Content uploaded by Evangelos Gazis
Author content
All content in this area was uploaded by Evangelos Gazis on Jul 18, 2014
Content may be subject to copyright.
Content uploaded by Evangelos Gazis
Author content
All content in this area was uploaded by Evangelos Gazis on Jul 18, 2014
Content may be subject to copyright.
IEEE Communications Surveys & Tutorials • 1st Quarter 2008
70
he continuing evolution of mobile communications has
spawned several radio technologies (e.g., orthogonal
frequency-division multiplexing [OFDM], code-division
multiple access [CDMA]) and mobile network architectures
(e.g., Third Generation Partnership Project [3GPP], 3GPP2)
over the last few years. This technological proliferation has, in
turn, brought on an unprecedented increase in the number of
wireless access standards and their associated protocol stacks.
The need for ever faster standardization cycles and the urgent
demand to support access to the Internet by these mobile net-
work architectures called for a “mix-and-match” approach to
the definition of the associated protocol stacks. As individual
protocols are typically specified with different assumptions in
mind, the end-to-end performance of these protocol stacks in
deployed mobile networks has not always been satisfactory.
Stratification, the composition mechanism for protocol
frameworks, renders each protocol layer impervious to the
functionality embedded within other protocol layers. Inside a
protocol stack, exchange of control and data information may
take place only between adjacent protocol layers and is sup-
ported by the concept of a service access point (SAP). A SAP
provides access to a selected subset of protocol functionality
via a precisely defined set of primitive operations. A particu-
lar protocol layer may offer and/or use more than one SAP,
depending on its function and the information it needs to
exchange with its adjacent protocol layers. This black box
paradigm lies at the heart of the standard open systems inter-
connection (OSI) reference model and has been the prevalent
design approach since the dawn of all modern networking
architectures [1]. In this respect, any attempt to violate the
OSI reference model is considered a cross-layer design. In the
same context, in wireless networks any abstract model that
rationalizes in a nontrivial manner the cross-layer interactions
from the physical to the transport layer in order to allow
information transfer across the layers’ stratification bound-
aries is also considered a cross-layer design. In most cases
cross-layer design jointly attunes a number of lower layers’
parameters (e.g., channel state information) to upper-layer
functions like transport and routing [2].
Regarding the underlying motivation, cross-layer design
addresses problems of wireless network performance whose
cause can be traced back to the original design assumptions
underpinning the architecture of the employed network archi-
tectures and their protocol stacks (i.e., the black box
paradigm). The well-known case of TCP’s performance over
wireless links is one of the most commonly cited applications
of cross-layer design, but it is not the only one. Recently,
numerous research efforts from around the globe have used
T
SURVEYS
IEEE
COMMUNICATIONS
The Electronic Magazine of
Original Peer-Reviewed Survey Articles
FOTIS FOUKALAS, VANGELIS GAZIS, AND NANCY ALONISTIOTI, UNIVERSITY OF ATHENS
ABSTRACT
Third-generation (3G) and beyond 3G mobile communication systems must pro-
vide interoperability with the Internet, increase throughput for mobile devices, and
optimize their operation for multimedia applications. The limited ability of tradition-
al layered architectures to exploit the unique nature of wireless communication has
fostered the introduction of cross-layer design solutions that allow optimized opera-
tion for mobile devices in the modern heterogeneous wireless environment. In this
article we present the major cross-layer design solutions that handle such problems,
and discuss cross-layer implementations with a focus on functional entities that sup-
port cross-layer processes and the respective signaling. In addition, we consider the
associated architectural complexity and communication overhead they introduce.
Furthermore, we point out the major open technical challenges in the cross-layer
design research area. Finally, we conclude our article with a summary of cross-layer
approaches developed thus far and provide directions for future work.
