Predictable 802.11 Packet Delivery from
Wireless Channel Measurements
Daniel Halperin
∗
, Wenjun Hu
∗
, Anmol Sheth
†
, and David Wetherall
∗†
University of Washington
∗
and Intel Labs Seattle
†
ABSTRACT
RSSI is known to be a fickle indicator of whether a wireless link
will work, for many reasons. This greatly complicates operation
because it requires testing and adaptation to find the best rate, trans-
mit power or other parameter that is tuned to boost performance.
We show that, for the first time, wireless packet delivery can be
accurately predicted for commodity 802.11 NICs from only the
channel measurements that they provide. Our model uses 802.11n
Channel State Information measurements as input to an OFDM re-
ceiver model we develop by using the concept of effective SNR. It
is simple, easy to deploy, broadly useful, and accurate. It makes
packet delivery predictions for 802.11a/g SISO rates and 802.11n
MIMO rates, plus choices of transmit power and antennas. We re-
port testbed experiments that show narrow transition regions (<2 dB
for most links) similar to the near-ideal case of narrowband, fre-
quency-flat channels. Unlike RSSI, this lets us predict the highest
rate that will work for a link, trim transmit power, and more. We use
trace-driven simulation to show that our rate prediction is as good
as the best rate adaptation algorithms for 802.11a/g, even over dy-
namic channels, and extends this good performance to 802.11n.
Categories and Subject Descriptors
C.2.1 [Computer-Communication Networks]: Network Archi-
tecture and Design—Wireless Communication
General Terms
Design, Experimentation
1. INTRODUCTION
Wireless LANs based on 802.11 are used almost everywhere,
from airports to zoos and in urban, suburban and rural areas. Mod-
ern wireless NICs provide a large and growing range of physical
layer configurations to obtain good performance across this range
of environments. With 802.11n, the latest version of the standard
that ships on most laptops, combinations of modulation, coding and
spatial streams offer rates from 6 Mbps to 600 Mbps [1]. Other im-
portant choices include transmit power, channel, and antennas.
For good performance, reliability and coverage, the physical layer
settings should match the RF channel over which the wireless sig-
nals are sent. This is evident in rate adaptation schemes [5, 10, 14,
28] that determine the highest rate for transmission, since a good
scheme has a large effect on throughput. Other work adapts trans-
mit power to reduce co-channel interference [17, 21, 25].
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
SIGCOMM’10, August 30–September 3, 2010, New Delhi, India.
Copyright 2010 ACM 978-1-4503-0201-2/10/08 ...$10.00.
In theory, it is simple to select the physical layer configuration
because this is directly determined by the specifics of the RF chan-
nel. The signal-to-noise ratio (SNR) is the gold standard for per-
formance in narrowband channels. Textbook formulas relate the
error rate of different modulations to the SNR [27]. The best rate or
required transmit power is then simple to compute.
In practice, 802.11 LANs have never used channel measurements
as more than a coarse indicator of expected performance. There
have simply been too many ways in which the observed measure-
ments and actual performance fail to match the predictions of the-
ory. For example, the most accessible channel measurement is re-
ceived signal strength indication (RSSI), which serves as a proxy
for the true SNR. RSSI measurements are samples that may vary
over packet reception, be mis-calibrated, or be corrupted by in-
terference, all of which are known to be issues in practice [6, 10,
22]. Even if RSSI were perfect, it does not reflect the frequency-
selective fading of 802.11 channels, which are not close to narrow-
band. Nor does it account for imperfect receivers that may greatly
degrade performance [3, 10]. Due to these factors, the minimum
RSSI at which a rate starts to work varies by more than 10 dB for
real links [22, 30, 31].
To reconcile these viewpoints, a form of guided search is widely
used in practice to select operating points [21, 24, 29]. Packet de-
livery is simply tested for a rate or transmit power to see how well
it works. If the loss rate is too high, a lower rate (or more power) is
used, otherwise a higher rate (or less power) is tested. SampleRate
is a well-known algorithm of this kind for finding transmit rates [5].
