1. Model-Driven Energy-Aware Rate Adaptation
M. Owais Khan, Vacha Dave, Yi-Chao Chen,
Oliver Jensen, Lili Qiu, Apurv Bhartia,
Swati Rallapalli
The University of Texas at Austin
MobiHoc 2013, Bangalore, India

2. Rate adaptation needs to energy-aware!
Motivation
Multi-antenna devices are becoming common
Offer diverse rate choices
# of antennas, modulation, coding, # of streams
Rate adaptation – beaten to death problem?
Large capacity gain, but significantly more energy!
Talk
Last few years have witnessed a prolific growth of multi-antenna devices or MIMO deployment. (routers, fancy smartphones, etc.)
These devices offer diverse rate choices (# of antennas, streams, coding, etc.) thereby providing ability to serve high throughput traffic.
This makes the rate adaptation problem increasingly interesting, and has/is being discussed widely in the wireless community
However, most existing work talks about this in the context of maximizing throughput/performance.
But with these multi-antenna devices energy consumption becomes a valid concern. (see table).
PIC – check if the phone actually has MIMO
Mode
Intel TX
Intel Rx
Single Antenna
1.28 W
0.94 W
Two Antennas
1.99 W
1.27 W
Three Antennas
2.10 W
1.60 W
Rate adaptation needs to energy-aware!

3. What’s the big deal? Fixed antenna systems are fairly simple
𝑒𝑛𝑒𝑟𝑔 𝑦 𝑐𝑜𝑛𝑠𝑢𝑚𝑒𝑑 ∝𝐸𝑇𝑇
Energy-aware rate adaptation becomes simple
Highest rate
Lowest ETT
Minimum energy!
Can this be applied to MIMO as well?
𝑒𝑛𝑒𝑟𝑔 𝑦 𝑚𝑢𝑙𝑡𝑖−𝑎𝑛𝑡𝑒𝑛𝑛𝑎𝑠 ≫𝑒𝑛𝑒𝑟𝑔 𝑦 𝑠𝑖𝑛𝑔𝑙𝑒−𝑎𝑛𝑡𝑒𝑛𝑛𝑎
Additional hardware and RF chains
But multiple data streams reduces transmission time!
Talk
The problem is significantly more challenging as compared to single antenna Wi-Fi.
In single ant., energy consumption is directly proportional to the ETT, which makes the problem simpler since,
Highest supported rate will have the smallest ETT
Max. throughput also min. energy consumption

4. Energy vs. Tx time: the trade-off
Reduce time by 68%!
Reduce time by 50%!
This figure compares the transmission time of a single antenna with that of using 2 or 3 antennas.
The plot is based on the energy model of Intel 5300 card, which I’ll be talking about later.
X-axis shows transmission time of a single antenna transmission
Y-axis shows the % of transmission time that 2 and 3 antenna MIMO transmissions must reduce to have the same energy as single antenna transmission.
From the figure, we see that for a single antenna transmission time of 0.2ms, using 3-antennas is beneficial only if we can reduce the transmission time by more than 68%. On the other hand, for a transmission time of 1.3ms, this reduction just needs to be 50%. So, in the best case scenario, the 3-antennas can reduce the transmission time by 66% (for same MCS as compared to single antenna) should lead to savings in energy!
But the rate and # of antennas used depend on multiple factors like channel condition, wireless card and frame size.
Exact rate and # of antennas depend on multiple factors
Channel condition, wireless card and frame size
No single setting to minimize energy
Single antenna ≠ minimum energy

5. Hence, our work!
Understand how energy consumption in these devices relates to these factors
Design an energy-aware rate adaptation scheme that can minimize energy!
And hence it becomes essential to have a comprehensive understanding of how energy consumption relates to these factors,
And design a rate adaptation algorithm that can minimize energy according to the current n/w condition and wireless device.

6. Contributions
Extensive power measurements for multiple n wireless adapters
Derive energy model based on power measurements
Propose an energy-aware rate adaptation scheme
Evaluate using simulation and testbed experiments

7. Why Model-Driven? Why not probing? Model-driven
Slow given the large search space w/ MIMO
Hard to accurately measure the power of probe frames
Model-driven
Estimate power consumption for each rate under the current channel condition
Directly select the one w/ lowest power

8. Power Measurement Setup
Multiple wireless cards
Intel 5300N series
Atheros 11n
Windows mobile smartphone
Monsoon power monitor
One reading/μs
Maximum power value every 200μs
Owais
Please add the picture
I got lost in the first line. And so will the audience.
‘Monitor’ need not be caps
Picture speaks a 1000 words 
Tell the cards that were used
The power monitor used (monsoon)

