1

Estimating Link Quality in 802.11 WLANs

Domenico Giustiniano∗ , David Malone† , Douglas J. Leith† and Konstantina Papagiannaki+
∗ Università degli Studi di Roma-Tor Vergata, Dipartimento di Ingegneria Elettronica; † Hamilton Institute, NUI

Maynooth, Ireland; + Intel Research, Pittsburgh.

Abstract— We propose a powerful new MAC/PHY cross-layer performance [7]. Similar problems occur in the presence of
approach to estimating link quality in 802.11 WLANs. Unlike pre- hidden nodes, e.g. see [8]. The availability of a measure of
vious approaches, we explicitly classify channel impairments into the loss rate specifically induced by channel noise would po-
noise-related losses, collision induced losses, hidden-node losses
and 802.11 impairments caused by exposed nodes and capture tentially allow much more effective rate adaptation algorithms
effects. Our approach distinguishes among these different types to be employed. Similarly, channel selection algorithms are
of impairments without requiring any modification to the 802.11 fundamentally related to channel impairments and typically
protocol and provides separate quantitative measures of the depend upon the availability of an appropriate link quality
severity of each one. Our approach is suited to implementation metric, which can then be optimised by a suitable search over
on commodity hardware and we demonstrate both a prototype
implementation and experimental assessments. available channels. Effective carrier sense adjustment is also
strongly dependent on link measurements.
The consideration of link quality measurements is par-
I. I NTRODUCTION ticularly topical since the trend towards increasingly dense
wireless deployments is creating a real need for effective
In this paper we consider how to estimate the link quality
approaches for channel allocation/hopping, power control, etc.
experienced by communicating stations in an 802.11 WLAN.
for interference mitigation [10], [11] while new applications
Link impairments (and so quality) are intimately linked to
such as mesh networks and media distribution within the home
MAC operation and so cannot be estimated purely on the
are creating new quality of service demands that require more
basis of PHY measurements such as signal-to-noise ratio
sophisticated approaches to radio resource allocation [12].
(SNR). High level measurements such as throughput and
In this paper we propose a powerful new MAC/PHY cross-
delay statistics are can have difficulty distinguishing between
layer approach to estimating link quality in 802.11 WLANs.
sources of channel impairment. Instead, a MAC/PHY cross-
Unlike previous approaches, we explicitly classify channel
layer approach is essential to understand the actual channel
impairments into noise-related losses, collision induced losses,
status and the impact of different performance impairments.
hidden-node losses and consider related issues of exposed
This can be readily seen, for example, from the fact that
nodes and capture effects. Our approach distinguishes among
frame loss due to collisions is a feature of normal operation
these different types of impairments and provides separate
in 802.11 WLANs and thus we need to distinguish losses due
quantitative measures of the severity of each type of impair-
to collisions and losses due to channel impairment. Similarly,
ment. We thus make available new measures that we expect
hidden nodes effects, exposed nodes, capture effects etc are
to be of direct use for rate adaptation, channel allocation, etc.
all associated with cross-layer issues.
Since we take advantage of the native characteristics of the
Despite the resulting difficulty of measuring link quality, the
802.11 protocol (such as timing constraints, channel busy de-
potential benefits arising from the availability of accurate and
tection and so on) — without requiring any modification to the
reliable link quality data are considerable. Tasks such as rate
802.11 protocol — our approach is suited to implementation
adaptation, channel allocation, contention window selection,
on commodity hardware and we demonstrate both a prototype
power control and carrier sense selection — essential for
implementation and experimental measurements. Indeed we
improving and optimizing the network performance — all
argue that it is vital to demonstrate operation in a real radio
depend crucially on the availability of suitable link quality
environment not only because of the difficulty of developing
measurements, and it is the current lack of such measurements
realistic radio propagation models but also because important
that underlies the poor performance of many approaches
impairments such as hidden-nodes and capture effects are af-
currently implemented in commodity hardware. For example,
fected by low-level issues (e.g. interactions between amplifier
at present rate adaptation is in practice commonly based on
and antenna design as well as radio propagation) that are
the number of transmission retries (e.g. a typical approach
difficult to model in simulations. We note that many of the
might involve lowering the rate after n retries and increasing
measurements presented are new and of interest in their own
the rate after m successful transmissions). However, since the
right.
number of retries is affected not just by channel noise but is
The paper is organized as follows. In Section II we review
also closely linked to the number of contending stations (with
related work and in Section III briefly review the 802.11 MAC
associated collision related losses), this can easily lead to poor
and then categorize the main link impairments. In Sections IV
We gratefully acknowledge the help of Richard Gass at Intel. Supported and V we introduce our estimation approach. We describe our
by Science Foundation Ireland grants IN3/03/I346 and 07/IN.1/I901. testbed setup in Section VI and present extensive experimental
 2

DIFS
measurements in Section VII evaluating this approach in a Time Slot
wide range of real radio environments. Finally we summarize PIFS
our conclusions in Section IX and give some insight on hidden DIFS SIFS
Busy medium TX Frame
node interference estimate in the appendix.
Defer Access Backoff
II. R ELATED WORK
Previous work on 802.11 channel quality estimation can
Fig. 1. DCF protocol summary.
be classified into three categories. First, PHY link-level ap-
proaches use SNR/RSSI to directly estimate the link quality.
Second, MAC approaches rely on throughput and delay statis- physical slot time. It aims to detect transmissions within
tics, or frame loss statistics derived from tranmsitted frames the interference range.
which are not ACKed and/or from signaling messages. Finally • NAV (Network Allocation Vector) timer at MAC level
cross-layer MAC/PHY approaches aim to combine information which is encapsulated in the MAC header of each 802.11
at both MAC and PHY layersl. frame and is used to accurately predict the end of a
Most work on PHY layer approaches is based on SNR received frame on air. It is naturally updated once per
and RSSI measurements [13], [14]. The basic idea is to a- packet and can only gather information from stations
priori map SNR measures into MAC channel quality estimates. within the decoding range. This method is also called
However, i) SNR/RSSI methods are not able to distinguish virtual carrier sense.
between different sources of channel impairment at the MAC The channel is detected as idle if the CCA detects the channel
layer (e.g. between collision and noise related losses), ii) as idle and the NAV is zero. Otherwise, the channel is detected
the mapping between measured SNR and delivery probability as busy. A station transmits when the backoff counter reaches
rate is generally specific to each link [9] and may be time- zero. The countdown process is illustrated schematically in
varying iii) the correlation between SNR/RSSI and actual Fig. 1. The 802.11 handshake imposes a half-duplex process
packet delivery rate can be weak [24]. whereby an acknowledgment (ACK) is always sent by the
With regard to MAC approaches, RTS/CTS signaling can be receiver upon the successful receipt of a unicast frame. The
used to distinguish collisions from channel noise losses [3], ACK is sent after a period of time called the Short InterFrame
[20]. However, such approaches can perform poorly in the Space (SIFS). As the SIFS is shorter than a DIFS, no other
presence of hidden nodes and other types of channel impair- station is able to detect the channel idle for a DIFS until the
ment. [22] considers an approximate MAC layer approach for end of the ACK transmission. If the transmitting station does
detecting the presence of hidden nodes but does not consider not receive the ACK within a specified ACK Timeout, or it
other types of channel impairment. detects the transmission of a different packet on the channel,
With regard to combined MAC/PHY approaches, early work it reschedules the packet transmission according to the given
related to the present paper is presented in [16], [17]. However, backoff rules. CW is doubled with successive referrals until a
this uses a channel busy/idle approach that is confined to maximum value (labeled as CWmax ) and is reset to the mini-
distinguishing between collision and noise related losses and mum value (labeled as CWmin ) after an ACKed transmission
does not allow consideration of hidden nodes or exposed node or once the maximum number of retransmission attempts is
and capture effects. reached.
In addition to the foregoing Basic Access mode, an op-
III. CSMA/CA P ROTOCOL AND L INK I MPAIRMENTS tional four way handshaking technique, known as Request-
To-Send/Clear-To-Send (RTS/CTS) mode is available. Before
A. CSMA/CA protocol
transmitting a packet, a station operating in RTS/CTS mode
In 802.11 WLANs, the basic mechanism controlling reserves the channel by sending a special Request-To-Send
medium access is the Distributed Coordination Function short frame. The destination station acknowledges the receipt
(DCF). This is a random access scheme, based on Carrier of an RTS by sending back a Clear-To-Send frame, after which
Sense Multiple Access with Collision Avoidance (CSMA/CA). normal packet transmission and ACK response occurs.
In the DCF Basic Access mode, a station with a new packet to The DCF allows the fragmentation of packets into smaller
transmit selects a random backoff counter in the range [0,CW- units. Each fragment is sent as an ordinary 802.11 frame,
1] where CW is the Contention Window. Time is slotted which the sender expects to be ACKed. However, the frag-
and if the channel is sensed idle the station first waits for ments may be sent as a burst. That is, the first fragment
a Distributed InterFrame Space (DIFS), then decrements the contends for medium access as usual. When the first fragment
backoff counter each PHY time slot. If the channel is detected is successfully sent, subsequent fragments are sent after a
busy, the countdown is halted and only resumed after the SIFS, so no collisions are possible. In addition, the medium is
channel is detected idle again for a DIFS. Channel idle/busy reserved using virtual carrier sense for the next fragment both
status is sensed via: at the sender (by setting the 802.11 NAV field in the fragment)
• CCA (Clear Channel Assessment) at physical level which and at the receiver (by updating the NAV in the ACK). This is
is based on a carrier sense threshold for energy detection, illustrated schematically in Fig. 2. Burst transmission is halted
e.g. −80dBm. CCA is expected to be updated every after the last fragment has been sent or when loss is detected.
 3