This approach is very effective for slowly varying channels and sim-
ple configurations (e.g., a few rates with fixed transmit power and
channel) since the best setting will soon be found.
However, search becomes less effective as channels change more
quickly and the configuration space becomes more complex. Both
of these factors are trends: 802.11 clients are increasingly used
when they are truly mobile, both walking and in vehicles; and NICs
that are now being deployed with 802.11n depend on multiple an-
tennas, which adds another dimension to and increases the size of
the search space. Also, tuning combinations such as rate and power
is much more complex.
For rate selection, recent work has made headway by measur-
ing symbol-level details of packet reception. In particular, SoftRate
uses the output of soft-Viterbi decoding for each symbol to estimate
the bit error rate (BER) [28]. This allows it to predict the effects on
packet delivery of changing the rate. AccuRate uses symbol er-
ror vectors for the same purpose [23]. However, these methods
are not defined for selecting other useful parameters, such as trans-
mit power, and they do not extend from 802.11a/g to 802.11n, e.g.,
when selecting antennas or numbers of spatial streams.
In this work, we return to the basic problem of using theory to
connect the performance of 802.11 NICs on real links to measured
channels in practice. The opportunity to make progress has arisen
for two reasons. First, 802.11n NICs measure the channel at the
OFDM subcarrier level to support MIMO (multiple antenna) op-
eration. They report this information in a standard Channel State
1
Information (CSI) format [1]. This provides a much richer source
of information than RSSI. Note that this CSI naturally applies to
802.11a/g rates because they are a subset of 802.11n rates. Sec-
ond, modern NICs use OFDM, which gives channel estimates that
are less susceptible to interference than spread spectrum (because
of lower correlation), and are calibrated. Both factors lead to more
meaningful measurements than in the past.
We use the CSI as input to a model of receiver processing that we
develop to predict packet delivery. Our model uses the concept of
an effective SNR for a multi-carrier channel [18], such as OFDM,
in which there are different subcarrier SNRs, plus approximations
for coding, interference between MIMO streams, and decoding al-
gorithms. It requires no per-link calibration and predicts delivery
for a wide range of configurations (including rates, transmit power,
antenna selection, and spatial streams) from a single CSI measure-
ment. We also expect it can be extended to new factors such as
beamforming for even wider applicability in the future.
We make two contributions in this paper. Our main contribution
is to show how to accurately predict the performance of commodity
802.11 OFDM NICs over real links using only the channel mea-
surements that the NIC provides. We believe this to be a first. Our
packet delivery model is evaluated with measurements over two sta-
tionary indoor wireless testbeds built from PCs and 3-antenna In-
tel 802.11a/g/n NICs. For a wide range of configurations, we can
predict whether a link will successfully deliver packets (>90%),
outside of a narrow (<2 dB for most links) uncertainty region that
is similar to behavior over the near-ideal channel of nodes con-
nected by a wire. This lets us consistently predict the best rate
to use over a channel, and perform other tasks such as trim ex-
cess transmit power. In contrast, RSSI often fails to reflect perfor-
mance by a wide enough margin that it does not reliably predict
the best rate or power setting, especially for dense modulation and
higher coding rates (transition >7 dB for 10% of the links). A key
factor in this improvement is the use of effective SNR to capture
frequency-selective fading, which is clearly visible in our testbeds.
No published work has explored effective SNR measures in 802.11
beyond simulation, to the best of our knowledge. Note that our ef-
fective SNR model does not predict the performance of links under
interference. However, our measurements show that its estimate
of interference-free link quality is robust to interfering transmis-
sions (§5). We also discuss ways to handle persistent interference.
Our method is practical and can be applied to many classic prob-
lems, including rate adaptation, transmit power tuning, and channel
and antenna selection. While we must leave most of this to future
work, we demonstrate how our model can inform rate adaptation.