9. Power Measurement Methodology
Measurements at both transmitter and receiver
Different configurations
Frame size ( bytes)
# of antennas
802.11n compliant data rates

10. Atheros Energy Measurements
Atheros Wi-Fi transmitter
Atheros Wi-Fi receiver
𝑬𝒏𝒆𝒓𝒈 𝒚 𝒄𝒐𝒏𝒔𝒖𝒎𝒆𝒅 ∝𝑬𝑻𝑻
Slope of the line depends on # of antennas

11. Intel Energy Measurements
Owais
Is this is the model? This is probably just the actual measurements!
Model is just table 2, isn’t it?
We need the learnings from the measurements
We need the table 2.
We also need the MAPE formula.
Intel Wi-Fi transmitter
Intel Wi-Fi receiver
𝑬𝒏𝒆𝒓𝒈 𝒚 𝒄𝒐𝒏𝒔𝒖𝒎𝒆𝒅 ∝𝑬𝑻𝑻
Slope of the line depends on # of antennas

12. Measurement-Driven Energy Model
Use least-square fitting to develop energy models
𝐸 𝑡𝑥 =𝐴 ×𝐸𝑇𝑇+𝐵
𝐸 𝑟𝑥 =𝐶 ×𝐸𝑇𝑇+𝐷
where 𝐴, 𝐵, 𝐶, 𝐷 vary for different wireless cards
Intel
Atheros 𝐴
0.24× 𝑛 𝑡𝑥 ×𝑀𝐼𝑀𝑂+1.02
0.38× 𝑛 𝑡𝑥 𝐵
0.045× 𝑛 𝑡𝑥
0.040× 𝑛 𝑡𝑥 𝐶
0.30× 𝑛 𝑟𝑥 +0.61
0.142× 𝑛 𝑟𝑥 +0.30 𝐷
0.064× 𝑛 𝑟𝑥
0.048× 𝑛 𝑟𝑥
Based on the observations, we develop simple energy models using least-squares fitting to find the coefficients that best match the energy consumption of the different wireless cards.
The parameters A..D vary across different wireless cards. You can see that the parameters although are similar but are not identical.

13. Error is consistently below 5%!
Validating the model
𝑀𝐴𝑃𝐸=𝑚𝑒𝑎𝑛( 𝑥− 𝑥 ′ 𝑥 )
𝑥 : actual energy consumption
𝑥 ′ : estimated energy consumption
Card
Transmission
Reception
Atheros
3.4%
1.3%
Intel
0.65%
1.4%
Phone
4.9%
3.6%
Now once have the model, we then want to see how accurate our model is. The table shows the mean absolute percentage error of our energy model vs. the measurement data, and we see that the error is consistently below 5% indicating a close match.
Error is consistently below 5%!

14. Energy Aware Rate Adaptation
Select rate for next transmission that minimized energy!
But this can be used easily for other objectives as well …

15. Channel State Information (CSI)
Specifies amplitude and phase between tx-rx pair
Measured for all subcarriers using preamble
Reported once per received frame
pp-SNR can be calculated as:
𝑆𝑁𝑅 𝑚 𝑀𝑀𝑆𝐸 = 𝐸 𝑠 𝑁 𝑡 𝑁 [ 𝐻 𝐻 𝐻+ 𝐸 𝑠 𝑁 𝑡 𝑁 −1 𝐼] −1
To achieve this goal, the protocol first obtains Channel State Information (CSI) as seen by the receiver.
IEEE n standard specifies how to calculate and report CSI.
CSI values are a collection of MxN matrices each of which specify the amplitude and phase between each pair of antennas.
Its measured for all subcarriers using the preamble and reported once per received frame.
Given the CSI values, the post-processed SNR can be calculated depending on the MIMO transmission mode used.
For instance, here is the expression used for spatial multiplexing mode where we use a MMSE equalizer to calculate the pp-SNR for the mth stream.
For brevity I’ve excused the formulas used for transmit diversity modes (look in the paper for CDD or STBC modes)

16. Compute loss rate Map pp-SNR to un-coded BER using known relationship
Convert un-coded BER to coded BER
Calculate frame error rate (FER)

17. Estimate energy consumption
AP or back-end server keeps table of energy models
Account for most commonly used Wi-Fi cards
Get the make/model of the Wi-Fi card
Explicit feedback or passive detection
Compute ETT based on frame loss rate (FER)
𝐸𝑇𝑇= 𝑠𝑖𝑧𝑒 𝑟𝑎𝑡𝑒 1 (1−𝐹𝐸𝑅)
Select the MCS with the minimum energy
Different variations of schemes are possible
Partial packet recovery (PPR) support
Only the ETT calculation changes (ref. paper)

18. Putting it all together
Measure CSI
Calculate pp-SNR
Calculate estimated loss rate
Compute ETT
Select rate minimizes energy!
Move this upward.