Original Frame around 40% (numbers from the model in [6]). We denote by
pc the probability that a transmitted data frame is lost due to
a collision.
HDR Frame Body CRC HDR Frame Body CRC 2) Hidden nodes: Frame corruption due to concurrent
transmissions other than collisions are referred to as hidden
Fragment 0 Fragment 1
node interference. We denote by ph,data the probability that
Fig. 2. Fragmentation of a 802.11 Frame. a data transmission fails to be received correctly due to
hidden node interference. Similarly, we denote by ph,ack the
probability that an ACK transmission is lost due to hidden
B. Link Impairments node interference. A lost data packet or a lost ACK both lead
to a failed transmission and so we combine data and ACK
In this section we categorize the main impairments that can
losses into an overall hidden node error probability ph .
affect transmissions between an 802.11 sender and receiver.
3) Noise errors: Frame corruption due to sources other
Before proceeding, it is important to emphasize that a two-
than transmissions by other 802.11 stations are referred to as
way (or four-way with RTS-CTS) handshake is used in 802.11.
noise losses. We denote by pn,data (respectively, pn,ack ) the
Hence, the quality of a link is determined by the channel
probability that a data (respectively, ACK) frame is lost due to
conditions at both the sender and the receiver stations. For
noise related errors. Since data and ACK losses both lead to
example, low link-quality at the receiver can mean that data
a failed transmission we lump these together into a combined
packets transmitted by the sender cannot be decoded at the
noise loss probability pn .
receiver. Similarly, low link-quality at the sender can mean that 4) Exposed nodes: Not all link impairments lead to frame
ACK packets transmitted by the receiver cannot be decoded loss. One such important issue is that the carrier sense mecha-
at the sender. It follows immediately that: nism used in 802.11 to sense channel busy conditions may
• Measuring the SNR (or other local properties) at either incorrectly classify the conditions. We denote by pexp the
the sender or receiver alone is insufficient to determine probability that a slot is erroneously detected as busy when
the link quality. Instead it is necessary to recognize the in fact a successful transmission could have been made. Such
intrinsically two-way nature of a link in 802.11 when errors lead to an unnecessary pause in the backoff countdown
measuring its quality. and so to a reduction in achievable throughput.
• Links are directional since data packets and ACKs may 5) Capture effect: A second impairment which does not
have different properties e.g. coding rate, duration, NAV cause losses is the so-called physical layer capture (PLC).
protection. Collisions and interference with transmissions Specifically, we denote by pplc the probability of successful
by other stations can therefore affect each end of a link reception of a frame when a collision occurs. This can occur,
differently. for example, when the colliding transmissions have different
• Since each station is typically located in a different received signal power — the receiver may then be able to
physical position, its local radio environment is generally decode the higher power frame. For example [15] shows
different from that of other stations. Hence we need to that for 802.11b PLC can occur when a frame with higher
measure the link quality between each sender-receiver received power arrives within the physical layer preamble of
pair individually. In particular, we cannot reliably infer a lower power frame. Our measurements — not shown here
the properties of one link from measurements taken on for lack of space — have confirmed this finding and found a
another link, even if the links share a common sender e.g. similar behavior for 802.11g. Differences in received power
the AP in an infrastructure mode WLAN. Further, due can easily occur due to differences in the physical location of
to the directional nature of link quality (see above) we the transmitters (one station may be closer to the receiver than
need to measure quality in each direction separately and others), differences in antenna gain etc. The physical layer
generally cannot use measurements from one direction capture effect can lead to severe imbalance of the network
to reliably infer the quality in the opposite direction. An resource and hence in the thoughputs achieved by contending
example illustrating this is shown later in the paper, see stations (and so to unfairness).
section VIII-B.
As we will see, the manner in which link impairments are IV. E STIMATING L INK Q UALITY
manifested is closely linked to the interaction between MAC Our aim is to develop an estimation framework capable
and PHY operation. We distinguish five main types of link of distinguishing the different types of link impairment and
impairment when using the 802.11 DCF. providing quantitative measurements of link quality. To do this
1) Collisions: Collisions are part of the correct operation of we make use of the key observation that these impairments are
CSMA/CA. A collision occurs whenever two or more stations intimately related to MAC operation. We therefore exploit the
have simultaneously decremented their backoff counter to 0 flexibility already present in the 802.11 MAC to enable us to
and then transmit. Note that collisions can only occur on data distinguish the impact of the different impairments.
packet transmissions. The level of collision induced packet Specifically, we make use of the following properties of the
losses is strongly load dependent. For example, 802.11b with 802.11 MAC:
four saturated nodes has a collision probability of around • Time is slotted, with well-defined boundaries at which
14% while with 20 saturated nodes the collision probability is frame transmissions by a station are permitted.
 4

