Evaluation of the Minstrel Rate Adaptation Algorithm

in IEEE 802.11g WLANs

1,2 1 1
Dong Xia Jonathan Hart Qiang Fu
1 2
School of Engineering and Computer Science Ministry of Social Development, PO Box 1556
Victoria University of Wellington, PO Box 600 Wellington 6140, New Zealand
Wellington 6140, New Zealand
Email: {dong.xia, jonathan.hart, qiang.fu}@ecs.vuw.ac.nz

Abstract—Rate adaptation varies the transmission rate of a Therefore, in this paper we present an experimental study
wireless sender to match the wireless channel conditions, in order
which analyzes the performance of Minstrel by comparing it
to achieve the best possible performance. It is a key component of
IEEE 802.11 wireless networks. Minstrel is a popular rate against fixed rates in a real-world IEEE 802.11g wireless
adaptation algorithm due to its efficiency and availability in testbed. Fixed rates provide a performance benchmark which a
commonly used wireless drivers. However, despite its popularity, good rate adaptation algorithm should be at least able to meet,
little work has been done on evaluating the performance of if not exceed. We therefore consider an evaluation against
Minstrel or comparing it to the performance of fixed rates. In this fixed rates to be an important tool in the analysis of rate
paper, we conduct an experimental study that compares the adaptation algorithm behavior.
performance of Minstrel against fixed rates in an IEEE 802.11g To fully understand the performance of Minstrel we
testbed. The experiment results show that whilst Minstrel
conducted experiments of three different scenarios: static
performs reasonably well in static wireless channel conditions, in
some cases the algorithm has difficulty selecting the optimal data
channel transmission, dynamic channels with fast variation,
rate in the presence of dynamic channel conditions. In addition, and dynamic channels with slow variation. This work builds
Minstrel performs well when the channel condition improves on our previous study which evaluates Minstrel [2]. In this
from bad quality to good quality. However, Minstrel has trouble paper we consider different scenarios to provide a more
selecting the optimal rate when the channel condition detailed evaluation. We have also further analyzed our results
deteriorates from good quality to bad quality. in order to provide deeper insights into Minstrel and rate
adaptation algorithm design. Our results show that whilst
I. INTRODUCTION
Minstrel performs well in static wireless channel conditions,
The IEEE 802.11 standards specify multiple data rates that the algorithm has difficulty selecting the optimal data rates in
can be used by a wireless sender, depending on the modulation more dynamic channel conditions. Furthermore, Minstrel
and coding schemes used at the PHY layer. High data rates performs well compared to fixed rates when the channel
transmit data faster than low data rates, however high data conditions improve from a poor quality starting point,
rates are more susceptible to bit errors. This means more however it has trouble selecting optimal rates when the
packets are lost on low quality wireless channels with high bit channel conditions deteriorate from a high quality starting
error rates (BERs). Low data rates take longer to transmit point.
packets over the link, however they are more resistant to bit The rest of the paper is organized as follows. We discuss
errors and are more likely to be successfully transmitted over a related work in Section II. Section III describes the Minstrel
poor quality wireless link. rate adaptation algorithm, and Section IV outlines our
Wireless channels are extremely variable and can be experiment methodology. Section V presents and discusses
our results, and finally Section VI concludes the paper.
affected by a number of different factors, such as interference
from other wireless devices, multi-path fading and signal
attenuation. As such, one of the key components of an 802.11 II. RELATED WORK
system is the rate adaptation mechanism, which adapts the The topic of rate adaptation has been heavily researched.
data rate used by a wireless sender to the wireless channel Previous research generally either targets the design of new
conditions. Rate adaptation is optimization problem: if we use rate adaptation algorithms [3]-[10], or compares the
a rate that is too high, many of the packets will be dropped due performance of different rate adaptation algorithms [11]-[13].
to bit errors, however if we use a rate that is too low, the The SampleRate algorithm [14] was previously considered one
wireless channel is not fully utilized. of the best publicly available algorithms, and is often used for
Rate adaptation is crucial to wireless network performance, comparison in these studies. Minstrel works similarly to
and therefore much research has been conducted on how to SampleRate, although it differs in the statistic used to measure
design effective rate adaptation algorithms. Minstrel [1] is a the success of each rate. SampleRate uses the expected
popular, recently-developed rate adaption algorithm, although transmission time, whereas Minstrel uses the expected