Our second contribution is a rate selection algorithm that is as good
as the best 802.11a/g rate adaptation algorithms and extends this
excellent performance to 802.11n. Our algorithm simply uses our
model to predict the highest rate for the channel, repeating to track
the channel over time. We use a trace-driven simulation to compare
it with SampleRate, which is widely used in practice, and SoftRate,
which has the best published performance. Our algorithm tracks
the best rate nearly as well as is possible, even for dynamic mobile
channels. It performs very well for MIMO rates, and supports en-
hancements such as transmit power trimming and antenna selection.
As far as we are aware, there is no other reported work on 802.11n
rate adaptation that is evaluated for real, 802.11 channels, and no
other rate adaptation algorithms that support these enhancements.
In the rest of this paper, we first motivate the need for better de-
livery predictions in §2, and then present our model in §3. Our ex-
perimental testbeds are described in §4, and our model is evaluated
in §5. In §6, we use simulation to study rate selection guided by
our model. §7 discusses related work, and §8 concludes the paper.
Modulation Coding Rate Data Rate (Mbps)
BPSK 1/2 6.5
QPSK 1/2 13.0
QPSK 3/4 19.5
QAM-16 1/2 26.0
QAM-16 3/4 39.0
QAM-64 2/3 52.0
QAM-64 3/4 58.5
QAM-64 5/6 65.0
Table 1: 802.11n single-stream rates.
2. MOTIVATION
Existing predictions of packet delivery for a given link are based
on its Received Signal Strength Indication (RSSI) value. This is
widely available as a proxy for the SNR. We characterize this map-
ping to motivate our research.
802.11 Setting. Our work applies to 802.11a/g/n radios that use
coded Orthogonal Frequency Division Multiplexing (OFDM). 20 or
40 MHz channels are divided into 312.5 kHz bands called subcar-
riers, each of which sends independent data simultaneously. Con-
volutional coding is applied across the bits for error correction and
bits are interleaved to spread them in frequency. Each subcarrier
in a packet is modulated equally, using BPSK, QPSK, QAM-16, or
QAM-64, with 1, 2, 4 or 6 bits per symbol, respectively. The data
rates depend on the combination of modulation and coding.
Our experimental platform uses 802.11n radios that operate on
20 MHz channels. The single-stream 802.11n rates are shown in
Table 1. The main innovation in 802.11n is the use of multiple an-
tennas for spatial multiplexing. By using MIMO processing, multi-
ple streams can be sent at the same time, each at the single-stream
rate, for higher overall rates. Note that the details of single-stream
802.11n differ slightly from 802.11a/g (optimized coding rates and
more data subcarriers), but in ways that are not material for our
work so that we can treat 802.11n as a superset of 802.11a/g.
Packet Delivery versus RSSI/SNR. Textbook analyses of modu-
lation schemes give delivery probability for a single signal in terms
of the signal-to-noise (SNR) ratio [8], typically expressed on a log
scale in decibels. This model holds for narrowband channels with
additive white Gaussian noise. It predicts a sharp transition region
of 1–2 dB over which a link changes from extremely lossy to highly
reliable. This makes the SNR a valuable indicator of performance.
RSSI values reported by NICs give an estimate of the total sig-
nal power for each received packet. From RSSI, the packet SNR
can then readily be computed using NIC noise measurements.
1
We
generated performance curves using SNR for a real 802.11n NIC
over a simple wired link with a variable attenuator and for a sin-
gle transmit and receive antenna. The result is shown for all single
antenna 802.11n rates in Figure 1(a). We observe a characteristic
sharp transition region for packet reception rate (PRR) versus SNR.
This is despite the relatively wide 20 MHz channel, 56 OFDM sub-
carriers, coding and other bit-level operations. This is the behavior
we want from a link metric in order to predict packet delivery.
In contrast, packet delivery over real wireless channels does not
exhibit the same picture. Figure 1(b) shows the measured PRR ver-
sus SNR for three sample rates (6.5, 26, and 65 Mbps) over all wire-
less links in our testbeds, using the same 802.11n NICs. The SNR
of the transition regions can exceed 10 dB, so that some links easily
work for a given SNR and others do not. There is no longer clear
separation between rates. This is consistent with other reported
1
We refer to the metric computed from RSSI and noise measure-
ments as the packet SNR, RSSI-based SNR, or simply RSSI.