• The standard data-ACK handshake is affected by all types B. Estimating Hidden Node Interference
of link impairment considered and a sender-side analysis We now require to distinguish frame losses due to hidden
can reveal any loss. node interference. To achieve this we exploit the fact that
• When fragmentation is enabled, second and subsequent
frames transmitted after a PIFS are protected from collisions
fragment transmissions are protected from collisions and since other transmissions must defer for a DIFS interval after
hidden nodes by the NAV values in the fragments and sensing the channel to be idle, with DIFS > PIFS. Although
ACKs. We treat hidden nodes that are unable to decode the PCF element is rarely implemented in 802.11 hardware,
either NAV value as channel noise. Instead of using the ability to transmit after a PIFS is commonly supported.
fragments, we could use TXOP packet bursting is used, Losses on PIFS frames are due either to noise or hidden node
although this is only available in 802.11e [2], and would interference. That is,
require the NAV value in the MAC ACK to be set.
RTS/CTS might also be used, but in practice can perform P[PIFS success] = A1 /T1 = (1 − ph )(1 − pn ), (3)
poorly — see the appendix.
• Transmissions occurring before a DIFS are protected where the station transmits T1 data frames after a PIFS and
from collisions. This is used, for example, to protect of these A1 are successful because an ACK is received. We
ACK transmissions, which are transmitted after a SIFS can now use our estimate of pn (based on fragment loss
interval. The 802.11 DCF also permits transmissions after measurements, see equation (2)), to allow estimation of the
a PIFS interval (with SIFS < PIFS < DIFS) and while probability ph of hidden node losses as:
the full 802.11 Point Coordination Function (PCF) is
rarely implemented on commodity interface cards, the ph = 1 − (A1 · TS )/(AS · T1 ) (4)
ability to transmit after a PIFS is widely available on
modern hardware (e.g. as part of the so-called multi- C. Estimating Collision Rate
media extensions that are a subset of 802.11e).
In the following sections we consider in more detail how Consider a station sending ordinary data packets (i.e. sent
these properties can be exploited to obtain powerful new after DIFS and not fragmented) to a given receiver. Suppose
measurements of link quality. that over some time period the station contends and transmits
data frames T0 times and of these A0 are successful because an
A. Estimating Noise Errors ACK is received. As discussed previously, the possible sources
of frame loss are: collisions, hidden nodes and noise errors.
Consider a station sending fragmented packets to a given
Assuming that these sources of frame loss are independent, if
receiver. Each fragment is immediately acked by the receiver
the station transmits the probability of success over the link
when it arrives, allowing detection of loss. Fragments are
is:
sent back to back with a SIFS interval between them. Hence,
second and subsequent packets are protected from collisions. P[success] = A0 /T0 = (1 − pc )(1 − ph )(1 − pn ). (5)
Importantly, fragment ACK frames update the NAV and so the
fragment-ACK handshake is akin to an RTS-CTS exchange Finally pc can be estimated from Eq. (5) and (3):
from the point of view of hidden nodes1 . Hence, second and
subsequent fragments are also protected from hidden node pc = 1 − (T1 · A0 )/(T0 · A1 ). (6)
collisions. That is, while the first fragment will be subject to
collisions, noise and hidden node errors, subsequent fragments V. I MPAIRMENTS THAT DO NOT LEAD TO FRAME LOSS
are only subject to noise errors and we have that
Section IV presents a straightforward approach for esti-
P[fragment success] = AS /TS = (1 − pn ), (1) mating the magnitude of those link impairments that lead to
where the station transmits TS second and subsequent data frame loss, namely collisions, hidden nodes and noise. The
frames and of these AS are successful because an ACK is estimates require only very simple measurements that are read-
received. We can therefore directly estimate the probability of ily available on commodity hardware. In this section we now
noise errors pn from the fraction of second and subsequent consider methods for estimating capture and exposed node
fragments with no ACK, effects. These impairments do not lead directly to frame losses,
but can nevertheless lead to unfairness in throughput/delay
pn = 1 − AS /TS (2)
between interfering stations.
Since the impact of noise losses is dependent on frame In order to estimate capture and exposed node effects we
length (longer frames typically having higher probability of make use of additional measurements. In particular, measure-
experiencing bit errors), we must select the fragment size to ments of channel idle and busy periods. Here idle/busy refers
be equal to the packet size used for regular data transmissions. to time as measured in MAC slots rather than in PHY slots. In
The frame loss rate estimated from fragment measurements the next section we discuss MAC slots in more detail. Then we
can then be reliably applied to estimate the loss rate for other discuss estimating capture and exposed node effects. Note that
transmissions. while these additional measurements offer further insight into
1 As already mentioned, we do not rely on RTS/CTS since it can perform the wireless environment, they are not necessary to estimate
poorly, see appendix. the basic quantities pc , pn and ph .
 5

A. MAC slots 1 2 3
Tx_succ
4 5 6 7
Other
8 9
Tx_unsucc
10
Other
11 12 13
Tx_succ
14 15 16
Other
17 18

The slotted CSMA/CA process creates well-defined bound- MAC Slots

aries at which frame transmissions by a station are permitted. Fig. 4. MAC slot boundaries at which transmissions are permitted. Different
The time between these boundaries we call MAC slots (as types of MAC slot are possible: idle slots (corresponding to PHY slots),
distinct from PHY slots). Considering operation from the busy slots due to transmissions by other stations (marked “Other”) and busy
slots due to transmissions the station of interest (marked “Tx ”). “Other”
viewpoint of a station, say station 1, we have the following transmissions include both successful and unsuccessful transmissions.
possibilities:
1) Station 1 has transmitted and received an ACK. We call
these slots successful transmissions. In effect we are estimating the number of collisions losses
2) Station 1 has transmitted, timed-out while waiting for an that we expect based on the carrier sense environment and
ACK and is about to resume its backoff. We call these comparing it with the actual collision rate. The discrepancy, if
slots unsuccessful transmissions. any, provides a measure of exposed node and capture effects –
3) Station 1 has seen the medium as idle and, if backoff both of which are associated with apparently busy slots during
is in progress, has decremented its backoff counter. We which a successful transmission can in fact take place.
call these idle slots. Note that the idle/busy measurements can also be used to
4) Station 1 has detected the medium as busy due to one estimate the collision probability when there are no exposed
or more other nodes transmitting, and has suspended its node or capture effects — see [16] and [17] — but this is not
backoff until backoff can resume. We call these slots possible in the more general setting considered here.
other transmissions, and include both successful and
unsuccessful transmissions of other stations. Note that VI. I MPLEMENTATION ON COMMODITY HARDWARE AND
each busy period is counted as a single slot, so these T ESTBED S ETUP
busy slots are closer to the MAC’s view than the PHY’s.
These events are illustrated (not to scale) in Fig. V-B. Trans- A. Implementation
missions by station 1 are only permitted at event boundaries. We have implemented the foregoing estimators using a com-
We also make the following assumptions: bination of driver and firmware modifications to commodity
Assumption 1. The probability that at least one other station network cards using the Atheros AR5212/AR5213 and Intel
transmits in an arbitrary slot does not depend on whether 2915ABG chipsets.
station 1 transmits or not. The proposed estimators are summarised in Table 3. The
Assumption 2. The collision probability is independent of the estimators of collision rate, hidden node and noise errors
backoff stage of station 1. described in Section IV can be implemented via straight-
With these assumptions, the probability of a collision is then forward driver modifications. In our work they have been
precisely the probability that at a slot boundary the channel is mainly tested on Atheros cards and the widely used MADWiFi
busy due to a transmission by one or more other stations. driver. To transmit frames after a PIFS interval we made use
We note that Assumptions 1 and 2 are reasonable in a of the WME (Wireless Multimedia Enhancements) features,
distributed random access MAC scheme such as CSMA/CA which allow dynamic adjustment of the TXOP, CWmin and
and, indeed, these assumptions are central to well-established AIFS parameters for each Access Category of 802.11e. In
models of 802.11 operation such as that of Bianchi [6] and particular, we created an access category with MAC settings
others (e.g. the nonsaturated heterogeneous model in [26]). CWMin=CWMax=AIFSN=TXOP=0. All traffic sent via the
queue associated with this access category is then transmitted
B. Capture and Exposed Nodes
using PIFS. A second access category and queue is defined for
Suppose there are R MAC slots in which our station does normal traffic. On this queue, data packets are fragmented in
not transmit and that I of these are idle. These quantities can two fragments, which is sufficient for assessing our estimator.2
be measured by appropriate sensing of the channel idle/busy By appropriately directing packets to these two queues we can
status. The classification of a MAC slot as idle/busy relies on collect statistics for the overall number of transmissions T0 ,
carrier sensing, using both carrier sensing mechanisms. Hence, T1 and TS and number of successful transmissions A0 , A1
this measurement is affected by exposed nodes and capture and AS (transmissions for which a MAC ACK is received).
effects whereby the carrier sense indicates that the channel In our implementation packets are allocated between queues
busy when in fact a transmission would be successful. at driver level, although other solutions are possible.
We therefore have that, The estimators in Section V require measurement of the
R−I number of R and I busy and idle MAC slots. This requires
pc + pexp + pplc = , (7)
R carrier sense information from the hardware. We modified
where pc is the collision probability, pexp the probability that the card firmware and microcode on cards using the Intel
the channel is sensed busy due to exposed node behavior and 2915ABG chipset to perform the necessary measurements and
pplc the probability that the channel is sensed busy due to to expose these to the driver. Our implementation implicitly
capture effects . Combining our estimate of pc from eq. (6) uses the same carrier-sense threshold as the rest of the MAC.
with the additional information in (7), we can estimate:
2 Note that other traffic configurations are possible, e.g. to fragment only
pexp + pplc = (T1 · A0 )/(T0 · A1 ) − I/R. (8) the PIFS traffic.
 6

Successful and unsuccessful TX slot counters Idle and other transmissions slot counters
T0 TX of normal traffic
T1 TX of PIFS traffic, first frag.
TS TX of subsequent frag.
A0 ACK of normal traffic
A1 ACK of PIFS traffic, first frag.
AS ACK of subsequent frag.
I idle slots
R slots we do not TX in
Probability of Estimator
pc collision 1 − (T1 · A0 )/(T0 · A1 )
pn noise interference err. 1 − AS /TS
ph hidden node err. 1 − (A1 · TS )/(AS · T1 )
pexp + pplc exposed and capture effect (T1 · A0 )/(T0 · A1 ) − I/R
Fig. 3. Summary of measurements used and proposed estimators.