there is currently very little literature studying its performance. throughput.
 Ancillotti et al. evaluated three different rate adaptation four rate-count pairs, named r0/c0, r1/c1, r2/c2, and r3/c3. A
algorithms in wireless mesh networks [15]. This evaluation packet is first transmitted at rate r0 for c0 attempts. If these
includes SampleRate, and there is some comparison between attempts are not successful, Minstrel transmits the frame at
rate adaptation algorithms and fixed rates. However, the focus rate r1 for c1 attempts. The process continues until either the
of the paper is on the correlation between the number of packet is successfully transmitted or ultimately discarded after
senders and the rate adaptation algorithms. The evaluation did (c0 + c1 + c2 + c3) unsuccessful transmission attempts. The
however show in a particular scenario that the performance of following section discusses how Minstrel chooses the r-values
some of the fixed rates was better than the evaluated rate in the retry chain.
adaptation algorithms.
Wong et al. proposed an algorithm called Robust Rate B. Rate Selection
Adaptation Algorithm (RRAA) [16], which was evaluated There are two categories of transmission: normal
against SampleRate and earlier algorithms such as Auto Rate transmission occurs 90% of the time and sampling
Fallback (ARF) and Adaptive Auto Rate Fallback (AARF). transmission occurs for the remaining 10% of packets. Table I
Acharya et al. proposed Wireless cOngestion Optimized gives a summary of the rate selection decisions.
Fallback (WOOF) [17], a rate adaptation algorithm that 1) Normal Transmission: During normal transmission the
attempts to identify congestion related packet losses. WOOF r-values in the retry chain are chosen as follows: r0 is set to
was evaluated against other solutions including SampleRate. the rate that achieves the highest expected throughput, r1 is
Much of this work was conducted before Minstrel was the rate with the second highest expected throughput, r2 is the
developed, so the authors were unable to compare with rate with the highest probability of success, and finally r3 is
Minstrel. As such, performance evaluations of Minstrel are set to the lowest available data rate.
scarce. 2) Sampling Transmission: Minstrel relies on having
The most closely related works are the evaluations done by accurate statistics about the success rate of transmissions at
Yin et al. [12][13]. These are the only previous evaluations of each data rate. Of course, it has to attempt to send packets at
Minstrel to our knowledge. In one paper the four rate
each data rate in order to have statistics on them. 10% of the
adaptation algorithms found in the MadWiFi driver (namely
data frames are sent as sampling transmissions, where a
ONOE, AMRR, SampleRate and Minstrel) are evaluated
random rate not currently in the retry chain is chosen to
against one another [12]. The second paper compares the
algorithms found in the mac80211 Linux driver framework sample. The the r-values are chosen as follows: r0 is set to
[13], namely Minstrel and PID. These evaluations use a wired whichever is higher out of the sample rate or the rate with the
testbed with co-axial cables to emulate wireless channels. This highest expected throughput, and r1 is set to whichever is
is a clever solution because the algorithms are evaluated in an lower. r2 and r3 remain the rate with the highest probability of
environment that is free of interference, which means the success and the lowest available rate respectively.
experiments are able to produce repeatable results. These C. Statistics Calculation
evaluations show that Minstrel performs far better than the
The final piece of the puzzle is how Minstrel calculates the
other algorithms. Therefore we chose not to include other
probability of success and expected throughput for each data
schemes such as SampleRate in our evaluation. Moreover, a
rate. Minstrel maintains the probability of successful
wired testbed cannot fully reproduce the complex channel
transmission at each data rate as an Exponentially Weighted
environments found in real-world 802.11 deployments,
Moving Average (EWMA). This probability is based on the
therefore we feel that real-world experimental studies are also
historical success rate of packet transmissions at each data
important for fully understanding rate adaptation behavior.
rate. This probability is used to estimate the throughput of
Whilst comparisons among different rate adaptation
each rate and the retry chain is re-evaluated based on this
algorithms are certainly very useful, most of these studies do
estimate every 100ms. In each 100ms sampling window, the
not directly compare against fixed rates. We feel that
success rate, RS, is calculated for each data rate based on the
comparing rate adaptation against fixed data rates is an
important step towards the design of effective rate adaptation TABLE I. MINSTREL RETRY CHAIN [1]
algorithms.
Sampling Transmission Normal
Rate
Transmission
III. THE MINSTREL ALGORITHM Random < Best Random > Best

