of the combining algorithm.

12. Failure decreases with NB and burst size
1.0
Frame size = 1500B
Probability of failure
0.8
0.6
NB = 2
0.4
NB = 4
To help us understand this tradeoff, we developed an analytical model that predicts how the block combining failure rate varies with N and also with the burstiness of errors in the channel.
This graph shows the results of our model.
The X axis is the burst length parameter in our model. Basically, the large the value, the longer the burst errors are in the channel.
Each curve shows the frame combining failure rate for each number of block divisions that is used.
We see that the failure rate drops rapidly as a function NB. This makes intuitive sense because higher number of divisions allow finer grained combining, which in turn reduces the chance of simultaneous block error.
We also see that failure drops rapidly wrt the burst error length. This makes sense because it reduces the chance for simultaneous block errors as the errors become clustered into fewer blocks.
In our implementation, we use this model to help us determine a good value for N. We picked N=6 in our implementation because we see a diminishing return for higher values.
0.2
NB = 6 …
NB = 16 10 20 30 40 50
Burst error length parameter

13. Flawed retransmission schemes
Conventional link-layer ACKs do not work
Final status known only to MRDC
Two levels of ACKs are redundant
Cannot disable link-layer ACKs
We saw how frame combining works and how it sometimes fails. In this case, the sender needs to retransmit a frame. How do we do this?
In order to know when to retransmit, you need to have a feedback mechanism.
The conventional scheme, using only link-layer ACK from the receiver won’t work, because the final status of frame combining is known only to the MRD controller.
Clearly, the MRD controller need to send a separate feedback to the sender. But this introduces two levels of ACK, one at the link-layer and one at the MRD layer. This seems a little redundant.
We can try to disable link-layer ACKs to eliminate this redundancy. But wireless networks overloads the link layer ACKs for signalling congestion and adjusting the contention backoff window. So we cannot disable link-layer ACKs either.

14. Request-for-acknowledgment (RFA) for efficient feedback
DATA IP IP
RFA
MRD
MRD
MRD-ACK
DATA
DATA
link
link
link
ACK
MRDC
Our solution is to introduce a request for acknowledgment protocol that allows the Sender to obtain feedback from the MRD Controller only when necessary.
1) In RFA, the sender first transmit a data frame and if the AP that the client is associated with receives the frame correctly, it replies with an ACK.
2) Upon receive this ACK, the sender knows that the infrastructure has successfully received the data frame, so it continues to send the next data frame in its queue.
3) Notice MRD does not automatically send an ACK, which eliminates the redundancy.
4) However, if the link-layer ACK is missing, the sender sends an RFA to ask for feedback from the MRDC about the status of the transmission. Notice that the sender HAS TO to send a RFA because the link-layer ACK can get lost. So only the sender knows if a link-level ACK is missing and therefore only the sender knows when it needs feedback from the MRD controller.
5) We’ve covered the basics of RFA. The paper has more detail about how to implement RFA efficiently with low overhead.
SKIP TO NEXT SLIDE
6) To improve efficiency, sender does not block on RFA --- transmits up to N outstanding frames to keep link utilized
7) RFA is embedded as a 1 bit value in the header of data frames (also for reliability)
MRDC replies with ACK feedback

15. MRD-aware rate adaptation
Standard rate adaptation does not work
Reacts only to link-layer losses from 1 receiver
Uses sub-optimal bit-rates
MRD-aware rate adaptation
Reacts to losses at the MRD-layer
When both frame combining and retransmission fail, the sender needs to adapt the bit rate.
By-in-large, many rate adaptation algorithm, such as the one we used, works by reacting to losses in the wireless link. However, if the standard algorithm were used in MRD, it will use a sub-optimal bit-rate because it ignores the fact that some frames may be recovered by the MRD-layer.
Our solution is to expose the MRD feedback information to the rate adaptation layer and make it react to losses at the MRD-layer.
Even though this is a simple modification, the solution has a deeper implication on link quality management.
The conventional wisdom in managing link quality is to have the transmitter adapt the bit rate first before allowing the sender to do a handoff and use a different path.
MRD takes an opposite approach, which allows a sender to use multiple simultaneous paths opportunistically, then adapt the bit rate. This strategy allows MRD to reduce loss variations in the channel, and reduce the burden for the rate adaptation layer to try to adapt with high sensitivity.
Implication: First use multiple paths,
then adapt bit rates.

16. Experimental setup ~20 m R2 R1 L
802.11a/b/g implementation in Linux (MADWiFi)
L transmits 100,000 1,500B UDP packets w/ 7 retries
auto bit rate (6, 9, 12, 18, 24, 36, 48, 54)
L is in motion at walking speed, > 1 minute per trial
Variants: R1, R2, MRD (5 trials each)
We conduct a real life experiment to measure how well MRD works.
We implemented MRD on Linux and set up an experiment using two receivers and a laptop sender.
During the experiment, the laptop sends 100,000 unicast 1500 UDP packets with 7 retries.
We let the sender adapt using all the available bit rates in a.
To produce variations in the channel, we set the laptop in motion by carrying it around and walking with it in the space shaded in blue during the experiments.
We ran 3 variants of the experiment. Two of them (R1 and R2) uses the conventional single-path scheme with standard rate adaptation and MRD, with the modified rate algorithm.