2
0 5 10 15 20 25 30
0
20
40
60
80
100
Packet−level SNR (dB)
PRR
6.5
13
19.5
26
39
52
58.5
65
(a) A wired 802.11n link with variable attenuation has a
predictable relationship between SNR and packet recep-
tion rate (PRR) and clear separation between rates.
0 5 10 15 20 25 30 35
0
20
40
60
80
100
Measured packet SNR (dB)
PRR
6.5
26
65
(b) Over real wireless channels in our testbeds, the transition
region varies up to 10 dB. This loses the clear separation be-
tween rates (and so only three rates are shown for legibility).
Figure 1: Measured (single antenna) 802.11n packet delivery
over wired and real channels.
measurements that show RSSI does not predict packet delivery for
real links [3, 22, 30, 31].
Impact of Frequency-Selective Fading. Many possible factors
cause the observed variability for real channels, including NIC cal-
ibration, interference, sampling, and multipath. Here, we look at
frequency-selective fading due to multipath, as our experiments
show this to be a major factor.
Multipath causes some subcarriers to work markedly better than
others although all use the same modulation and coding. These
channel details, and not simply the overall signal strength as given
by RSSI, affect packet delivery. Figure 2 illustrates this with the
measured subcarrier SNRs for four different links in our testbed
averaged over a 5-second run. All links are shown at the closest
transmit power level, in steps of 2 dB, to 80% packet delivery when
using the 52 Mbps rate. However, the fading profiles vary signifi-
cantly across the four links. One distribution is quite flat across the
subcarriers, while the other three exhibit frequency-selective fading
of varying degrees. Two of the links have two deeply-faded subcar-
riers that are more than 20 dB down from the peak.
These links harness the received power with different efficien-
cies. The more faded links are more likely to have errors that must
be repaired with coding, and require extra transmit power to com-
pensate. Thus, while the performance is roughly the same, the
most frequency-selective link needs a much higher overall packet
SNR (30.2 dB) than the frequency-flat link (16.5 dB). This differ-
ence of almost 14 dB highlights why RSSI-based SNR does not re-
liably predict performance. Fading and its effects are well-known.
However, it is rare to see data that shows fading for real links and
NICs because it has been difficult to measure.
Impact of multiple streams. The use of multiple antennas adds an-
other dimension to the problem of predicting packet delivery. While
5
15
25
35
45
-28 -14 0 14 28
SNR (dB)
Subcarrier index
PRR 83%, SNR 30.2dB
PRR 78%, SNR 27.1dB
PRR 74%, SNR 18.2dB
PRR 80%, SNR 16.5dB
Figure 2: Channel gains on four links that perform about
equally well at 52 Mbps. The more faded links require larger
RSSIs (i.e., more transmit power) to achieve similar PRRs.
we do not present further motivating data here, we briefly note that
this makes the problem more difficult, not simpler. To begin with,
there is now an RSSI for each receive antenna. This makes it dif-
ficult to know which RSSI or function of RSSIs to use to predict
delivery even when there is a single spatial stream. When multiple
streams are sent simultaneously, they interfere on the channel. The
MIMO processing used to separate them depends on the details of
the channel, and less of the signal will be harnessed if the RF paths
are correlated. This adds variability that exacerbates fading effects.
3. PACKET DELIVERY MODEL
Our goal is to develop a model that can accurately predict the
packet delivery probability of commodity 802.11 NICs for a given
physical layer configuration operating over a given channel. We
want our model to be simple and practical, so that it can be readily
deployed, and to cover a wide range of physical layer configura-
tions, so that it can be applied in many settings and for many tasks.
In particular, the scope of our model is 802.11n including multiple
antenna modes, of which single antenna 802.11a/g is a subset. This
scope is sufficient for many current and future networks. We model
delivery for single packet transmission only, leaving extensions for
interference and spatial reuse to future work.
