Contention and traffic load-aware association in IEEE 802.11 WLANs: Algorithms and implementation

Abstract

Efficient association of a station with the appropriate access point has always been a challenging problem. The standard approach of considering only the Received Signal Strength, has recently been substituted by more efficient schemes that consider channel conditions, cell population etc. However, in spite of the large variety of approaches, several factors that determine to a large extent user throughput after association with an access point have been overlooked. In this work, we propose innovative metrics on which association should be based. First, we capture the contention from one-hop and interference from two-hop neighbors that is inherent in IEEE 802.11 WLAN environments. Second we include the PHY transmission rate and show preference to higher rates that reduce the above effects. Third, unlike most relevant approaches, we define an activity factor that reveals the anticipated activity due to backlogged traffic. We devise an association protocol suite, through which messages containing the information above are passed between the AP and the user to support association decisions for the uplink and downlink. We implement the proposed mechanism using the MAD-WiFi open source driver and moreover show through experiments in a wireless testbed that it significantly improves user performance in real conditions.

Full text

Page 1

Contention and Traffic Load-aware Association in IEEE 802.11 WLANs:
Algorithms and Implementation
Stratos KeranidisThanasis KorakisIordanis Koutsopoulos Leandros Tassiulas
Department of Computer and Communication Engineering, University of Thessaly, Greece
Centre for Research and Technology Hellas, CERTH, Greece
E-mail: {efkerani, korakis, jordan, leandros}@uth.gr
Abstract—Efficient association of a station with the appropriate
access point has always been a challenging problem. The standard
approach of considering only the Received Signal Strength,
has recently been substituted by more efficient schemes that
consider channel conditions, cell population etc. However, in
spite of the large variety of approaches, several factors that
determine to a large extent user throughput after association
with an access point have been overlooked. In this work, we
propose innovative metrics on which association should be based.
First, we capture the contention from one-hop and interference
from two-hop neighbors that is inherent in IEEE 802.11 WLAN
environments. Second we include the PHY transmission rate and
show preference to higher rates that reduce the above effects.
Third, unlike most relevant approaches, we define an activity
factor that reveals the anticipated activity due to backlogged
traffic. We devise an association protocol suite, through which
messages containing the information above are passed between
the AP and the user to support association decisions for the
uplink and downlink. We implement the proposed mechanism
using the MAD-WiFi open source driver and moreover show
through experiments in a wireless testbed that it significantly
improves user performance in real conditions.
Index Terms—Wireless communications, Association, Handoff,
MAC, IEEE 802.11
I. INTRODUCTION
In IEEE 802.11 WLANs, each station (STA) has to first
associate with an access point (AP), before it can start trans-
mitting data to other nodes in the network. This association
procedure consists of four phases. During the first phase, a
STA has to discover the networks in its vicinity before it can