CROSS-LAYER DESIGN PROPOSALS FOR
WIRELESS MOBILE NETWORKS:
A S
URVEY AND TAXONOMY
1ST QUARTER 2008, VOLUME 10, NO. 1
www.comsoc.org/pubs/surveys
1553-877X
IEEE Communications Surveys & Tutorials • 1st Quarter 2008
71
cross-layer solutions to improve the performance of wireless
communication systems and protocol stacks in selected appli-
cation areas. These range from the case of TCP’s operation
over wireless networks to more advanced topics such as maxi-
mizing the amount of users per service area (i.e., per radio
cell) and adapting the encoding of multimedia content accord-
ing to the state of the wireless channel.
Nowadays, the potential of cross-layer design for improving
critical performance aspects of modern wireless networks is
widely recognized, and the ability to support cross-layer inter-
action patterns throughout the protocol stack is considered an
important property of beyond 3G mobile communication sys-
tems. The present article surveys existing cross-layer design
applications and summarizes their key properties in a compre-
hensive taxonomy of cross-layer design approaches. The rest
of the article is organized as follows. We present the architec-
tures and models proposed so far within the concept of cross-
layer design in wireless networks and identify key functional
entities in cross-layer management procedures. We present a
taxonomy into which existing cross-layer design efforts are
classified and identify their salient functional properties. We
present and categorize cross-layer signaling approaches. We
identify open technical challenges associated with cross-layer
design proposals and provide a qualitative investigation. We
provide a comparative discussion of cross-layer design solu-
tions and a summary of the lessons learned from the survey.
Finally, we conclude the article with directions for future
work.
CROSS-LAYER ARCHITECTURES,
M
ODELS AND ENTITIES
As so eloquently stated in [3], it is architecture that facilitates
the decomposition of a system’s functions into modular com-
ponents that operate in unison to realize its purpose. By
employing components optimized for specific purposes, a
modular architecture allows for gains in performance and
transparent system upgrades. At a design level, modularity is
achieved by abstracting the functionality offered by each mod-
ule via appropriate interfaces [4]. For instance, the OSI refer-
ence model encapsulates protocol functionalities into entities
(i.e., protocol layers) that form the modules of the proposed
architecture, while abstracting the internal details of each
module through the interfaces it exposes to its adjacent layers
[1].
Cross-layer design (CLD) allows communication to take
place even between nonadjacent layers through additional
entities introduced into the system’s architecture. However,
there is no reference model that specifies the functionality
each new entity (i.e., module) must realize in a cross-layer
design solution. To address this, [5] has proposed a model
for determining the functionality that each new CLD entity
might support. This model introduces four different planes
that extend across the protocol layers of the OSI reference
model in a visually vertical manner. Each of these so-called
coordination planes encapsulates the behavior of a CLD
algorithm or protocol targeted at solving a specific prob-
lem. In wireless mobile devices these problems include
security, mobility, quality of service (QoS), and adaptation
of the wireless link, thus leading to four coordination
planes.
The security plane (Fig. 1) — The security plane coordi-
nates encryption protocols and security technologies across
different layers. Thus far, several encryption methods are
available at various protocol layers. SSH and SSL provide
end-to-end encryption at the transport and the application
layer, and IPSec provides an end-to-end encryption at the
network layer; in IEEE 802.11a/b/g wireless networks, Wired
Equivalent Privacy (WEP) has been superseded by Wi-Fi
Protected Access (WPA) for encryption. If each layer, inde-
pendent of other layers, carries out encryption, unnecessary
duplication of encryption functionality occurs, thus consum-
ing more power, wasting valuable processing resources, and
degrading network performance. Hence, there is the problem
on which a CLD technique focuses is to determine which par-
ticular protocol layer should perform encryption. Thus, if the
use of encryption schemes offered by different layers is coor-
dinated by the security plane pertaining to CLD, the selec-
tion of a single encryption scheme suitable for whatever
security requirements apply is possible — and, of course,
desirable.
The QoS plane (Fig. 1) — Several QoS solutions have
been proposed so far involving various protocol layers, such as
RTP and TCP receiving QoS information from the applica-
tion layer, and the integrated services (IntServ) and differenti-
ated services (DiffServ) architectures developed by the
Internet Engineering Task Force (IETF) support IP QoS.