r0 Best rate1 Random rate Best rate

There are three parts to the Minstrel algorithms: the retry
chain mechanism, the rate decision process and the statistic r1 Random rate Best rate Second best rate
calculations. Best probability
2
r2 Best probability Best probability
A. Retry Chain r3 Base rate
3
Base rate Base rate
The Minstrel rate adaptation algorithm uses a mechanism
called a multi-rate retry chain, which enables it react to short- 1
The data rate that gives the highest throughput
term variations in channel quality. The retry chain consists of 2
The data rate that has the highest rate of successful transmissions
3
The lowest available data rate
historical observation of packet successes and failures as in
(1), where NS is the number of packets transmitted
successfully at the data rate and NT is the total number of
packets attempted at the data rate.
RS = NS / NT (1)
P(t + 1) = RS × (1 – α) + P(t) × α (2)
The RS value is then used to alter the measured value for
the probability of success for each data rate using the EWMA
expression (2). The EWMA parameter α is used to determine
how much weight is given to the RS value from the new
sampling period. The default value of α is 0.25 which means
the new probability of success is comprised of 75% of the new
sample and 25% of the previous probability of success. Finally
Minstrel calculates an expected throughput for each data rate,
T, as in (3).
T = Psuccess × (B / t) (3)
The expected throughput T is the number of bytes B
transferred in time t (i.e. the raw throughput of the rate) scaled
by the probability of success Psuccess which is maintained by the
EWMA expression (2). This results in an expected throughput
for each rate which is based on the station’s previous
observations of the proportion of packets that have been
successfully transmitted at the data rate.

IV. EXPERIMENT METHODOLOGY

A. Testbed Platform
Figure 1 shows the network topology of the testbed. A
wireless Access Point (AP) is connected via Ethernet to a
sender PC. We use iperf to send UDP traffic from the sender
PC to a laptop receiver, which is wirelessly associated with the
AP. The AP and laptop use wireless cards containing the
Atheros AR5414 chipset. The AP is based on the MikroTik
Routerboard R52 platform and runs the Linux-based OpenWrt
backfire 10.03 [18] operating system with a 2.6.32 kernel. The
server PC and client laptop run Ubuntu 10.04 with a 2.6.23
kernel. The AP used 802.11g mode.
This real-world evaluation was conducted in an office
building on a university campus. We are aware of other
wireless network operating in the area. While interference
from these networks can affect our experiment system, it also
reflects the complex nature of a real-world wireless
environment and the fact that wireless networks often do
operate in the vicinity of one another. It is important to know
how wireless systems behave in the presence of interference,
 Figure 2. Throughput with a static channel Figure 4. Throughput with rapid channel variation

Figure 3. Minstrel rate selection breakdown in a static channel Figure 5. Minstrel rate selection breakdown with rapid channel variation