join a Basic Service Set (BSS). This process is called scanning
and can be either passive or active. In passive scanning, a
STA scans all available channels and listens to information
periodically broadcasted by the APs in their beacon frames.
In active scanning, a STA tries to find the BSSs in its vicinity
by transmitting a Probe Request frame on each channel of the
channel list. APs respond by sending Probe Response frames.
Having collected these frames, the STA decides which AP
it will associate with, in the second phase. According to the
standard [1], AP selection is based on the Received Signal
Strength Indication (RSSI). A STA simply selects the AP from
which it has received the strongest signal during the scanning
process. In the third phase, the STA has to follow the authenti-
cation process if the selected AP follows some authentication
mode. Finally, the STA sends an Association Request frame
to the selected AP and sequentially the AP responds with
an Association Response frame. If the Association Response
frame is received with a ”successful” status value, the STA is
now associated with the AP.
The rest of the paper is organized as follows. In the
remaining of this section the state of the art related work is
presented and a summarization of our contribution follows. A
detailed analysis of our metric definition follows in section
II. Details about the proposed algorithms and their implemen-
tation are provided in section III. The configuration of our
experiments, concerning the testbed and the methodology used
is then described in section IV. In section V, we experimentally
evaluate the performance of our implementation. Finally, in
section VI, we present the conclusion and discussion of future
work.
A. Related Work
The performance of the standard AP selection policy has
been extensively studied and it is well known that it leads to
inefficient use of the network resources [2],[3]. In addition,
due to the asymmetric nature of the wireless medium, this
policy becomes unsuitable, as RSSI is an indicator just for the
downlink channel and not for the uplink. An association mech-
anism considering signal to interference and noise (SINR)
per connection, as well as asymmetric traffic was proposed
in our previous work [4]. Although our approach considered
uplink channel conditions as well, thus offering a significant
improvement, it was not able to lead to the best available
throughput performance.
One of the major issues studied among relevant works
has been the proper definition of AP load. The authors in
[5], proposed an AP selection policy that estimates AP load
based on instantaneous measurements of the transmission rate
and the fraction of time an AP acquires the channel for its
transmissions. However, this model faces the disadvantage
of considering only downlink traffic and therefore assumes
that channel contention is only among APs. Another common
assumption of works on the field has been to denote AP load as
a factor reflecting the AP’s inability to satisfy the requirements
of its associated users [2],[6]. Another approach followed in
[7], bases association decisions on a metric denoted as airtime
cost, which considers both uplink and downlink traffic as well
as AP load. The above approaches, have the common char-
acteristic of considering the effect induced by transmissions
only of associated users in the AP load estimation.
However, since the IEEE 802.11 MAC layer is based on
contention, the efficiency of an AP is not only dependent on