Developed according to the OSI reference model, these solu-
tions do not support cross-layer communications, and QoS
requirements are not conveyed to layers further along the
protocol stack. In the time-varying wireless environment, how-
ever, the need to communicate protocol state information
from the physical and link layers to the application layer, and
to exploit it for improved QoS (e.g., in real-time data flows) is
compelling. The provision of QoS information between non-
adjacent protocol layers requires a cross-layer design. Hence,
the QoS coordination plane must facilitate the communication
of QoS information and coordinate the provision of QoS
across multiple layers.
The mobility plane (Fig. 1) — Mobility supports the
movement of wireless terminals from one service area to
another through handovers to appropriate radio access points
(i.e., cellular base stations). There are two handover cate-
gories: horizontal handover, where the mobile device moves
between access points of the same technology, and vertical
handover, dealing with mobile device movements between
access points of different technologies. In both cases, upper
layers must be able to mitigate the effects of handover, so
mobility-related functionality must support the generation of
notifications about handovers [18]. That will facilitate a
smooth — and, ideally, seamless — transition of the mobile
device’s applications to the new wireless technology. To this
end, the mobility coordination plane would take care of
adapting the upper-layer services to the underlying wireless
technologies.
■ Figure 1. A cross-layer coordination model [5].
Security
Mobility
Wireless link adaptation
Quality of service
Application
Transport
Network
Link
IEEE Communications Surveys & Tutorials • 1st Quarter 2008
72
The wireless link adaptation plane (Fig. 1) — This plane
addresses effects specific to the wireless link, i.e., channel fad-
ing, bit error rate (BER) variations, and transmission delays.
These properties can affect the performance of upper layers,
particularly that of TCP, which erroneously considers packet
losses attributed to the instant state of the wireless channel as
being caused by congestion in the end-to-end path. Several
cross-layer solutions to indicate the actual cause of packet
losses occurring at lower layers have been proposed thus far
(e.g., the TCP-sleep protocol identifies losses related to chan-
nel fading effects). In this case the automatic repeat request
(ARQ) protocol at the data link layer tries retransmissions.
Obviously, retransmission of lost packets from both TCP and
ARQ could halve the congestion window and, thus, the uti-
lization of the wireless link. To avoid such rate degradations,
the coordination between TCP and ARQ protocols is neces-
sary [17].
Another important CLD aspect is the management of
cross-layer interactions in a way that can guarantee the sys-
tem’s smooth operation. To this end, the aforementioned
management model specifies an interlayer coordination man-
ager responsible for the central coordination of CLD process-
es. In general, CLD introduces management entities that
operate as either an optimizer of performance or a scheduler
of some kind, depending on the problem at hand. Such an
entity may reside within the protocol stack of the affected sys-
tem, in which case it is considered an internal entity, or in an
external network node. In the former case, the internal entity
may be either an interlayer entity that coordinates the opera-
tion of all protocol stack layers or a set of intralayer entities,
each of which is collocated with a protocol layer (Fig. 2). In
the case of external entities, these may be centralized and
hosted by a specific network node or distributed over several
network nodes.
TYPES OF CROSS-LAYER MANAGEMENT ENTITIES
Internal Interlayer Entities
Interlayer Cross-Layer Manager (Fig. 2) — Reference [5]
proposed a central interlayer coordination manager that
applies cross-layer algorithms in any protocol stack layer. The
coordination manager receives notifications for events occur-
ring at protocol layers and is thus aware of the specific state
any protocol layer is in. For instance, in the TCP case the
congestion window and BER threshold state variables are
used to trigger the connection initiated and link lost events,
respectively.