Model Design. The structure of our model is simple: given 1) the
current state of the RF channel between transmitter and receiver,
and 2) a target physical layer configuration of the NIC, it predicts
whether that link will reliably deliver packets in that configuration.
For the first piece of input, we use 802.11n Channel State In-
formation (CSI). The CSI is a collection of MxN matrices H
s
in
which each describes the RF path (SNR and phase) between all
pairs of N transmit and M receive antennas for one subcarrier s.
It is reported by the NIC in a format specified by the standard [1],
with details in §4.2. An 802.11n NIC can probe a receiver to gather
CSI, or use channel reciprocity to learn CSI from a received packet.
The CSI is a much richer source of information than the RSSI, and
it gives us the opportunity to develop a much more accurate model.
The second form of input is the target physical layer configura-
tion for which we want to predict delivery. This is specified as the
choice of transmit and receive antennas, transmit power level, and
transmit rate (as the combination of modulation, coding, and num-
ber of spatial streams). Other choices, such as beamforming, could
be added in the future. The only restriction is that the CSI includes
the antennas and subcarriers used in the target configuration.
For the model output, we define that the link will work, i.e., will
reliably deliver packets, if we predict ≥90% packet reception rate.
We do not try to make predictions in the transition region during
which a link changes from lossy to reliable. Predictions there are
3
OFDM
Demodulator
Deinterleaver
Convolutional
Decoder
Descrambler
(0)
Received
signal
MIMO Stream
Separation
Separated signals
for each spatial stream
(1)
Scrambled,
coded bits
(3)
(2)
Scrambled,
interleaved,
coded bits
(4)
Scrambled bits
(5)
Received
bitstream
Packet
processing
Figure 3: The 802.11n MIMO-OFDM decoding process. MIMO receiver separates the RF signal (0) for each spatial stream (1).
Demodulation converts the separated signals into bits (2). Bits from the multiple streams are deinterleaved and combined (3) followed
by convolutional decoding (4) to correct errors. Finally, scrambling that randomizes bit patterns is removed and the packet is
processed (5).
Modulation Bits/Symbol (k) BER
k
(ρ)
BPSK 1 Q
√
2ρ
QPSK 2 Q
√
ρ
QAM-16 4
3
4
Q
p
ρ/5
QAM-64 6
7
12
Q
p
ρ/21
Table 2: Bit error rate as a function of the symbol SNR ρ for
narrowband signals and OFDM modulations. Q is the standard
normal CDF.
likely to be variable, and simply knowing when the link starts to
work is useful information in practice.
802.11 Packet Reception. The model must account for the action
of the 802.11 receiver on the received signal. This is a complex pro-
cess described in many pages of the 802.11n specification [1]. Our
challenge is to capture it well enough with a fairly simple model.
We begin by describing the main steps involved (Figure 3).
First, MIMO processing separates the signals of multiple spatial
streams that have been mixed by the channel. As wireless chan-
nels are frequency-selective, this operation happens separately for
each subcarrier. The demodulator converts each subcarrier’s sym-
bols into the bits of each stream from constellations of several dif-
ferent modulations (BPSK, QPSK, QAM-16, QAM-64). This hap-
pens in much the same way as demodulating a narrowband channel.
The bits are then deinterleaved to undo an encoding that spreads
errors that are bursty in frequency across the data stream. A paral-
lel to serial converter combines the bits into a single stream. For-
ward error correction at any of several rates (1/2, 2/3, 3/4, and 5/6)
is then decoded. Finally, the descrambler exclusive-ORs the bit-
stream with a pseudorandom bitmask added at the transmitter to
avoid data-dependent deterministic errors.
Modeling Delivery. We build our model up from narrowband de-
modulation. Standard formulas summarized in Table 2 relate SNR
(denoted ρ) to bit-error rate (BER) for the modulations used in
802.11 [8]. CSI gives us the SNR values (ρ
s
) to use for each sub-
carrier. For a SISO system, ρ
s
is given by the single entry in H
s
.