Page 2

the STAs associated with it, but also on other neighboring
stations and their activity.
Trying to address this issue the authors in [8] consider AP
load over all neighboring nodes. The new scheme incorporated
both the effects of associated and contending nodes in its
throughput estimation algorithm. However, this approach was
restricted in considering only downlink transmissions and
setting fixed transmission rates, neglecting the importance of
rate adaptation mechanisms.
All the above approaches follow the assumption of fully
saturated traffic, which considers that all users transmit and
require service at all times. In [9], the authors suggest that
APs should assign an activity level estimator to their associated
STAs based on observations of their traffic intensity. Neverthe-
less, this approach does not manage to characterize the traffic
intensity of neighboring nodes that belong to adjacent cells,
although these contend for channel usage or even interfere
with transmissions in the cell under consideration. Moreover,
they suggest that an Inter-AP protocol is required, that is used
to collect activity estimations about all STAs in the WLAN
and feed this information to a central entity that calculates
the optimal association scheme, considering aggregate WLAN
throughput. However, such centralized approaches can only
apply to centrally managed deployments, which is not the
general case.
One more issue that has not received much attention in the
association process, is the effect of hidden node terminals,
which appears very often in dense WLANs. In a later work
in [10], a metric is proposed, that comprises contention
and interference as well. The authors trying to estimate the
effect of interfering nodes, use a factor that captures the
error probability due to collisions, considering it as a value
proportional only to the number of STAs associated to each
AP and STAs that belong to neighboring cells and operate
on the same channel. The disadvantage of this assumption
is that it does not consider APs transmitting on downlink as
potential interference. In addition, their approach is not able to
distinguish between nodes that just contend for channel usage
and nodes that appear hidden.
B. Our Contribution
In this paper, we propose a novel approach that resolves the
issues mentioned above not as individual parameters but in a
joint manner. We contribute by developing a comprehensive
metric, that is based on estimation of end user throughput
in 802.11 infrastructure networks. In order to capture the
asymmetric nature of the wireless medium, we estimate per-
formance both on uplink and downlink channel.
As a first contribution we encapsulate in our throughput
estimation formula the effect of contention. In contradiction
to the aforementioned approaches, we state that AP load
should be considered over all neighboring nodes, due to the
shared nature of the medium. The IEEE 802.11 medium
access is performed by the distributed coordination function
(DCF), that is based on the CSMA/CA protocol. This medium
access control (MAC) protocol provides all compliant nodes
with the same chance to access the medium and transmit
frames in the long term. As a result every node in a WLAN
shares the medium with its neighboring nodes. Moreover,
due to the multi-rate capability at the physical layer (PHY),
supported by rate adaptation mechanisms, the transmission
duration of a frame depends on the transmission rate selected
by the transmitter. As a result, transmitters that use low PHY
rates, capture the medium for longer duration, muting their
surrounding nodes during their transmissions. The combined
effect of shared medium in accordance with the multiple
PHY rates used, can cause the well known 802.11 ”anomaly
phenomenon”, where low transmission rate STAs negatively
affect high bit rate ones [11]. The result is that all STAs finally
get throughput of the same order of magnitude. Consequently,
we have to take a step further than the previous approaches
and take into account transmissions of all active nodes in
STA?s neighborhood, in accordance with their transmission
rates, in order to estimate the levels of contention and extend
the definition of load.
Another key contribution of the proposed scheme is its
ability to adapt to the varying traffic patterns that each corre-
sponding node follows. Thus it manages to adapt to realistic
traffic conditions. We state that activity estimation should be
performed by each individual node and this information can
be exchanged through neighboring WLANs, by using specially
generated for this purpose management frames. Through this
approach, all nodes are able to detect transmissions in adjacent
cells in a distributed way.
As a third contribution we investigate how simultaneous
transmissions of hidden nodes affect user performance. Due to
the shared medium, transmissions of interfering hidden nodes
can cause collisions and erroneous receptions, that lead to de-
creased packet delivery ratios (PDR) at the receivers. Counter
to relevant approaches, we incorporate in our proposed metric
the effects of contending and interfering nodes separately.
Our mechanism, integrating all the above features, results in
algorithms proposed both for the association and the handoff
procedure. One more important contribution is that we move
one step further than simulation and implement the proposed
algorithms using open source drivers and also validate their
performance in a wireless testbed, to evaluate the performance
in real world settings.
II. SYSTEM MODEL AND METRIC DEFINITION
We consider an IEEE 802.11 based WLAN that consists of
a large number of APs and STAs. We use M to denote the
set of APs that define a network coverage area. We assume
that there is a set of available channels, denoted by C. Each
APj ∈ M operates on a single predefined channel cj ∈ C,
where C denotes the set of non-overlapping channels that the
operating band offers. The coverage areas of multiple APs
may be overlapping. Within the network coverage area resides
a set of mobile STAs, denoted by N, which tend to stay in
the same physical locations for long time periods. At any time
instant, a STAi∈ N chooses to associate with a single APj
∈ Mi, where Midenotes the set of APs that operate in the
vicinity of STAi. We use Njto denote the set of STAs that
are associated with APj.