Interlayer Cross-Layer Optimizer (Fig. 2) — Reference [6]
proposed a CLD architecture that jointly optimizes the opera-
tional parameters of multiple layers via a cross-layer optimizer
(CLO) entity responsible for optimizing N layers based on
abstracted layer parameters. The abstraction of the layer
parameters reduces the number of parameters the CLO needs
in order to optimize the layer functionality. The benefit of this
approach is that it provides a technology-independent way of
interacting with each protocol layer. Consequently, the CLO
can be deployed in heterogeneous networks comprising differ-
ent wireless technologies and access systems. Layer abstrac-
tion identifies the parameters that expose the capabilities of
the corresponding layer, thus enabling calculation of the prop-
er values that optimize a specific objective function by the
CLO. For instance, in the case of audiovisual transmission,
the objective function to optimize may be the average peak
signal-to-noise ratio (PSNR) that translates to the video quali-
ty perceived by a user. The performance criterion of the cross-
layer optimization is the average PSNR between the encoded
and displayed video stream, calculated through the rate dis-
tortion (RD) factor. The CLO employs a reconfiguration pro-
cedure to distribute the values of the abstracted parameters to
■ Figure 2. The different entities which (a, b) coordinate cross-layer management procedures in a protocol stack; and (c, d) which pro-
cess cross-layer information in a network deployment.
Physical
a) Internal intralayer entities c) External centralized entities
b) Internal interlayer entities d) External decentralized entities
CLE
CLE
BS
Link
CLE
MT
MT
CLE
CLE
MT
CLE
BS BS
BS
IEEE Communications Surveys & Tutorials • 1st Quarter 2008
73
the corresponding protocol layers. Each protocol layer is then
responsible for matching the abstracted parameters and values
into its own (i.e., internal) parameters that adapt its mode of
operation. This approach incurs communication overhead due
to the cross-layer information (i.e., the RD information) being
conveyed from the video server to the CLO as well as some
processing overhead during the reconfiguration process.
Internal Intralayer Entities
Intralayer Cross-Layer Optimizer (Fig. 2) — Reference [7]
introduced a model for designing and implementing cross-
layer feedback to allow direct communication between any
pairs of layers in the protocol stack. This CLD model is called
ÉCLAIR and consists of the following modules:
• The tuning layer (TL) provides an interface for invoking
control information at a particular protocol layer. As
control information specifies the behavior of protocols,
the actual protocol behavior can be changed by properly
manipulating its control information.
• The optimizing subsystem (OSS) activates optimization
algorithms. The OSS collects control information from
the TL through the protocol optimizers and adapts the
protocol’s behavior during runtime. To this end, the OSS
contains a set of protocol optimizer (PO) entities. A PO
implements an algorithm that addresses a specific cross-
layer optimization. Hence, several specialized PO entities
can be implemented and deployed according to the opti-
mization purposes.
The ÉCLAIR architecture has been used in a feedback
loop to control the bandwidth of running applications by tun-
ing the receiver window of each TCP connection. In the
ÉCLAIR architecture, applications set up the desired TCP
window size to advertise their restrictions on network through-
put. ÉCLAIR assigns a priority to each application and calcu-
lates the appropriate receiver window based on that priority.
More specifically, the TL for the TCP layer (TCPTL) uses the
priority to calculate the receiver window for each application.
In this particular approach, changes in the protocol stack will
only affect TL entities and the functionality of PO entities;
hence, a PO does not depend on protocol layer code. All the
same, the additional function calls between the OSS and the
TL through the several POs incur internal overhead. Howev-
er, ÉCLAIR achieves the following design goals:
• Minimal or zero processing overhead within the protocol
stack since the OSS is executing concurrently with the
protocol stack.
• Several PO entities can be dynamically deployed and
ported to multiple technologies.
Intralayer Cross-Layer Scheduler (Fig. 2) — Reference [8]
proposed a cross-layer adaptation framework for 802.16e
orthogonal frequency-division multiple access (OFDMA) sys-
tems. It strives to achieve the highest system performance by
exploiting cross-layer information between the medium access
control (MAC) and physical layers. The MAC layer consists
of the scheduling and resource allocation components that
comprise a MAC scheduler and resource controller, respec-
tively. The scheduler’s algorithm determines the number of
packets that could be transmitted to each user. The resource
controller allocates the frequency bands for each user by using
a channel-aware subcarrier allocation algorithm. In addition, a
user grouper organizes individual users into groups according
to the subchannel type of the 802.16e OFDMA standard.