We will also cross-validate a number of our results based on ST1

the number of CRC errors, CRCerr, observed at a receiving ST2

STA. This counter has been also retrieved from the microcode AP1 H H−RX

in Intel cards, and driver code in Atheros cards. This cross-

ST3
validation is described in detail in Section VI-C.
Fig. 5. Topology for hidden node tests.
B. Testbed setup
To evaluate the estimators we performed experimental mea- ST2
surements over a wide range of network conditions, of which
ST2 AP1 ST1 AP2
we present a subset here. Our testbed consists of Soekris
net4801 devices running Linux and configured in infrastructure ST3

mode. Stations transmit 1400 byte UDP packets to an AP

equipped with a NIC using the Intel 2915ABG chipset or Fig. 6. Topology for exposed node tests.
Atheros AR5213 chipset, according to the specific test. Unless
otherwise specified, the physical rate is set to 6 Mbps in each
station, time slots are set to 20 µs on both Intel and Atheros
isolation; the link from the hidden node to the hidden
NICs and the carrier sense threshold for the Intel NICs was set
receiver is of high quality in isolation; a link can not
to −80dBm, while the carrier sense level used with the Atheros
be established from the transmitter to the hidden node;
NICs is the default value (set in the binary component —
losses occur when the hidden node operates at the same
HAL — of the Atheros MADWiFi driver, and thus not acces-
time as the transmitter.
sible/modifiable). In all experiments, automatic rate selection
• Exposed nodes Exposed nodes are investigated via a
and the RTS/CTS mechanism are disabled unless otherwise
setup with up to two interfering WLANs, as depicted
stated. Antenna diversity functionality is also disabled (see
in Fig. 6. In more detail, ST 1 and ST 2 are associated
[25] and therein), together with any proprietary mechanisms at
to AP 1 (WLAN 1), while ST 3 and ST 4 are associated
MAC level. External interference levels are measured using a
to AP 2 (WLAN 2). In WLAN 1 we verify that i) ST 1
spectrum analyzer. Link impairments are generated as follows:
receives the signals from WLAN 2 (ST 3 and ST 4) at
• Noise errors In the testbed we modify the signal-to- higher strength than the carrier sense threshold ii) the
noise ratio of a link by a combination of adjusting the ST 1 → AP 1 link3 is of higher signal quality than the
physical separation of stations and/or adjustment of the ST 3 → AP 1 and ST 4 → AP 1 links, so that AP 1 may
transmit power used. In this way we can roughly control successfully decode any signal from ST 1, despite the
conditions to allow investigation of the ability of the interference from WLAN 2.
proposed estimator to measure the level of frame losses • Capture effects Capture effects are studied using the setup
due to noise errors on a link. illustrated in Fig. 7. Two stations ST 1 and ST 2 are
• Collisions The level of collision induced losses is adjusted associated to AP 1. We verify that the ST 1 → AP 1 link
by varying the number of contending stations and their is of higher signal quality than the ST 2 → AP 1 link
offered traffic load. such that transmissions by ST 1 are successfully received
• Hidden nodes Hidden node effects are evaluated using at AP 1 even when they collide with transmissions by
scenarios based on the setup illustrated in Fig. 5. We ST 2 i.e. ST 1 can capture the channel.
have a number of transmitting nodes and a receiver.
The hidden node transmits to an independent receiver.
We ensure that the following conditions hold: the link
from the transmitter to our receiver is of high quality in 3 We denote by A → B a link with data sent from A to B
 7

ST2
PSDU, and so we assume that channel noise never
AP1
results in a PHY error, but instead results in a CRC
ST1
error.
3) Hidden nodes Finally, consider the impact of hidden
Fig. 7. Topology for physical layer capture tests. nodes. The receiver will see a certain number of hidden
node errors as simple collisions, when a hidden node
CRC32 error and a ordinary node select the same slot, as illustrated
at point 1 in Fig. 8. These will contribute to pc . However,
MAC MSDU CRC hidden-node transmissions beginning in later slots (i.e.,
Header
after an ordinary node has already started) may result in
more complex errors. In our experiments we use 802.11g
packet 1 PLCP PSDU transmissions with a PLCP of 20µs and the 802.11b
compatible slot length of 20µs. For this setup, shown in
Fig. 8, we expect all of the hidden node errors that are
packet 2 (Hidden) PLCP PSDU
not simple collisions to result in CRC errors, because
the hidden node will not transmit until after the PLCP
1 2 3 4 5 6 7 has been transmitted.
PHY slots(20µsec) Thus, the CRC errors seen at the receiver satisfy:
Fig. 8. Hidden node errors for an 802.11 frame (not to scale). CRCerr
= pn + ph + pc2 − (pn + ph )pc1 (10)
R−I
−(pn + ph )pc2 ≈ pn + ph + pc2 (11)
C. Cross-Validation of Frame Loss Impairments
To help validate the sender-side link quality measurements where CRCerr is the number of CRC32 errors and R − I is
obtained using the estimator in the previous section, in our the number of busy MAC slots seen at the receiver.
experimental tests we also make use of the following inde-
pendent measurements, obtained at the receiver-side. VII. E XPERIMENTAL ASSESSMENT
The 802.11 frame consists of a PLCP (Physical Layer In this section we present experimental measurements to
Convergence Preamble) and MAC payload called the PSDU explore the practical utility of the proposed estimators. We
(Physical Service Data Unit). Each PSDU is protected with argue that experimental testing is vital when assessing link
a 32 bit Cyclic Redundancy Check (CRC checksum). At the quality estimators since issues such as complex radio propa-
PHY level, errors in frame reception can be classified as either gation effects, real antenna behavior, front-end amplifier issues
PHY or CRC errors: etc can all have an important impact on performance yet are
• an error occurs on the PLCP preamble or header. We call difficult to capture accurately in simulations. Experimental
these PHY errors. testing also highlights implementation issues, demonstrates
• the PLCP is correctly decoded but the PSDU CRC fails: the practicality of operation on commodity hardware, and
we call this a CRC32 error. Note that the presence of generally helps to build greater confidence in the viability of
a CRC32 error notification on a received frame implies the proposed approach.
that no errors occurred in the PLCP.
In the present work we analyze the count of CRC32 errors
A. Collisions only, no noise, no hidden nodes
for our validation measurements, that is we consider when
collisions, channel noise and/or hidden nodes result in CRC We begin by considering a simple scenario with a clean
errors: channel and no hidden nodes. A low level of RF interference
1) Collisions First, note that in a collision two or more is confirmed by spectrum analyzer. We vary the number of
transmit stations have chosen the same PHY slot to start contending wireless stations so as to vary the collision rate.
transmission. We assume that a receiver station will not Each station generates traffic at a rate of 300 fps (frame
only observe this as a busy slot, but that it will also per seconds), which is sufficient to saturate the network, for
detect either a PHY error or, in the case of physical an interval of 600s. 10% of the transmit traffic is generated
layer capture in the PLCP, a CRC error. We split the through the PIFS queue, while the rest is sent through the BE
probability of collision, queue.
Fig. 9 shows the measured estimates of pc , ph , and pn ,
pc = pc1 + pc2 , (9) averaged over the experiment. We can immediately make a
where pc1 is the probability of a collision resulting in a number of observations:
PHY error and pc2 the probability of a collision resulting • The collision probability pc increases with the number of
in a CRC error. Thus pc2 collisions will be observed by stations, as expected.
the CRC estimator. • The noise loss probability pn , estimated from measure-
2) Noise errors Second, consider channel noise. Typically ments on subsequent fragments, is negligible, as ex-
the PLCP is sent at a substantially lower rate than the pected.
 8