Higher data rates always have a higher chance of being

V. RESULTS
sampled. As Figure 3 shows, even when the transmission
power is low, a significant proportion of packets that are still
A. Static Channel attempted at high data rates. This means that Minstrel has
Figure 2 shows the throughput of Minstrel and each of the better statistics on the high data rates because it uses them
fixed rates. The transmission power is static during each more often than the low data rates. However, Minstrel’s
experiment, and we conduct experiments at a number of tendency to sample and select higher data rates doesn’t incur a
different power levels. The throughput of Minstrel decreases significant performance cost in stable channel conditions.
as the transmission power decreases. In addition, Minstrel B. Rapid Channel Variation
achieves comparable performance to the best-performing fixed
rate at all power levels. This experiment also shows that In this section, we study the performance of Minstrel in
reducing the transmission power has a significant performance dynamic channel conditions by periodically switching between
impact on the higher data rates, because more frames are lost high and low transmission power levels.
due to the lower quality wireless channel. Figure 4 shows the performance of fixed rates and Minstrel
Figure 3 shows a percentage breakdown of the data rates in relation to the duration of the low transmission power. As
attempted by Minstrel in the static channel experiment at each the duration of the low quality channel increases, the
different power level. When the transmission power is set to performance of Minstrel and the high fixed rates (54, 48, and
17 dBm (the highest quality channel), Minstrel mostly 36 Mbps) suffers significantly. This poor performance is
attempts to use 54 Mbps and 48 Mbps. However, at the lower expected for the high rates, as they drop more packets due to
transmission power levels the channel drops more packets, and bit errors in a low quality wireless channel. The performance
Minstrel attempts the lower data rates more often. When the of Minstrel and high rates stabilizes after the low power
transmission power is set to 1 dBm, more than 70% of duration increases past 5 seconds. In contrast, the performance
attempts are made at the 12 Mbps and 18 Mbps rates. By the of the low rates is relatively consistent even as the low power
nature of Minstrel sampling algorithm, if the randomly duration increases. We expect that sudden changes would
selected rate is higher than the current best rate, the random reduce Minstrel’s performance suffer, however it is surprising
rate is used first, otherwise the random rate is used second. to see that Minstrel is outperformed in most cases by the 12
Mbps and 18 Mbps fixed rates, which unable to adapt at all.
 use the optimal rate a greater proportion of the time when the
update period is longer.
From looking at the rate distribution (not shown), we can
see that Minstrel attempts to use low data rates more often as
the transmission power decreases. However, fixed rates still
outperform Minstrel, suggesting that Minstrel makes incorrect
rate selection as the channel quality deteriorates.
Table II shows the rate most frequently used by Minstrel
(outside the brackets), and the fixed rate that achieves the
highest ideal throughput (inside the brackets) for each
transmission power update period at different power levels.
The ideal throughput, TI, is calculated as in (4) as the data rate
multiplied by the ratio of successful attempts (NS) to total
attempts (NT).
Figure 6. Throughput with gradually deteriorating channel conditions rates becomes costly in
The rate breakdown highly dynamic channel
statistics for Minstrel are conditions.
shown in Figure 5. In
Figure 4, we have C
observed that the best .
throughput is achieved by
the 54 Mbps and 48 Mbps G
rates when the duration of r
low power is low (2-3 a
seconds). When the d
duration is greater than 4 u
seconds, the 12 Mbps and a
18 l
Mbps rates have the best
throughput. In Figure 5, it C
can be h
seen that Minstrel most a
frequently selects 54 Mbps n
over the other rates for n
transmission. This makes e
Minstrel comparable to the l
best rates (54/48 Mbps) in
terms of throughput when V
the low power duration is a
short. As the duration r
increases, Minstrel uses the i
54 Mbps rate less a
frequently and the 18 t
Mbps rate more frequently. i
However Minstrel does not o
use the n
18 Mbps and 12 Mbps rates In this experiment, we
enough, which leads to the further analyze the behavior
poor performance we of Minstrel in dynamic
observed. The rightmost wireless environments.
bar in Figure 5 shows the However in this scenario the
case when the cycle consists channel conditions gradually
of high power for 2 seconds change from bad to good or
and low power for 10 vice versa so that Minstrel
seconds, in other words a has time to adapt. The
5:1 ratio of low power transmission power increases
duration to high power or decreases gradually, and
duration. However the the time period between
figure shows that Minstrel power level updates is varied
attempts low and high from 2 seconds to 5 seconds.
rates almost equally, when it Figure 6 shows the
clearly should use low rates throughput of Minstrel
a much greater percentage of and fixed rates
the time. This experiment as the transmission power
indicates that Minstrel’s decreases from 17 dBm to 1
tendency to use higher dBm, with different update
 use the optimal rate a greater proportion of the time when the
update period is longer.
From looking at the rate distribution (not shown), we can
see that Minstrel attempts to use low data rates more often as
the transmission power decreases. However, fixed rates still
outperform Minstrel, suggesting that Minstrel makes incorrect
rate selection as the channel quality deteriorates.
periods. The 18 Mbps TI = Table II showsn the rate most frequently used by Minstrel
and 24 Mbps fixed rates Rate × n
(NS / e
outperform Minstrel, l
NT)
although the performance (4)
difference between these c
two fixed rates and Minstrel should strive o
n
Minstrel becomes less to use the rate with d
significant as the the highest throughput i
duration of update period most often, however t
i
increases. When the during these experiments o
update period is short the it often fails to do so, as n
channel is relatively indicated by the bold s
dynamic, but as the entries in Table II(a). T
2s 3s 4s 5s
update period lengthens Minstrel often makes X
u u u u
the channel becomes incorrect rate choices p p p p
relatively static. Minstrel when the power level is 1 24 24 24 12
7 (1 (1 (1 (1
updates its retry chain low and when the 1 24 18 18 18
every 100ms, which means transmission power update 5 (1 (1 (1 (1
1 24 18 18 24
it updates 20 times in a 2 period is short. Again, 3 (2 (1 (1 (2
second period and 50 Minstrel tends to select a 1 24 18 24 24
times in a 5 second rate higher 1 (2 (2 (2 (2
9 24 24 36 36
period. Minstrel is able to (2 (2 (3 (3
TABLE II. RATE MOST 7 36 36 36 36
FREQUENTY ATTEMPTED BY (3 (3 (3 (3
MINSTREL (OUTSIDE BRACKETS) 5 36 36 36 36
VS. RATE WITH HIGHEST IDEAL (3 (3 (3 (3
THROUGHPUT (INSIDE 3 36 36 48 48
BRACKETS) (3 (3 (4 (4
1 48 48 48 48
T (4 (4 (4 (4
2 3 4 5 (
X
s s s s b
)
1 36 36 36 36
7 (3 (3 (3 (3
1 36 36 36 36 I
5 (3 (2 (3 (3 m
1 36 24 36 36 p
3 (2 (2 (3 (3 r
1 24 24 24 24 o
1 (2 (2 (2 (2 v
9 24 24 24 24 i
(1 (1 (2 (2 n
7 24 24 24 24 g
(1 (1 (1 (2
5 24 24 24 24
(1 (1 (1 (1 c
3 24 24 24 12 h
(1 (1 (1 (1 a
1 24 24 18 12 n
(1 (1 (1 (1 n
(
a e
) l