The scheduler works in conjunction with the resource con-
troller to increase the achievable throughput of users based
on the channel quality information (CQI). Moreover, a hybrid
ARQ protocol is deployed for purposes of link adaptation.
The use of HARQ aids in the selection of the modulation and
the coding scheme. Although the involvement of HARQ
increases overall throughput, it also introduces a considerable
amount of overhead when a large number of retransmissions
occur. In 802.16e systems, retransmissions are associated to
certain control messages that allow a mobile terminal to iden-
tify the correct packets in a data burst. However, these control
messages can accrue up to 60 percent of the resources that
should be allocated to mobile users. Consequently, allocation
of channel resources should take into account in the amount
of control messages and retransmissions.
External Entities
External Radio Scheduler (Fig. 2) — Reference [9] pro-
posed a scheduling strategy for wideband CDMA (WCDMA)
systems such as the Universal Mobile Telecommunication Sys-
tem (UMTS). This strategy exploits cross-layer information to
improve system performance in terms of capacity and delay. It
assigns users to priorities based on short-term channel varia-
tions instead of using only long-term ones. In a WCDMA sys-
tem the radio base station (BS) provides each user with a
transmission power P
i
(t). The BS sets the available transmis-
sion power P
T
and, using a downlink fast control mechanism,
notifies each wireless device of the minimum transmission
power. Due to the slowly varying radio channel conditions, the
power fluctuates around an average value.
By merging the rapid channel fluctuations of WCDMA, [9]
proposed a scheduling scheme that prioritizes radio transmis-
sions using a function that exploits these rapid channel fluctu-
ations. This function takes into account not only the channel
state (e.g., as typical multiuser diversity does), but also the
channel variation experienced by each user. These priorities
are evaluated for the downlink channel. To handle downlink
radio conditions, a radio scheduler located in the BS need not
use signaling to invoke power-related information since it uses
fast power control information from the power control mecha-
nism. Hence, in principle, the proposed scheme does not
affect system performance, and its deployment in UMTS
mobile networks could increase the number of served users.
Simulation results quantify the realistically achievable gain of
this strategy as up to 30 percent in capacity and 35 percent
reduction in average channel access times.
External Centralized Cross-Layer Optimizer (Fig. 2) —
Reference [10] proposed a CLD solution to address QoS pro-
visioning over IP-based CDMA networks. The authors pro-
posed a centralized cross-layer scheduler located at the BS
that interacts with mobile terminals to exchange information
regarding its traffic, power level, etc.eteras. This cross-layer
scheduler supports a Dynamic Weight Generalized Processor
Sharing (DWGPS) scheduling scheme according to which a
video frame from the application layer is compressed to sever-
al batches of link layer (LL) packets according to its priority.
To this end, the mobile terminal sends the batch class and
batch size to the BS. The BS is also aware of the maximum
tolerable delay over the wireless link as denoted by the time-
out value.
This proposal takes into account the multiuser diversity
gain that denotes the ratio of average transmission power for
an LL packet. A piece of information also considered is the
good/bad threshold F, since the bad channel state of a batch
affects the backoff probability. Consequently, this threshold
must be carefully set to avoid degrading the effectiveness of
the backoff functionality. Whether a channel is in a bad or
good state is a critical issue that must be estimated carefully,
IEEE Communications Surveys & Tutorials • 1st Quarter 2008
74
since in CDMA systems a mobile terminal does not transmit
only when it has the best channel quality but also when its
channel gain is no more than F dB less than the average
value. Hence, the backoff probability must be small when the
timer value for transmission is low, in which case the batch
must be transmitted urgently. Moreover, the channel fading
rate affects the backoff probability and consequently the
timer’s value.
Distributed Cross-Layer Optimizer — The majority of
cross-layer designs focus on single-hop wireless networks (i.e.,
cellular networks). The aforementioned approaches concern
the conventional case where a single access point serves
mobile devices using radio cells. On the other hand, cross-
layer design for resource allocation has already been applied
in multihop wireless networks. Reference [11] discusses cross-
layer design algorithms that operate in a distributed fashion.