Estimator in absence of Interference 1 transmit node with a low snr link

40 100
pc Normal traffic, first fragment
ph Normal traffic, second fragment
35 pn

Probability of Error (%)

PIFS traffic
Estimator Value (%) 80
30

25 60
20

15 40

10
20
5

0 0
1 2 3 4 5 6 7 50 100 150 200 250 300
Number of Stations Time (sec)
(a) Measured loss rate of first and second fragments
Fig. 9. Estimates of pc , ph , and pn vs. number of contending stations. Clean and PIFS traffic.
channel, no hidden nodes.
Estimators with a link with low SNR
80
tx1,err
tx2,err
• The hidden node loss probability ph is consistently low, rx1,err
70 rx2,err

Estimators Value (%)

as expected.
Although a simple test scenario, it is nevertheless encourag- 60
ing that these initial tests indicate correct operation of the
estimators. In particular, the ability to distinguish collision 50
losses from noise and hidden node effects. We confirm this
in more detail in the following sections by varying the level 40

of noise and hidden node losses over a wide range of operating

conditions. 30
10 20 30 40 50 60 70 80 90 100
Time (sec)
(b) Cross-validation of measured noise loss rate.
B. Channel noise only, no collisions, no hidden nodes
Fig. 10. Mesured loss rates for Low SNR link, no collisions, no hidden
To explore the impact of channel noise, we begin in this nodes. tx1,err is loss rate for first fragment transmissions, tx2,err loss rate
section by considering a setup with one transmitting and one for second fragments (an estimate of pn ), rx1,err the error rate measured at
receiving station and thus no collisions or hidden nodes (more the receiver for first fragments, rx2,err the rate for second fragments.
complex setups with noise, collisions and hidden nodes are
considered in later sections). The transmit physical rate is
fixed to 12 Mbps and sending rate at 300fps, which saturates and tx2,err measurements (operating at the transmitter). As
the transmit queue. The link is adjusted to have low SNR expected, Fig. 10(b) shows that the two estimators report very
and thus a high noise error rate, according to the testbed similar statistics for first and subsequent fragments, as the only
setup described in section VI-B. Recall that noise losses are errors present are noise errors4 .
measured via the loss rate for subsequent fragments. Fig. 10(a)
plots the measured loss rate for first and second fragments on
C. Hidden nodes only, no collisions, no noise
normal traffic and PIFS traffic. It can be seen that the loss
rates are all similar, as expected in the absence of collisions We now consider estimation of hidden node losses, again
and hidden nodes. This data also helps to confirm that the loss starting with a simple setup in this section in order to help
rate measured on second fragments is a good indicator of the gain clear insight into performance but with more complex
noise loss rate experienced by other types of traffic. situations considered in later sections.
As further validation of correct operation of the estimator, Fig. 11 reports the experimental results for a setup with only
we classify the loss percentage of transmitted/received frames, one transmitter and one receiver (and so no collisions) and with
respectively, one hidden node, the offered load at the transmitter and hidden
• tx1,err = (T0 −A0 )/T0 i.e. the loss rate for first fragment node being 300fps. As before, measurements at the transmitter
transmissions are validated against independent measurements taken at the
• tx2,err = (TS − AS )/TS i.e. the loss rate for second and receiver. It can be seen that while the first fragment in a burst
subsequent fragments experiences a high error rate, the second fragment has a very
• rx1,err = CRCerr0 /(R − I) i.e the rate of CRC errors low error rate. That is, as we expect, hidden node errors are
at the receiver for first fragments (CRCerr0 ) limited to the first fragment sent in a burst, while second
• rx2,err = CRCerrS /(R − I). i.e the rate of CRC errors fragments are protected from these errors. It is interesting to
at the receiver for subsequent fragments (CRCerrs ). observe that in this experiment the channel characteristics were
The measurement tx2,err is our proposed estimator for pn ,
4 Note that for this validation the receiver needed to use fragment and retry
the frame loss rate due to noise errors. Note that the rx1,err
bits in the PSDU to distinguish first and subsequent fragments. These bits
and rx2,err measurements are obtained by an entirely inde- may have been corrupted. Interestingly, despite the uncertainty in these bits,
pendent estimator (operating at the receiver) from the tx1,err the estimates are quite satisfactory.
 9

Estimators with a Hidden node Estimator with 2 transmitting nodes (one hidden)
50
tx1,err ph
tx2,err 100 pc
rx1,err pc (ph=0,pn=0)
40 rx2,err pn
Estimators Value (%)
80

30
60

%
20
40

10 20

0 0
10 20 30 40 50 60 70 80 90 100 20 40 60 80 100 120 140
Time (sec) Time (sec)
(a) Hidden node and one transmitting station.
Fig. 11. Hidden nodes, clean channel, no collisions. tx1,err is loss rate
for first fragment transmissions, tx2,err loss rate for second fragments (an Estimator with 3 transmitting nodes (one hidden)
estimate of pn ), rx1,err the error rate measured at the receiver for first 100
ph
fragments, rx2,err the rate for second fragments. pc
pc (ph=0,pn=0)
80 pn

Probability (%)
slowly varying, as can be seen from the peak in loss rate after 60

around 30s.
Note that the transmitter and receiver estimators report 40

different error rates. This can be explained as follows: while

20
measurements indicate that the number of CRC errors mea-
sured at the receiver is roughly the same as the number of
0
retries measured at the transmitter, the number of busy slots 20 40 60 80 100 120 140
Time (sec)
is measured to be higher at the receiver because the hidden (b) Hidden node and two transmitting stations.
node’s transmissions can be heard at the receiver.
Fig. 12. Estimator values for pc , ph and pn in the presence of collisions,
hidden nodes and high SNR (low noise).
D. Collisions and hidden nodes, no noise
Having validated the individual components of the estimator
in basic scenarios, we now consider more complex situations caused by the third station transmissions, which are overheard
with a mix of link impairments. In this section we consider a by both hidden nodes, thus decreasing the number of hidden
link with both collision losses and hidden node interference. node transmissions and hence the hidden node interference
In the experiments, the offered load at all stations is 300fps. probability.
Firstly, we again use a setup with a pair of stations that
behave as hidden nodes transmitting to one AP. Fig. 12(a)
plots estimates of pc , ph , and pn locally measured on one of
the hidden node stations. It can be seen that ph is estimated
at a high value, as expected due to the severe hidden node E. Collisions, hidden nodes and noise
interference in this example. The noise loss rate pn is correctly
estimated as being close to zero. The collision loss rate pc is Finally, we consider a link suffering from all three loss
correctly estimated at a value very close to that measured with inducing impairments: collisions, noise and hidden node in-
two contending stations and no noise or hidden nodes (marked terference. The scenario is illustrated in Fig. 13. We have
as pc (ph = 0, pn = 0) in the figure, with the value taken from three contending stations (stations 1, 2 and H), a pair of
the measurements in Fig. 9). This is an encouraging result as which behave as hidden nodes (stations 1 and H), and with
it clearly demonstrates the ability of the proposed estimation a noisy channel between station 1 and its receiving station.
approach to effectively distinguish the different sources of Each station sends saturated traffic. Measurements gathered
frame loss, even under complex conditions. on station 1 are summarized in Fig. 14. It can be seen that
Fig. 12(b) plots similar measurements, but now with a pair the collision loss rate pc is estimated at a value very close to
of stations that behave as hidden nodes plus one station which that measured with three contending stations and no noise or
can be heard by all the other stations, for a total of three hidden nodes (marked as “pc (ph = 0, pn = 0)” in the figure
contending stations with saturated traffic. Again, the noise with the value taken from the measurements in Fig. 9). That
loss rate pn is correctly estimated as being close to zero and is, the estimator is able to successfully distinguish collision
the collision loss rate is correctly estimated as being close related losses from noise and hidden node related losses. It
to that with three stations and no hidden nodes (marked on can also be seen from the figure that there is a high level of
plot, with value taken from Fig. 9). The hidden node loss errors caused by noise and hidden node interference, with loss
rate ph is estimated at a high value, albeit somewhat lower rates of approximately 65% and 75% respectively, providing
than in the previous example (60% against 80%). This is a demanding test of our estimator.
 10