In OFDM, decoding is applied across the demodulated bits of
subcarriers. If we assume frequency-flat fading for the moment,
then all the subcarriers have the same SNR. The link will behave
the same as in our wired experiments in which RSSI reflect real
performance and it will be easy to make predictions for a given SNR
and modulation combination. We can use Figure 1(a) to measure
the fixed transition points between rates and thus make our choice.
Frequency-selective fading complicates this picture as some weak
subcarriers will be much more likely to have errors than others that
are stronger. To model a link in this case, we turn to the notion of an
effective SNR. This is defined as the SNR that would give the same
error performance on a narrowband channel [18]. For example,
the links in Figure 2 will have effective SNR values that are nearly
equal because they perform similarly, even though their RSSIs are
spread over 15 dB.
The effective SNR is not simply the average subcarrier SNR; in-
deed, assuming a uniform noise floor, that average is indeed equiv-
alent to the packet SNR derived from the RSSI. Instead, the effec-
tive SNR is biased towards the weaker subcarrier SNRs because it
is these subcarriers that produce most of the errors. If we ignore
coding for the moment, then we can compute the effective SNR by
averaging the subcarrier BERs and then finding the corresponding
SNR. That is:
BER
eff,k
=
1
52
X
BER
k
(ρ
s
) (1)
ρ
eff,k
= BER
−1
k
(BER
eff,k
) (2)
We use BER
−1
k
to denote the inverse mapping, from BER to SNR.
We have also called the average BER across subcarriers the effec-
tive BER, BER
eff
. SoftRate estimates BER using internal receiver
state [28]. We compute it from channel measurements instead.
Note that the BER mapping and hence effective SNR are func-
tions of the modulation (k). That is, unlike the RSSI, a particular
wireless channel will have four different effective SNR values, one
describing performance for each of the modulations. In practice, the
interesting regions for the four effective SNRs do not overlap be-
cause at a particular effective SNR value only one modulation will
be near the transition from useless (BER ≈0.5) to lossless (BER
≈0). When graphs in this paper are presented with an effective SNR
axis, we use all four values, each in the appropriate SNR range.
For 802.11n, we also model MIMO processing at the receiver.
To do this we need to estimate the subcarrier SNRs for each spa-
tial stream from the channel state matrix H
s
. Although the stan-
dard does not specify receiver processing, we assume that a Min-
imum Mean Square Error (MMSE) receiver is used. It is compu-
tationally simple, optimal and equivalent to Maximal-Ratio Com-
bining (MRC) for a single stream, and near optimal for multiple
streams. All of these make it a likely choice in practice. The SNR
of the i
th
stream after MMSE processing for subcarrier s is given
by ρ
s,i
= 1/Y
ii
− 1, where Y =
H
H
s
H
s
+ I
−1
for i ∈ [1, N ]
and N xN identity matrix I [27]. For MIMO, the model computes
the effective BER averaged across both subcarriers and streams.
Coding interacts with the notion of effective SNR in a way that
is difficult to analyze. One challenge is that the ability to correct
bit errors depends on the position of the errors in the data stream.
To sidestep this problem, we rely on the interleaving that random-
izes the coded bits across subcarriers and spatial streams. Assum-
ing perfect interleaving and robust coding, bit errors in the stream
should look no different from bit errors for flat channels (but at a
4
50 ft
Figure 4: Our indoor 802.11n testbeds, T1 and T2. T1 consists of 10 nodes spread over 8 100 square feet, and T2 consists of 11 nodes
spread over 20 000 square feet. The nodes are placed to ensure a large number of links between them, a variety of distance between
nodes, and diverse scattering characteristics.
8
12
16
20
24
-28 -14 0 14 28
SNR (dB)
Subcarrier index
BPSK
QPSK
QAM-16
QAM-64
Packet SNR
Subcarrier SNRs
Figure 5: Sample faded link showing the packet SNR and ef-
fective SNRs for different modulations. BPSK has the lowest
effective SNR, but it needs less energy to decode.
lower SNR). Thus our estimate of the effective BER in Eq. (1) will
accurately reflect the uncoded error performance of the link. Our
algorithm now proceeds as in the case of a flat-fading channel de-
scribed above: we take the computed effective SNR value and use
the measurements from a flat-fading link (Figure 1(a)) to determine
transmission success or failure. As in CHARM [10], we support
different packet lengths with different SNR thresholds.