Page 3

Each node of the network n ∈ M ∪ N, has a set of
neighbors, that reside in its sensing area and operate on the
same channel with n . This set of ”1-hop” neighbors, that can
be either APs or STAs, is denoted by An.
Based on the discussion of the previous section, we notice
that the throughput, which is experienced by a node in an IEEE
802.11 network, depends apart from channel quality also on
the transmission of frames by other nodes in the network and
its selected PHY rate. The authors of [12], based on the well
known analysis of Bianchi [13], have shown that when there
are multiple transmitters with different PHY rates, that lie in
the contention domain of node n, which uses a PHY rate of
Rn, each node of the network enjoys an equally shared value
of throughput, that is approximated as:
Tn=
1
1
Rn
+
|An|
?
k=1
1
Rk
(1)
where Rk denotes the PHY rate that each node k ∈ An
uses while transmitting. This equation ignores the overhead
resulting from the 802.11 MAC mechanism. The deficiency
mentioned here, is not important in our analysis, as we
use this equation to decide about the association that can
lead to the best available throughput performance and not in
order to calculate the actual resulting end user throughput.
Moreover, this equation considers saturated traffic conditions
and requires that all traffic flows consist of equal packet
lengths. Throughout this paper, we follow the assumption of
equal packet lengths, but later in our analysis we transform
the above equation, in order to capture realistic varying traffic
conditions. The above expression is based on the estimation
of channel usage time that each transmitter node gains access
to the medium, given the existence of other 802.11 nodes
operating on its channel and transmitting in its vicinity. We
modify the above equation, which refers to the general case
of a network with multiple flows generated between 802.11
compliant nodes, to fit our needs about the special case of
infrastructure 802.11 networks.
A. Contention Effect
We start forming our formula by considering only saturated
downlink traffic. Assume the usual case, where an APj ∈
M has |Nj| associated STAs and serves them with downlink
traffic. We also consider that in the vicinity of APj , there
are |Aj| ”1-hop” neighboring nodes operating on the same
channel and contending to capture the channel. An AP has an
equal probability among its contending nodes to capture the
medium for its own transmissions and in each such instant
it uses the medium to transmit to only one of its associated
STAs. The AP’s service rate is equally shared between the
associated STAs, if the number of frames destined to each
STA is equally distributed among them. Supposing this equal
distribution and the upcoming association of STAiwith APj,
the mean PHY rate that APjuses when transmitting downlink
traffic to its associated STAs can be approximated as follows:
Rj= (
|Nj|
?
m=1
Rjm+ Rji)/(|Nj| + 1)
(2)
where Rjmand Rjidenote the PHY rate that APjuses when
transmitting to each STAm ∈ Nj and STAi accordingly.
Concluding, in order to estimate the equally shared value of
transmitted bits destined to each STAi while associated to
APj, we transform expression (1) as:
Tdown
ij
=
1
(1
Rj
+
|Aj|
?
k=1
1
Rk) · (|Nj| + 1)
(3)
Thus, we capture the effect of equal sharing of the AP?
service rate among its |Nj| associated users, while we assume
that APj uses a mean value of PHY rate when transmitting
to all its associated STAs.
js
B. Hidden-Node Effect
As previously stated, these transmissions are still subject
to frame errors, due to interference at the receivers’ side.
The factor that plays the key role in interference, is the
effect of hidden terminals, which appears very often in dense
WLANs. In a simple downlink scenario with one APj and
one associated STAi, the set of nodes that appear hidden to
transmissions of APj, consists of nodes existing in the ”1-
hop” neighborhood of STAi, that do not belong in the ”1-
hop” neighborhood of APj. We call this set of nodes as the
”2-hop” neighbors of the transmitter, denoted as Bj, equal to
the relative complement of set Ajin Ai(Ai\Aj). Since the ”2-
hop” neighbors of the transmitter APjdo not sense its ongoing
transmissions, collisions occur, leading to decreased packet
delivery ratio (PDR) and consequently result in throughput
decrease. The negative effect of the hidden-node problem
is proportional to the number of ”2-hop” neighbors (|Bj|)
and their transmissions duration. This effect can be imported
in expression (3) to model the decrease in performance as
follows:
Tdown
ij
=
1
(1
Rj
+
|Aj|
?
k=1
1
Rk) · (|Nj| + 1) +
|Bj|
?
l=1
1
Rl
(4)
C. Traffic Intensity Estimation
In most cases, the consideration of saturated traffic is
not realistic. Practically, nodes run different applications that
generate traffic with varying rates. Assuming that all nodes
in a network generate traffic with the same rate can only
estimate performance under the worst case scenario. In order
to model realistic scenarios we have to characterize each
transmitter according to the traffic pattern that it follows. For
this purpose, we define an activity indicator, denoted as fn,
for each node n ∈ M ∪ N. Each node n measures the rate
of packets arriving to its transmission queue during a constant
time interval, capturing its arrival rate (λn). Moreover it can
estimate its affordable service rate (µn) using expression (1),