Estimator with 3 contending stations (two exposed) Estimator with 4 contending stations (two exposed)
30 30
pexp pexp
pc pc
25 pc (1 tx,pexp=0) 25 pc (2 tx,pexp=0)

Probability (%)

Probability (%)
20 20

15 15

10 10

5 5

0 0
20 40 60 80 100 120 140 160 180 200 20 40 60 80 100 120 140 160 180 200
Time (sec) Time (sec)
(a) 3 Stations (two exposed) (b) 4 Stations (two exposed)

Estimator with 5 contending stations (two exposed) Estimator with 6 contending stations (two exposed)
30 30
pexp pexp
pc pc
25 pc (3 tx,pexp=0) 25 pc (4 tx,pexp=0)
Probability (%)

Probability (%)
20 20

15 15

10 10

5 5

0 0
20 40 60 80 100 120 140 160 180 200 20 40 60 80 100 120 140 160 180 200
Time (sec) Time (sec)
(c) 5 Stations (two exposed) (d) 6 Stations (two exposed)

Fig. 15. Collision and exposed node probability vs. number of stations associated with AP 1. Network topology as in Fig. 6.

VIII. E STIMATING E XPOSED N ODE AND C APTURE

E FFECTS
ST2 A. Exposed nodes

Low SNR
An exposed node is a sender station that senses the channel
ST1 AP1 H to be busy when, in fact, the channel at the receiver is idle
and thus a successful transmission could have been made. A
typical scenario is illustrated in Fig. 6. Here, ST 3 and ST 4
send data to AP 2 while ST 1 sends data to AP 1. Sender
ST 1 overhears the data transmissions by ST 3 and ST 4 and
senses the channel to be busy. This is incorrect, however,
Fig. 13. Topology for hidden node and noisy interference with contending
stations. since the physical separation between ST 3 and ST 4 and AP 1
means that transmissions by ST 1 would in fact be received
corrected at AP 1 even when ST 3 and ST 4 are transmitting.
ST 1 therefore defers its backoff countdown unnecessarily and
Three transmit nodes (one hidden) and low snr link
100 its throughput suffers.
ph
pc We implemented the topology in Fig. 6 in our testbed. ST 3
pn
80 pc (ph=0,pn=0) and ST 4 send 300 fps traffic to Access Point AP 2, while ST 1
uses the same channel to send 20fps traffic to AP 1 and station
Probability (%)

60
ST 2 300fps to AP 1. The channel is clean with no noise losses.
In addition to measuring pc , pn and ph as before, we now also
40
measure the total number of MAC slots R and the number I
20
of slots which are detected idle. The value of (R − I)/R is a
measure of the proportion of slots which the MAC detects to
0 be busy via carrier sense. The collision probability pc provides
20 40 60 80 100 120 140 160 180
Time (sec)
a measure of the proportion of slots that are actually busy (in
Fig. 14. Link quality estimation with collisions, noise losses and hidden
the sense that a transmission in that MAC slot would result in
nodes. a collision). The difference between (R − I)/R and pc then
provides a measure of how exposed a node is.
 11

Our measurements for this situation are shown in Fig. 15(a). Exposed node probability as function of CCA

We show the collision probability pc estimated using our 14

pexp
pc
technique and a fixed value measured without an exposed 12

Estimator Value (%)

node (labeled “pc (1tx, pexp = 0)”). It can be seen that these
10
probabilities are low and close together. In this situation,
measurements indicate that ST 1 senses the channel to be busy 8

around 10% too often i.e. pexp = 10%. This suggests that ST 1 6

may freeze its backoff counter unnecessarily for about 1 in 10 4

MAC slots 2
Fig. 15(b)–(d) show the corresponding measurements as the
0
number of stations associated with AP 1 is increased. It can be -90 -85 -80 -75 -70 -65 -60 -55

seen that, as expected, pc increases in line with measurements CCA (dBm)

in Fig. 9 without exposed nodes. The exposed node probability Fig. 16. Exposed node probability pexp vs. carrier sense threshold.
14000
pexp is consistently measured as lying between 5% and 10%,
although the relative impact of pexp decreases as the number 12000

of stations increases.
10000
To further explore our ability to sense exposed node effects,
we recall that exposed node effects are intimately related to

MAC delay (us)

8000

the choice of carrier sense threshold used. In this scenario the 6000
carrier sense mechanism is too sensitive and ST 1 senses the
channel busy too often. This effect is illustrated in Fig. 16 4000

which plots the estimated pexp vs. choice of carrier sense 2000
threshold for ST 1 in the setup of Fig. 6. As expected, it can
MAC delay
be seen that the exposed node probability pexp has the highest 0
-55 -60 -65 -70 -75 -80 -85 -90

value for carrier sense thresholds in the range −90dBm to CS (dBm)

−80dBm. At around −75dBm, the value of pexp decreases Fig. 17. MAC delay vs. carrier sense threshold.
as the impact of ST 3 disappears (confirmed by inspection of
packet traces). Finally, moving the carrier sense threshold up
to −55dBm, the effect of ST 4 also disappears and ST 1 is no
Fig. 18(a) illustrates the impact of physical layer capture.
longer exposed (again, confirmed by detailed packet traces).
It can be seen that ST 1 benefits from a lower than expected
Also shown in Fig. 16 is the measured collision probability pc .
probability of collision. In particular, while with a total of
It can be seen that this slightly increases as the carrier sense
five contending stations we expect a pc around 19% (based
threshold is increased, which is to be expected as the backoff
on measurements without capture)the measured collision rate
countdown of ST 1 is becoming of shorter duration. The
at ST 1 is only around 8%. The difference of 11% is a
benefits of using a suitable choice of carrier sense threshold are
direct measure of the capture effect advantage experienced by
illustrated in Fig. 17, which plots the estimated MAC delay5
ST 1. To help validate the accuracy of this measurement, we
at ST 1. It can be seen that the MAC delay is halved when
took the same measurements with the carrier sense threshold
the carrier sense threshold is increased to −55dBm instead of
increased to −60dBm — this change will not affect capture
−85dBm.
but would eventually highlight the presence of exposed node
A full carrier sense tuning algorithm would naturally be
effects in our setup (see previous section). As can be seen
more complex and is beyond the scope of the present paper.
from Fig. 18(b), we find that the estimates of pc and pplc are
However, this example does demonstrate the value and feasi-
almost unchanged, confirming the absence of exposed node
bility of being able to make this type of measurement.
effects in these tests.
We now further explore our ability to measure the impact
B. Physical Layer Capture of the capture effect. Note that decreasing the transmission
power at ST 1 should reduce the capture effect. We confirm
Physical layer capture occurs when colliding transmissions
this experimentally in Fig. 19 which presents measurements
have different received signal power. It may then happen that
of pc and pplc versus the transmit power at ST 1. As expected,
the transmission with highest power is successfully decoded
we can see that the capture probability pplc is greatest at the
even though it collides with another transmission. To assess the
highest transmit power of 20dBm and that pplc decreases to
ability of our estimator to measure this effect, we configured
zero as the transmit power is reduced to 0dBm. Observe that,
our testbed as shown in Fig. 7. Station ST 1 sends data packets
as might be expected, pc + pplc remains roughly constant as
to AP 1 at 20 fps. In addition we have four other contending
the transmit power is varied, with a value around the expected
stations transmitting data to AP 1 at 300 fps, but with lower
probability of collision for five saturated stations.
received signal power that ST 1.
Note that by reducing the transmit power a ST 1 we gain
5 The mean time between a packet arriving at the head of the interface queue a double benefit: not only is electrical power consumption
and being successfully transmitted. is reduced plus radio interference with adjacent WLANs,
 12