17. MRD improves throughput
18.7 Mbps
2.3x Improvement
Throughput (Mbps)
8.25 Mbps
This plot shows the throughput in each of the 5 trials, represented by the different color bar, for the single path communication schemes and MRD.
The average throughput for the better single path communication scheme is 8.25 Mbps. MRD, in contrast, achieved an average of 18.7 Mbps, which represents a 2.3x improvement. R1 R2
MRD
Each color shows a different trial

18. MRD maintains high bit-rate
Fraction of transmitted frames
Frame recovery data
(% of total losses at R1)
via R %
frame combining 7.3%
Total %
MRD managed to outperform the conventional schemes. Now, we explain why it performs so well.
We analyzed the trace and plot the distribution of bit rates used to transmit each data frame in each scheme.
First, we see that the conventional schemes selected a variety of bit rates due to the highly varying channel.
In contrast, the MRD scheme selected a bit rate of 36 Mbps for almost 80% of the transmitted frames, contrasting to the conventional scheme where 80% of the frames were transmitted using less than half that bit rate.
The reason why MRD was able to maintain such a high bit rate is because it was able to hide frame losses from the rate adaptation algorithm. MRD was able to recover roughly 42% of all loss frames seen at R1 through the alternate path via R2 and an additional 7% through frame combining, recover reducing the number of loss frames by about 50%.
Raw FLR was 35%.
LOWER VARIANCE MEANS BETTER FOR TCP 6 9 12 18 24 36 48 56
Selected bit rate (Mbps)

19. Delay Analysis 10-4 10-3 10-2 10-1 1 Fraction of delivered packets
User space implementation caused high delay
Next, we want to examine how MRD affects packet delivery delay.
The delay for the conventional scheme spreads over the millisecond range because the laptop sender used a variety of different bit rates.
MRD delivered about 50% of packets within a very low delay because it was able to use high bit rate and recovered corrupt frames without a retransmission.
However, roughly 30% packets were delivered with a higher delay than the single transmission schemes. The reason is that these packets required a retransmission. Although there is delay associated with RFA during retransmission, we believe a large part of this delay is caused by our user-space implementation of the MRDC and the large reorder buffer inside the MRDC for in order packet delivery. So we expect delay to reduce with better implementation of our system.
10-4
10-3
10-2
10-1 1
One way delay (10x s)

20. Related work Physical layer spatial diversity techniques
Antenna diversity
MIMO/802.11n
Macro-diversity in CDMA networks
Retransmission with memory [Sindhu ’77]
Opportunistic forwarding [Biswas ‘05][Jain ’05]
Bit rate selection (AARF, RBAR, MiSer, OAR, Sample Rate)
There are a number of physical layer diversity that exploits diversity in the wavelength-scale. In contrast, MRD exploits diversity in a macroscopic scale from APs that are deployed throughout the environment.
CDMA cellular exploits macro-diversity by allowing a handset to communication with multiple base-stations. But our design is completely different.
Our frame combining technique is based on an old idea called retransmission with memory, which combines copies of retransmitted frames for error correction. We basically use the same idea to combine copies of a frame collected from different receivers.
There’s been recent work on opportunistic forwarding, which takes advantage of spatial diversity in the multi-hop mesh network setting. Because it is a multi-hop environment, these work focus on distributed coordination whereas MRD takes advantage of the wired-backbone and a centralized controller to do things like frame combining.
Finally, there’s a lot of related work in rate adaptation. They all assume that communication happens between a single transmitter and a single receiver. We offered a solution that appears to work well for MRD but finding an optimal rate adaptation algorithm for MRD remains an open problem (e.g., improve autorate by using unmatched bit info.)
SKIP
Last year, we published a paper on a system that performs fine-grained path selection in wireless networks. The idea behind FGPS is to switch downlink transmissions among a set of nearby access points on a frame-per-frame basis. Thus, FGPS also improves loss resilience using radio diversity. However, FGPS exploits diversity in the transmit side of the system, whereas MRD facilitates diversity in the receiver side. The two approaches are complimentary and in our future work, we plan to evaluate how the two can be combined to improve performance.
CDMA: optimization for CSMA-based wireless local area networks
There are a number of physical layer techniques that use diversity to improve communication. Almost all of today's cards have multiple diversity antenna on them to allow it to get better signal reception. However, the card is limited to using one antenna at a time. In contrast, MRD uses not just multiple antenna, but multiple radios to allow simultaneous receptions of a given transmission.
Recently, the IEEE is incorporating advanced physical-layer coding techniques into the WiMAX and standards to spread data over multiple input and output antenna to boost performance. However, these techniques aim to mitigate path-dependent effects localized at a single receiver. In contrast, MRD to distribute the reception of transmissions over access points that are spread across over a wide area. Thus, the two are really complimentary to each other, each aims to mitigate path-dependent effects at a different scale.

21. Summary Design of Multi-Radio Diversity WLAN
Block-based frame combining
Request-for-acknowledgment protocol
MRD-aware rate adaptation
Analysis of block-based frame combining
Experimental evaluation
MRD reduces losses by 50% and improves throughput by up to 2.3x
In summary, our research contribution is the design of MRD, which incorporates
frame combining for error recovery,
RFA for retransmissions, and
A MRD-aware rate adaptation mechanism
We presented a model that allows us to choose a parameter that balances the tradeoff between complexity and effectiveness of the block-based frame combining algorithm
We’ve implemented the MRD on a testbed and show that MRD…