Page 4

which is approximately the rate at which packets leave its
queue for transmission. In the case that the rate of packets
arriving to the transmission queue is higher than the rate at
which packets leave the queue, only the number of backlogged
packets increases, while the traffic injected in the network
remains constant.
Based on the above, n estimates its maximum affordable
traffic rate, by setting its activity indicator fnas follows:
fn= min{λn,µn}
Each node of the network announces its fn to all its
neighbors. This way, every node that receives reports about
ongoing transmissions in its neighborhood, manages to create
a list of all its’ ”1-hop” neighbors and their corresponding
activity indicators. Moreover, it has to detect its ”2-hop”
neighbors. For this purpose, all nodes have to exchange their
lists of neighbors. The above activity estimation procedure is
performed by each node of the network, either operating as
an AP or as a STA. We now use the activity indicators to
transform equation (4) as follows:
(5)
Tdown
ij
=
1
(fj
Rj
+
|Aj|
?
k=1
fk
Rk) · (|Nj| + 1) +
|Bj|
?
l=1
fl
Rl
(6)
where Rj is now calculated regarding the percent of traffic
destined to each individual STAm∈ Nj and STAiaccord-
ingly.
A similar approach can be used for uplink communication
as well. When a STA transmits on uplink, all of its frames are
destined to the AP it is associated with. The transmitter, STAi
in this case, shares the medium with its ”1-hop” neighbors
(Ai), while its ”2-hop” neighbors (Bi) are the nodes that are
located in the AP’s neighborhood but not in the STA’s (Aj\
Ai). For the uplink case, we arrive at the following expression:
Tup
ij=
1
fi
Rij
+
|Ai|
?
k=1
fk
Rk
+
|Bi|
?
l=1
fl
Rl
(7)
III. PROPOSED ALGORITHMS
A. Association Mechanism
The above analysis concludes in two expressions (6) and (7),
that estimate throughput performance for uplink and downlink
communications accordingly. In our model, we have assumed
that APs are statically assigned predefined channels, that
do not change during operation time. We do not consider
channel allocation in this work, since our focus is on devising
throughput-efficient access point association mechanisms. As
long as the operating channel remains constant for the APs,
they can constantly monitor their ”1-hop” and ”2-hop” neigh-
borhoods. On the other hand, the STAs are able to change
their operating channels by performing handoffs between APs
that operate on different channels. This way, the set of detected
neighbors depends on the channel the STA operates on. During
the scanning period, each STA has to remain on each channel
for duration equal to the Neighbor Reports’ interval, in order to
collect all the reports transmitted by its neighbors. We denote
this time period as tr. At the end of each tr, each STA has
to store the list of neighbors that it detects on each channel.
We use Ac
hop” neighbors, detected by STAi on each channel c. This
scanning procedure is repeated |C| times, so that each STA
can estimate its neighbors on all the available channels.
Generally in wireless communications, downlink connec-
tions dominate the overall communication load. However most
real-time applications such as VoIP or video conferencing
require suitable Quality of Service (Qos), in both the uplink
and the downlink. We indicate the uplink-to-total-link ratio as
urand similarly for downlink as dr. Each STA can determine
its own ratios, concerning the type of application it is running.
By using these factors, STAiunder association, can calcu-
late the combined metric, considering the achievable perfor-
mance when it is associated with APj, as follows:
iand Bc
i, to denote the sets of ”1-hop” and ”2-
Ttotal
ij
= ur· Tup
ij+ dr· Tdown
ij
(8)
Having calculated the above metric, considering every APj
∈ Ai, STAi estimates the potential performance both on
uplink and downlink for each available association and then
decides to associate with the AP, that provides the maximum
calculated metric. A brief pseudocode description of the im-
plemented association algorithm is given in Table Algorithm
1.
B. Handoff Mechanism
The above analysis should be extended in the handoff
mechanism. A handoff in 802.11 is the process that allows
a STA to change the AP that is associated with, because it
detects degradation of the communication quality. According
to the IEEE 802.11 standard, when a STA moves away from
the AP it is associated with, the SNR of the link drops, and
if the Cell Search Threshold is reached, the MAC Layer Scan
function starts to search for potential APs. The Cell Search
Threshold is not explicitly defined in the standard. Implemen-
tation of appropriate triggering mechanisms is typically left to
the wireless card manufacturer, and is therefore proprietary.
As previously explained, several more factors than the signal
strength affect communication quality. The key feature that
our mechanism supports is the consistent monitoring of all
these factors jointly, by calculating the proposed metrics. This
way, each STA can monitor the throughput performance that
its current association offers and consequently decide whether
a handoff to another availble AP is required. In the proposed
scheme, the triggering of the scanning procedure is based on a
throughput percent threshold denoted by H1, instead of RSSI-
based thresholds.
Moreover, in our model we introduce a periodical scan-
ning window, during which each STA triggers the scanning
procedure, so as to be able to estimate potential performance
considering APs that operate on the other available channels.
We define as tm the period of this periodical procedure. In
addition, we set a time threshold denoted by H2, to determine
the validity of the results in our scan cache. This threshold