Capture effect with CCA=-80 dBm Capture effect as function of transmission power
30 30
pc + pplc (ML estimator) pc + pplc
pc (Frame losses) pplc
25 pplc 25 pc

Estimator Value (%)

Probability (%)

20 20

15 15

10 10

5 5

0 0
20 40 60 80 100 120 140 20 15 10 5 0
Time (sec) Transmission power (dBm)
(a) CCA=−80dBm
Fig. 19. Measurements of capture effect vs. transmit power.
Capture effect with CCA=-60dBm
30
pc + pplc (ML estimator)
pc (Frame losses)
25 pplc IX. C ONCLUSION
In this paper we consider how to estimate the link quality
Probability (%)

20
experienced by communicating stations in an 802.11 WLAN.
15 We make the key observation that link impairments (and
10
so quality) are intimately linked to MAC operation and so
cannot be estimated purely on the basis of PHY measurements
5 or high level measurements. We propose a powerful new
0
MAC/PHY cross-layer approach to estimating link quality
20 40 60 80 100 120 140 in 802.11 WLANs. Unlike previous approaches, we explic-
Time (sec)
(b) CCA=−60dBm itly classify channel impairments into noise-related losses,
collision induced losses, hidden-node losses consider related
Fig. 18. Demonstrating capture effect estimation. Results are shown for two
different values of carrier sense threshold, to confirm the absence of exposed
issues of exposed nodes and capture effects. Our approach
node effects in these tests. Network setup is as in Fig. 7. distinguishes between these different types of impairments and
provides separate quantitative measures of the severity of each
type of impairment. We thus make available new measures
that we expect to be of direct use for rate adaptation, channel
but the capture effect is removed and thus fairness restored
allocation, etc. and demonstrate how the measurements might
between contending stations. The effect on fairness of tuning
be applied in carrier sense tuning and power control. Since we
the transmit power can be analyzed in more detail by looking
take advantage of the native characteristics of the 802.11 pro-
at the probability of collision for each node in the network.
tocol (such as timing constraints, channel busy detection and
We carried out tests with ST 1 transmitting at 20 fps plus four
so on) — without requiring any modification to the standard
other stations with saturated traffic. Table I summarizes the
— our approach is suited to implementation on commodity
experimental measurements obtained. We can see that decreas-
hardware and we demonstrate both a prototype implementation
ing the transmit power at ST 1 increases its the probability
and experimental measurements. Indeed we argue that it is
of collision. Meanwhile, the other nodes maintain a roughly
vital to demonstrate operation in a real radio environment
constant collision probability pc , thus improving fairness in
not only because of the difficulty of developing realistic RF
the network. Note that pc is not identical at all stations due to
propagation models but also because important impairments
remaining capture effects at stations other than ST 1 (power
such as hidden-nodes and capture effects are affected by low-
asymmetries arise due to antenna tolerances, differences in
level issues (e.g. interactions between amplifier and antenna
physical location, etc.). Adjustment of the transmit power at
design as well as radio propagation) that are difficult to
all stations, could restore fairness.
model in simulations. We note that many of the measurements
presented are new and of interest in their own right.
node 1 node 2 node 3 node 4 node 5
TX power (dBm) pc + pplc (%) pc (%) pc (%) pc (%) pc (%) pc (%)

16 18.8 2.3 14.9 11.0 17.3 15.9 A PPENDIX : REMARKS ON HIDDEN NODES
13 18.4 5.5 13.6 12.4 18.1 16.3
10 18.0 9.9 14.5 10.9 17.6 16.1
A. Performance of RTS/CTS with hidden nodes
7 17.6 11.9 14.3 12.3 17.3 16.0 In this paper we make use of the packet fragmentation
4 17.5 15.6 12.1 12.7 17.7 16.1
functionality in 802.11 to mitigate hidden node effects. Of
1 17.5 17.1 14.1 10.6 17.8 16.3
course it is more common to consider use of RTS/CTS
TABLE I
handshaking for this purpose and in principle the behavior
FAIRNESS WITH POWER TUNING .
should be similar. However, in practice we found a number
of basic difficulties with the use of RTS/CTS handshaking for
this purpose.
 13

Firstly, consider an experiment with 7 stations transmitting CRC errors for a Hidden node
100
traffic at 300 frame per second (fps) without noise and hidden CRCerr0/(T0-A0)

node interference. In Fig. 20(a) we plot the probability of 98

collision with and without RTS/CTS (labeled as rts − pc and 96

Percentage (%)
no rts−ptot respectively). The RTS/CTS collision probability
94
is estimated from the number of missed CTS frames. To
confirm the absence of noise interference, we have also plotted 92

the overall probability of error (labelled rts−ptot ), which also 90

takes into account the number of missed ACK over sent Data 88
frame. Thus in this basic case, it can be seen that RTS/CTS
86
reliably estimates the probability of collision. 0 10 20 30 40 50 60 70 80 90 100

Now consider a scenario with a hidden node. As a baseline Time (sec)