19. Evaluation Trace-driven simulator Testbed
Static and mobile channel traces using Intel 5300
Written in python
Testbed
Uses the Intel 5300 card
Iwlwifi driver is modified to support rate adaptation

20. Simulation Methodology
Developed in Python using real CSI traces
Different schemes are supported
Sample Rate with MIMO
Effective SNR
Maximum throughput
Minimum energy
Minimum Energy with throughput constraint
Talk
The simulator was coded in Python and uses real CSI traces captured from the Intel 5300 cards. We change the data rate on each of the frames using the various rate adaptation schemes.
For comparison, we consider the following rate adaptation schemes:
Sample rate with MIMO – default sample rate, but extended to work with MIMO functionality. It probes the network at a random rate every 10 frames and selects the rate that min. transmission time (incl. retx time). The goal here is to max. throughput without considering energy at all.
Effective SNR – selects data rate based on eff. SNR derived from CSI values. Computes eff. SNR for the frame and aims to max. throughput.
Max. throughput – pics MCS that maximized tput.
Min. energy – pics MCS that min. energy while keeping frame delivery > 90%.
Min. energy with throughput constraint. – ensures throughput is no lesser than X% of the max. throughput

21. Intel Transmitter (Static)
14%-24%
17%-31%
1-10%
1-2%
10%-22%
1-19%
0-1%
1-2%
Energy (nJ/bit)
Throughput (Mbps)
Talk
First we evaluate the performance under static network conditions. The traces here consist of 2000 CSI samples.
As we can see, compared to the scheme that maximizes throughput,
The energy-aware rate adaptation scheme consumes 14-24% less energy for the Intel card.
Throughput loss for the card is 10-22%.
Compares w/ Eff. SNR and Sample-rate, min. energy reduces transmitter energy by 17-31%. Throughput loss is 1-19%. Energy saving is higher and throughput saving is lower in these cases because these schemes are not optimal for either throughput or energy.
ET-Put X balances throughput and energy. For example, compared w/ Max.-Tput scheme, ETput80 minimizes energy whilst ensuring at least 80% of max. throughput. It saves energy of upto 10% for Intel transmitter, while throughput reduction is within 1%.
The Oracle schemes know the exact CSI of the next frame and eliminate the performance degradation caused by the prediction error. We can see that CSI prediction error causes only 1-2% more energy consumption and 1-2% throughput reduction, indicating that the impact of prediction error is small.
Minimum Energy provides significant power savings

22. Intel Receiver (Static)
25-35%
26-42%
1-10%
1-4%
10%-26%
1-23%
0-1%
1-5%
Energy (nJ/bit)
Throughput (Mbps)
Talk
First we evaluate the performance under static network conditions. The traces here consist of 2000 CSI samples.
As we can see, compared to the scheme that maximizes throughput,
The energy-aware rate adaptation scheme consumes 25-35% less energy for the Intel card.
Throughput loss for the card is 10-26%.
Compares w/ Eff. SNR and Sample-rate, min. energy reduces transmitter energy by 26-42%. Throughput loss is 1-23%. Energy saving is higher and throughput saving is lower in these cases because these schemes are not optimal for either throughput or energy.
ET-Put X balances throughput and energy. For example, compared w/ Max.-Tput scheme, ETput80 minimizes energy whilst ensuring atleast 80% of max. throughput. It saves energy of upto 10% for Intel transmitter, while throughput reduction is within 1%.
The Oracle schmes know the exact CSI of the next frame and eliminate the performance degradation caused by the prediction error. We can see that CSI prediction error causes only 1-4% more energy consumption and 1-5% throughput reduction, indicating that the impact of prediction error is small.

23. Intel Receiver (Mobile)
29-31%
34-40%
1-16%
2-6%
15-19%
2-13%
0-2%
3-6%
Energy (nJ/bit)
Throughput (Mbps)
Talk
First we evaluate the performance under static network conditions. The traces here consist of 2000 CSI samples.
As we can see, compared to the scheme that maximizes throughput,
The energy-aware rate adaptation scheme consumes 14-24% less energy for the Intel card.
Throughput loss for the card is 10-22%.
Compares w/ Eff. SNR and Sample-rate, min. energy reduces transmitter energy by 17-31%. Throughput loss is 1-19%. Energy saving is higher and throughput saving is lower in these cases because these schemes are not optimal for either throughput or energy.
ET-Put X balances throughput and energy. For example, compared w/ Max.-Tput scheme, ETput80 minimizes energy whilst ensuring atleast 80% of max. throughput. It saves energy of upto 10% for Intel transmitter, while throughput reduction is within 1%.
The Oracle schmes know the exact CSI of the next frame and eliminate the performance degradation caused by the prediction error. We can see that CSI prediction error causes only 1-2% more energy consumption and 1-2% throughput reduction, indicating that the impact of prediction error is small.