Page 5

Algorithm 1 ASSOCIATION MECHANISM
Require: TIME := |C| ∗ tr
Require: INCOMING NEIGHBOR REPORT OF EACH k ∈ Ac
Require: INCOMING BEACON OF EACH j ∈ Mi
Require: TIME OF CHANNEL
Ensure: ASSOCIATION DECISION FOR STAi
i
1: while TIME < |C| ∗ tr do
2:
for c ∈ C do
3:
WAIT IN RECEIVE MODE FOR tr
4:
for k ∈ Ac
5:
COLLECT NEIGHBOR REPORT OF k
6:
SAVE NEIGHBOR LIST OF k, Rk,fk
7:
end for
8:
CALCULATE LIST OF Ac
9:
for j ∈ Mido
10:
if cj= c then
11:
COLLECT BEACON OF APj
12:
ESTIMATE RijUsing RSSIj
13:
SAVE Nj
14:
CALCULATE LIST OF Aj,Bj
15:
CALCULATE Tdown
(7)
16:
CALCULATE Ttotal
17:
end if
18:
end for
19:
end for
20: end while
21: for j ∈ Mido
22:
STAiSELECTS APjThat Maximizes (8)
23: end for
24: STAiASSOCIATES WITH APj
ido
i,Bc
i
ij
,Tup
ij
Using (6),
ij
Using (8)
is used to avoid the overhead induced by inefficient scanning
procedures that lead to simiral results. A brief pseudocode
description of the implemented handoff algorithm is given in
Table Algorithm 2.
C. Implementation Details
In this section we describe the key challenges encountered
in the driver implementation and the corresponding solutions.
For the implementation of our mechanism, we used the
MAD-WiFi open source driver [14]. Our proposed mechanism
assumes that each node is able to receive information about
ongoing transmissions from all the BSSs taking place on the
channel it is operating on. However, all packets received by
the network adapter are filtered out, so that the ones with a
destination address different than the local MAC address of the
adapter are discarded. Only unicast packets that are destined
to the adapter’s MAC address, multicast and broadcast packets
can be captured. A solution for our needs would be the passive
approach of capturing unicast data packets that would contain
the required information in specially generated fields in the
header of the data packets had to be avoided, due to the
large amount of information that has to be exchanged between
STAs.
Algorithm 2 HANDOFF MECHANISM
Require: TIME := k ∗ tm, k ∈ N
Require: HANDOFF THRESHOLDS H1, H2
Require: ASSOCIATED WITH AP0WITH Told
Require: OPERATION ON CHANNEL c = cold
Require: INCOMING NEIGHBOR REPORT OF EACH k ∈ Ac
Require: BEACON OF AP0
Ensure: HANDOFF DECISION
i0
i
1: while TIME < k ∗ tm do
2:
for k ∈ Ac
3:
SAVE NEIGHBOR LIST OF k, Rk,fk
4:
end for
5:
CALCULATE LIST OF Ac
6:
CALCULATE LIST OF Ac
7:
CALCULATE Tdown
i0
8:
CALCULATE Tnew
i0
9:
if (Tnew
i0
- Told
10:
if Scan Invalid (H2) then
11:
Scanning Procedure()
12:
end if
13:
Association Decision()
14:
end if
15: end while
16: if Scan Invalid (H2) then
17:
Scanning Procedure()
18: end if
19: Association Decision()
ido
i,Bc
0,Bc
i0Using (6), (7)
Using (8)
i0)% > H1 then
i
0
,Tup
Instead of using this scheme, an active information approach
has been followed. More specifically, the first modification we
made in the driver is the generation of an 802.11 broadcast
frame, that is transmitted periodically. This special control
packet, called Neighbor Report, includes the PHY rate used
in the last transmission of each node and its activity indicator,
computed using expression (5). We further modified the driver
in a way that each node n that receives these broadcast frames,
estimates its ”1-hop” neighborhood and subsequently creates
a list with the MAC addresses of each node k ∈ An , their
PHY rates and their fk accordingly. Upon the reception of
Report packets by all its ”1-hop” neighbors, node n estimates
its ”2-hop” neighborhood, as described in the previous section.
A third modification of the driver was the extension of the
Beacon frames transmitted by the APs, by adding an extra
field that contains the number of associated users and its
mean PHY transmission rate, as calculated upon observation
according to the percent of traffic that is destined to each
associated STA. In addition, the PHY rates (Rjiand Rij) that
will be used by STAifor its communication with each APj
are estimated considering the strength of Beacon and Probe-
Response Frames, transmitted by the neighboring APs. Our
final modification was made in the scanning procedure, where
we set the interval that STAs have to remain on each channel
equal to the Neighbor Reports’ interval (tr).