we collect data when two transmitting stations are within Fig. 21. Hidden nodes, clean channel, no collisions. tx1,err is loss rate
for first fragment transmissions, tx2,err loss rate for second fragments (an
one another’s carrier sense region. As expected we see a low estimate of pn ), rx1,err the error rate measured at the receiver for first
collision probability of around 7%, see Fig. 20(b) (line labelled fragments, rx2,err the rate for second fragments.
rst − no hi). Now, we move the transmitters so that they
are hidden from one another. In the absence of RTS/CTS,
we measure a high error probability of around 82% (labelled are attributed to both nodes choosing to transmit in the same
norts) which is mainly caused by hidden node errors. If we slot thus leading to PHY errors, as we expect.
enable RTS/CTS, the error probability drops, but not to the
expected value of 7%. Instead, we have a residual error of R EFERENCES
about 52% (line labelled rts in Fig. 20(b)). That is, in presence
[1] IEEE 802.11 WG, IEEE Std 802.11, 1999 edition “IEEE standard for
of hidden nodes the RTS/CTS estimator is still subject to Information technology - Telecommunications and information exchange
considerable hidden node interference. between systems - Local and metropolitan area networks. Specific
In order to understand this behaviour, we note that the requirements. Part 11: Wireless LAN Medium Access Control (MAC)
and Physical Layer (PHY) specifications”, 1999.
hidden node will defer its transmission if it overhears the CTS [2] IEEE Std 802.11e, “Amendment to STANDARD [for] Information
from the receiver before sending its frame. We can calculate Technology - Telecommunications and Information Exchange Between
when this occurs. Our tests used an 802.11g PHY. Station 1 Systems - LAN/MAN Specific Requirements - Part 11: Wireless
Medium Access Control (MAC) and Physical Layer (PHY) specifi-
sends an RTS frame (duration 48µs), the receiver waits for cations: Medium Access Control (MAC) Quality of Service (QoS)
a SIFS (duration 16µs) and finally a CTS frame is sent by Enhancements”, 2005.
the receiver. Thus the hidden station would need to leave the [3] DJ Leith, P Clifford, “A Self-Managed Distributed Channel Selection
Algorithm for WLANs”, Proc. IEEE RAWNET, Boston, 2006.
medium idle for at least 64µs in order to receive the CTS [4] A Mishra, et al. “A Client-driven Approach for Channel Management
frame. This is much longer than the PHY slot duration of in Wireless LANs”, Proc. IEEE INFOCOM, Barcelona, 2006.
20µs for mixed mode 11b/g. Indeed if the backoff counter [5] B Kauffmann et al. “Measurement-Based Self Organization of Inter-
fering 802.11 Wireless Access Networks”, Proc. IEEE INFOCOM,
of the hidden node is less than 3 when the other station Anchorage, 2007.
begins its RTS transmission, then the hidden node will make [6] G Bianchi, “Performance analysis of IEEE 802.11 distributed coordina-
a transmission that corrupts the CTS frame. tion function”, IEEE J. Sel Area Comm, 18(3):535–547, Mar. 2000.
[7] K Ramachandran et al., “Scalability analysis of Rate Adaptation Tech-
In order to verify this dependency on the PHY slot duration, niques in Congested IEEE 802.11 Networks: An ORBIT Testbed Com-
in Fig. 20(c) we show measurements when the PHY slot is parative Study”, Proc. IEEE WoWMoM, 2007.
increased to 40µs. As expected, the probability of error in the [8] S Wong, et al., “Robust Rate Adaptation for 802.11 Wireless Networks”,
Proc. ACM MobiCom, 2006.
presence of RTS/CTS is reduced. While a longer slot time can [9] C. Reis, R. Mahajan, M. Rodrig, D. Wetherall, J. Zahorjan
be used in our testbed to mitigate this issue, in practice our “Measurement-Based Models of Delivert and Interference”, Sigcomm
results indicate that RTS/CTS is unsuitable for estimating the 2006.
[10] R. Gummadi, D. Wetherall, B. Greenstein, S. Seshan, “Understanding
collision probability in the presence of hidden nodes. and Mitigating the Impact of RF Interference on 802.11 Networks”,
Sigcomm 2007.
[11] I. Broustis, J. Eriksson, S. Krishnamurthy, M. Faloutsos “Implications
B. CRC errors with hidden nodes of Power Control in Wireless Networks: A Quantitative Study”, Proc.
PAM, 2007.
In section VI-C we introduce a model to cross-validate [12] Bruno, R.; Conti, M.; Gregori, “Mesh networks: commodity multihop
measurements by counting CRC errors. When considering ad hoc networks”, Proc. IEEE Communications Magazine, March 2005.
hidden node errors, we note that for mixed mode 11b/g PHY [13] D Qiao and S Choi, “Goodput Enhancement of IEEE 802.11a Wireless
LAN via Link Adaptation”, Proc. IEEE ICC, 2001.
errors are only generated when a hidden node and an ordinary [14] I. Haratcherev, K. Langendoen, R. Lagendijk and H. Sips, “Hybrid Rate
node select the same slot to begin a transmission. Hidden-node Control for IEEE 802.11”, Proc. ACM, MobiWac, 2004
transmissions beginning in later slots (i.e., after an ordinary [15] A Kochut, et al., “Sniffing out the correct physical layer capture model
in 802.11b”, Proc. IEEE ICNP, 2004.
node has already started, see Fig. 8.) result in CRC errors. To [16] D Malone, et al. “MAC Layer Channel Quality Measurement in 802.11”,
confirm this for our setup, we took 2 hidden nodes transmitting IEEE Comms Let., 11(2):143–145, Feb. 2007.
at 300 fps. Fig. 21 shows the fraction of retry errors at the [17] D Giustiniano, et al. “Experimental Assessment of 802.11 MAC Layer
Channel Estimators”, IEEE Comms Let., 11(12):961–963, Dec. 2007.
transmitter that are mapped into CRCerr frames at the receiver. [18] P Chatzimisios, et al., “Effectiveness of RTS/CTS handshake in IEEE
We see a consistent level of about 91%. The remaining 9% 802.11a wireless LANs”, Electronic Letters, 40(14):915–916, Jul. 2004.
 14

collisions - node 2 hidden node - slot=20 microsec hidden node - slot=40 microsec
100 100 100
no rts-ptot no rts no rts
rts-pc rts rts
rts-ptot rts-no hi rts-no hi
80 80 80

60 60 60

ptot (%)

ptot (%)
%

40 40 40

20 20 20

0 0 0
10 20 30 40 50 60 10 20 30 40 50 60 10 20 30 40 50 60
Time (sec) Time (sec) Time (sec)
(a) Without hidden nodes. (b) With a hidden node and time slot equal to 20µs. (c) With a hidden node and time slot equal to 40µs.

Fig. 20. Estimating pc with RTS/CTS under various conditions.

[19] ST Sheu, et al., “The impact of RTS threshold on IEEE 802.11 MAC Doug Leith Doug Leith graduated from the Univer-
protocol”, Proc. Int’l Conference on Parallel and Distributed Systems, sity of Glasgow in 1986 and was awarded his Ph.D.,
Dec. 2002. also from the University of Glasgow, in 1989. In
[20] J Kim, et al. “CARA: Collision-Aware Rate Adaptation for IEEE 802.11 2001, Prof. Leith moved to the National University
WLANs”, Proc. IEEE INFOCOM, 2006. PLACE of Ireland, Maynooth to assume the position of SFI
[21] G Bianchi, I Tinnirello, “Kalman Filter Estimation of the Number PHOTO Principal Investigator and to establish the Hamilton
of Competing Terminals in an IEEE 802.11 network”, Proc. IEEE HERE Institute (www.hamilton.ie) of which he is Director.
INFOCOM, 2003. His current research interests include the analysis
[22] KJ Yu, et al., “A novel hidden station detection mechanism in IEEE and design of network congestion control and dis-
802.11 WLAN”, IEEE Comms Let., 10(8):608–610, Aug. 2006. tributed resource allocation in wireless networks.
[23] M Heusse, et al. “Idle sense: an optimal access method for high
throughput and fairness in rate diverse wireless LANs”, Proc. ACM
SIGCOMM, 2005.
[24] D Aguayo, et al. “Link-level measurements from an 802.11b mesh
network”, Proc. ACM SIGCOMM, 2004.
[25] D Giustiniano, et al., “An explanation for unexpected 802.11 Outdoor
Link-level Measurement Results”, Proc. IEEE INFOCOM, to appear
2008.
[26] D Malone, et al., “Modeling the 802.11 distributed coordination function
in non-saturated heterogeneous condition”, IEEE ACM T. Network,
15(1):159–172, 2007.

Domenico Giustiniano Domenico Giustiniano grad-

uated from the University of Palermo in 2003, where
he worked as a researcher until November 2004.
Then he moved to the University of Roma Tor Konstantina Papagiannaki Konstantina Papagian-
PLACE Vergata where he is currently a Ph.D. candidate in naki received her first degree in electrical and com-
PHOTO Telecommunication Engineering under supervision puter engineering from the National Technical Uni-
HERE of Prof. G. Bianchi. He visited the TAIT department, versity of Athens, Greece, in 1998, and her Ph.D.
University of Ulm, Germany during his master thesis PLACE degree from the University College London, U.K.,
and he spent 2007 as a visiting Ph.D. student at PHOTO in 2003. From 2000 to 2004, she was a member
Hamilton Institute, Ireland. His interests include HERE of the IP research group at the Sprint Advanced
signal processing and 802.11 channel quality. Technology Laboratories. She is currently with Intel
Research in Pittsburgh and holds an adjunct faculty
position in the Computer Science Department at
Carnegie Mellon University.

David Malone David Malone received B.A.(mod),

M.Sc. and Ph.D. degrees in mathematics from Trin-
ity College Dublin. During his time as a postgrad-
uate, he became a member of the FreeBSD devel-
PLACE opment team. He is a research fellow at Hamilton
PHOTO Institute, NUI Maynooth. His interests include math-
HERE ematics of networks, network measurement, IPv6
and systems administration. He is a co-author of
O’Reilly’s “IPv6 Network Administration”.