24. Intel Receiver with PPR
26% to 28%
~9%
26% to 29%
~9%
Energy (nJ/bit)
Throughput (Mbps)
Min. Energy saves 26-28% energy over Max. Throughput. The throughput loss is 26-28%.
ETput80 saves energy 9-21% while the throughput degradation is < 9%.
Moreover comparing the PPR schemes w/ non-PPR schemes, we can see that PPR schemes improve the energy by 6-23% by
extracting correct symbols from partially corrupted frames.
Minimum Energy provides significant power savings

25. Always use single-antenna systems?
One of the interesting questions that can arise is that, should we just resort to using a Single Antenna!
This can be more interestingly analyzed through this graph under the context of packet size.
As we can see, for 1000 byte packets, Min-Energy always selects single-antenna. However, as the frame size increase, we see the multiple-antennas coming into picture a lot more. E.g. with 5000-byte packets, we see that it becomes more advantageous to use multiple antennas to minimize energy.
Multiple-antennas provide energy savings for larger frames because for small frames, the preamble size dominates the total transmission time. Hence, using multiple antennas only results in small reductions in ETT, which does not offset the additional energy required to power up multiple antennas.
As the frame size increases, using multiple antennas leads to larger reductions in ETT, which more than offsets this additional energy.
Single Antenna performance decreases with increasing packet size

26. Testbed Implemented scheme on Intel Wi-Fi link 5300 driver
Used tool in [Halperin10] to extract CSI from driver
Static channel
200 UDP Packet of 1000 bytes each transmitted
Results averaged over 10 runs
Mobile channel
Receiver moves away from transmitter at walking speed
Results averaged over 5 runs

27. Static Channel 19% 6% 28% 16% 24% 11% 22% 2% Energy (nJ/bit)
Throughput (Mbps)
Now we see the throughput and energy consumption numbers for static experiments.
As we can see,
Min-Energy reduces the energy consumption by 19% for the transmitter and 28% for the receiver. The throughput reduction is 24% for tx, and 22% for the rx.
ETPutX smoothly trades-off between the two objectives.
ETPut80 reduces energy by 6% at an 11% throughput loss at the transmitter. For the receiver, ETPut80 reduces the energy by 16% with a throughput reduction of 2%.

28. Energy savings do not degrade with the channel!
Mobile Channel
This figure shows how the MCS changes over a mobile experiment for MaxTput and MinEnergy.
MCS 0-7: use 1 antenna, 8-15 uses 2 antennas and uses 3 antennas. In each case, num_spatial_streams = num_antennas.
In region 1, where the channe is good, MaxTput transmits using all 3 antennas at MCS 22. Since MinEnergy tries to minimize energy, it uses MCS 6, the highest 1-antenna rate that can be supported by the current channel. 16.9% energy savings!
As the receiver moves away from the transmitter, the channel degrades and MaxTput goes to MCS 14, while Min-Energy still uses MCS 6. Energy improvement reduces to 11.9% because MCS 14 consumes less energy.
In region 3, MaxTput goes from MCS 14 to MCS 12 (2 antennas). MCS 12 consumes more energy than MCS 14 because fewer antennas (higher ETT). Min_Energy continues at MCS 6 thereby giving savings of 21%. In region 4, this increases to
Interesting to note that even though channel degrades continuously, the energy savings do not follow the trend. In fact, region 2 has the least savings while region 3 has the highest.
Energy savings do not degrade with the channel!

29. Related Work
Models based on data size [Carvahlo04], empirical study [Bala09]
Neither considers effects of multiple antennas, data rates, tx power
Study power consumption under different parameters[Halperin10]
Do not develop energy model
Energy measurement and Models
Extensively studied [Bicket05, Holland01, Sadeghi02, Wong06, etc.]
None of these schemes consider minimizing energy
Energy based rate adaptation [Li12]
Limited effectiveness of probing-based approach
Rate Adaptation
Power Saving Mode Optimization [Napman10, Sleepwell11, E-mili11]
Complementary to our work
Power Savings

30. Conclusion Collect and analyze extensive power measurements
Derive simple energy models for transmission/reception
Develop model-driven energy-aware rate adaptation scheme
Experimentally show significant energy savings possible
14-37% over existing approaches
PPR extensions can be even better

31. Thank You! apurvb@cs.utexas.edu
Questions ???
Thank You!