D c
e o
t n
e d
r i
i t
o i
r o
a n
t s
i
n
g

c
h
a
 However it does give Minstrel the ability to quickly make use
of higher rates when the channel conditions improve.

VII. ACKNOWLEDGEMENT
We are grateful to InternetNZ for its support in this project.

REFERENCES
[1] Minstrel Rate Adaptation Algorithm Documentation, http://madwifi-
project.org/browser/madwifi/trunk/ath_rate/minstrel/minstrel.txt
[2] D. Xia, J. Hart and Q. Fu, “On the performance of rate control algorithm
rd
minstrel”, Proceedings of the 23 IEEE International Symposium on
Personal, Indoor and Mobile Radio Communications (PIMRC), 9-12
September 2012, Sydney, Australia.
[3] A. Kamerman and L. Monteban, “WaveLAN II: A High-
performance Wireless LAN for the Unlicensed Band”, Bell Labs
Figure 7. Throughput with gradually improving channel conditions Technical Journal, vol. 2, no. 1, Fall 1997, pp. 118-133.
[4] M. Lacage, M.H. Manshaei and T. Turletti, “IEEE 802.11 rate
than the optimal rate. However, as the update period increases, adaptation: a practical
th
A
Minstrel’s selection of approach”, Proceedings C
rates becomes more of the 7 M
accurate. This
explains why Minstrel’s International Symposium on
Modeling, Analysis and
performance improves Simulation of
with the longer update Wireless and Mobile
Systems (MSWiM), 4-6
period. October 2004, Venice, Italy,
Figure 7 shows the pp. 126-134.
throughput of fixed rates and [5] ONOE Readme,
http://madwifi-
Minstrel when the project.org/browser/madwifi/trunk/
transmission power ath_rate/onoe
increases from 1 dBm to [6] B. Sadeghi, V. Kanodia, A.
Sabharwal and E. Knighlty,
17 dBm. Minstrel has “Opportunistic media access
superior performance over for multirate ad hoc
the fixed rates in all cases. networks”, Proceedings of
th
the 8 Annual International
This is the positive outcome Conference on Mobile
of the algorithm which Computing and Networking
(MobiCom), 23-28 September
aggressively samples higher 2002, Atlanta, Georgia, pp.
rates over lower rates. This 24-35.
means that as soon as the [7] J. Kim, S. Kim, S. Choi
and D. Qiao, “CARA:
channel conditions collision-aware rate
improve, Minstrel is adaptation for IEEE 802.11
WLANs”. Proceedings of the
quickly able to take 25
th
IEEE International
advantage and use higher Conference on Computer
rates. Table II(b) again Communications
(INFOCOM),
shows the rate most often 23-29 April 2006, Barcelona,
used by Minstrel and the rate Spain.
with the highest ideal [8] D. Qiao and S. Choi, “Fast-
responsive link adaptation for IEEE
throughput in each 802.11
situation. We can see that in WLANs”, Proceedings of
this scenario when the IEEE International
channel conditions are Conference on
Communications (ICC), 16-20
improving, Minstrel May 2005, Seoul, Korea,
becomes more accurate in vol. 5, pp.
choosing the 3583-3588.
[9] H. Rahul, F. Edalat, D.
rate with the highest ideal Katabi and C. Sodini, “Frequency-
throughput. Moreover, as the aware rate
update th
period increases, Minstrel adaptation and MAC An
has more time to adapt protocols”, Proceedings nu
and is able of the 15 al
L
to choose the correct rate U
more accurately. S
I
O
V N
I
In this paper, we
.
evaluate the
performance of Minstrel
C against fixed rates in a
O
N real-world IEEE 802.11g
C wireless testbed. Our
 However it does give Minstrel the ability to quickly make use
of higher rates when the channel conditions improve.