Note that this procedure differs from the typical approach of
simulation-based analyses [11, 15, 19], that instead map the un-
coded BER estimate such as we compute to a coded BER esti-
mate by means of a simple log-linear approximation. They then
use the coded BER estimate, and the length of the target transmis-
sion, to directly compute the packet delivery rate of the link. We
believe our method of thresholding the effective SNR is better be-
cause it directly accommodates variation in the receiver implemen-
tation. Different devices may have different noise figures, a measure
of how much signal strength is lost in the internal RF circuitry of
the NIC. They may implement soft Viterbi decoders with more or
fewer soft bits for their internal state, or indeed might do hard de-
coding instead. A receiver could use the optimal Maximum Like-
lihood MIMO decoder that has exponential complexity for small
constellations like BPSK, but revert to the imperfect but more ef-
ficient MMSE at higher modulations. All of these can be easily
expressed, albeit maybe approximately, as (perhaps modulation-
dependent) shifts in the effective SNR thresholds. In contrast, chang-
ing these parameters in the simulation approach involves changing
the internals of the calculation.
Protocol Details. Effective SNR calculations can be performed by
either receiver or transmitter, and each has advantages. For it to
make decisions, the transmitter must know the receiver’s thresholds
for the different rates; these are fixed for a particular model of NIC
and can be shared once, e.g., during association. The transmitter
also needs up-to-date CSI: either from feedback or estimated from
the reverse path. Alternately, the receiver can request rates and se-
lect antennas directly using the new Link Adaptation Control field
of any 802.11n QoS packet [1, §7.1.3.5a]. This obviates sending
CSI, but the calculation instead requires that the transmitter share
its spatial mappings, i.e. how it maps spatial streams to transmit an-
tennas. These are likely to change less frequently than the channel,
if at all. Finally, when operating in either mode with fewer trans-
mit streams than antennas, the transmitter must occasionally send a
short probe packet with all antennas to measure the full CSI.
Summary and Example. Combining the above steps, our model
consists of the following: 1) CSI is obtained and a test config-
uration is chosen; 2) the MMSE expression is used to compute
per-stream, subcarrier SNRs from the CSI for the test number of
streams; 3) the effective SNR is computed from the per-stream,
subcarrier SNRs for the test modulation; and 4) the effective SNR
is compared against the pre-determined threshold for the test mod-
ulation and coding to predict whether the link will deliver packets.
As an example, Figure 5 shows the CSI for a SISO link (steps 1–
2) as a fading profile across subcarriers, with the computed effective
SNRs for all modulations (step 3). These effective SNRs are com-
pared with pre-determined thresholds (step 4, see §5) to correctly
predict that the best working rate will be 39 Mbps. Note that these
effective SNRs are well below the RSSI-based packet SNR that is
biased towards the stronger subcarriers (note the logarithmic y-axis
scale). This link does a poor job of harnessing the received power
because it is badly faded, so its RSSI is a poor predictor of rate.
Applications can use this model to find useful configurations
without sending packets to test them. For example, the highest rate
can be predicted by running the model for all candidate rates and
selecting the best working rate. Alternatively, we could predict the
minimum transmit power to support a rate.
4. TESTBEDS
We conduct experiments using two stationary wireless testbeds
deployed in indoor office environments, T1 and T2 (Figure 4). T1
consists of 10 nodes spread over 8 100 square feet. T2 is less dense
by comparison with 11 nodes over 20 000 square feet. Each testbed
covers a single floor of a multi-story building and has a variety of
links in terms of maximum supported rate and line-of-sight versus
multi-path fading. We conduct mobile experiments using laptops
that interact with testbed nodes and are configured in the same way.
5