VII. ACKNOWLEDGEMENT
We are grateful to InternetNZ for its support in this project.

International Conference on REFERENCESMAC layer rate International

results show that Mobile Computing and
performance comparable control Workshop
to the best performing mechanisms”, 18 on
Minstrel can achieve a Networking fixed rates
(MobiCom), 20-25 September when the wireless channel Quality of Service (IWQoS),
2009, Beijing, China, pp. 193- 16-18 June 2010, Beijing,
204. conditions are relatively China.
[10] M. Vutukuru, H. static. With static channels, [13] W. Yin, P. Hu, J. Indulska
Balakrishnan and K. and K. Bialkowski,
Minstrel can usually “Performance of mac80211
Jamieson, “Cross-layer
wireless bit rate adaptation”,
accurately identify the best rate control mechanisms”,
Proceedings of the ACM rates to use. However, in Proceedings of the 14th ACM
International Conference on
SIGCOMM Conference on the presence of dynamic Modeling, Analysis and
Data Communication, 17-21 channel conditions, Minstrel
August 2009, Barcelona, Simulation of Wireless and
Spain, pp. 3-14. is significantly outperformed Mobile systems (MSWiM ),
[11] D. Qiao, S. Choi and K.G. by certain fixed rates. This is 31 October – 4 November
Shin, “Goodput analysis and 2011, Miami, Florida, pp. 427-
due to Minstrel’s tendency to 436.
link adaptation for IEEE
802.11a wireless LANs”, select rates that are too high [14] J. Bicket, “Bit-rate Selection
IEEE Transactions on Mobile for the current channel in Wireless Networks”, MSc
thesis, Dept. of Electrical
Computing, vol. 1, no. 4, conditions. By analyzing Engineering and Computer
October 2002, pp. 278-292. Minstrel’s rate choices we
[12] W. Yin, K. Bialkowski. J. Science, MIT, February 2005.
Indulska and P. Hu, “Evaluations can see than Minstrel [15] E. Ancillotti, R. Bruno and M.
Conti, “Experimentation and
of madwifi attempts to use high rates too performance evaluation of
often in a rapidly changing rate adaptation algorithms in
wireless channel. We also wireless mesh networks”,
th
Proceedings of the 5 ACM
looked at gradual changes Symposium on Performance
in channel quality, and Evaluation of Wireless Ad
Hoc, Sensor, and Ubiquitous
discovered that Minstrel has Networks (PE-WASUN), 27-
poor performance when the 28 October 2008, Vancouver,
wireless signal strength Canada, pp. 7-14.
[16] S.H. Wong, H. Yang, S. Lu
deteriorates from high to and V. Bharghave, “Robust
low. Conversely, Minstrel rate adaptation for 802.11
outperforms fixed rate when wireless networks”,
th
Proceedings of the 12
the signal strength Annual International
increases from low to high. Conference on Mobile
Computing and Networking
Again, this is due to (MobiCom), 24-29 September
Minstrel’s tendency to 2006, Los Angeles, CA, pp.
select high rates. This 146-157.
[17] P.A.K. Acharya, A. Sharma,
behavior is undesirable when E. Belding, K.C. Almeroth and
the channel conditions are K.
deteriorating, and Papagiannaki, “Rate
adaptation in congested
Minstrel pays a wireless networks through
performance penalty. real-time measurements”,
IEEE Transactions on Mobile
Computing, vol.
9, no. 11, November 2010 pp.
1535-1550.
[18] OpenWrt Backfire Readme,
http://backfire.openwrt.org/10.03/RE
ADME