A decentralized scheduler is employed to schedule L links
simultaneously satisfying the interference constraints imposed
by the distributed network model and the associated general
interference model. Such links are dominated by interference;
consequently, they may suffer significant capacity losses in
packet transmission and reception. Naturally, the introduction
of a distributed and decentralized scheduling scheme intro-
duces additional signaling overhead that,ultimately depends
on the actual network size and topology [2].
Table 1 depicts the types of cross-layer management enti-
ties dealing with cross-layer optimization, scheduling, and
information management in general. It also presents the asso-
ciated overhead according to the reviewed works.
CROSS-LAYER SOLUTIONS FOR
MOBILE COMMUNICATIONS
CLD solutions have been proposed to address different prob-
lems that arise due to the evolution of wireless mobile com-
munications. The provision of Internet services over mobile
communication networks has been the driving force in this
evolution. On the other hand, the need to serve the largest
possible population of users in next-generation mobile com-
munication networks can be satisfied through cross-layer
design that exploits valuable properties of the wireless chan-
nel. Cross-layer design optimization solutions can provide
improved QoS to the mobile terminal for its multimedia
applications. In the following we present a (nonexhaustive) list
of cross-layer solutions developed for wireless mobile commu-
nications.
IMPROVING THE PERFORMANCE OF
TCP OVER WIRELESS NETWORKS
Many CLD solutions have been introduced in order to
improve TCP’s performance over wireless links. Drawing the
problem in outline, TCP invokes error and congestion control
mechanisms such as the retransmission of TCP segments and
■ Table 1. Types of cross-layer functional entities and their impact.
Management entity Objectives Incurred overhead
Internal interlayer entities
Interlayer cross-layer manager
Cross-layer design implementation based
on an internal interlayer cross-layer manag-
er, that manages the entire protocol stack.
N/A
Interlayer cross-layer optimizer
A cross-layer optimizer that is responsible
for optimizing N layers with respect to an
application-oriented function.
It causes external overhead when the cross-
layer information is passed from the net-
work to the terminal.
Internal intralayer entities
Intralayer cross-layer optimizer
Each layer includes protocol optimizers that
employ the appropriate algorithm for opti-
mization purposes.
It incurs internal overhead due to the addi-
tional function calls.
Intralayer cross-layer scheduler
The scheduler and the controller can jointly
improve users’ throughput based on the
channel quality information (CQI).
The control messages and retransmissions
that the MS needs in locating its packet
within a burst incur external overhead.
External entities
External radio scheduler
The radio scheduler prioritizes radio trans-
missions based on the channel state and
the channel variation of each user.
By employing fast power control informa-
tion in base stations, external overhead due
to signaling is avoided.
External centralized cross-layer optimiz-
er
The MS sends to the BS the batch class, the
batch size and the maximum tolerable
delay indicating the timeout value.
The information exchange overhead is not
significant. However, the good/bad state for
a MS is a critical issue and must be estimat-
ed carefully.
External distributed cross-layer optimiz-
er
The decentralized scheduler is employed to
schedule L links satisfying simultaneously
the interference constraints imposed by the
distributed network model.
A decentralized scheduling approach intro-
duces signaling overhead depending on the
network size and topology.
IEEE Communications Surveys & Tutorials • 1st Quarter 2008
75
the reduction of the congestion window whenever losses are
detected, even though these may not be a result of congestion
(i.e., losses caused by data corruption in the wireless medi-
um). The inability of TCP to correctly identify the cause of
packet losses is tackled by indicating explicitly either network
congestion or transmission errors. In addition, the properties
of underlying technologies mainly at the physical and MAC
layers are exploited in an effort to improve TCP’s perfor-
mance [2]. The following subsections present the relevant
cross-layer solutions.
Indicating Network Congestion — Wireless link reliability
is questionable due to transmission errors. Link-layer mecha-
nisms that tackle this deficiency are forward error correction
(FEC), ARQ, and HARQ. The FEC mechanism enables the
receiver to detect and correct errors [12]. As opposed to FEC,
ARQ does not provide any error detection or correction, but
solely grants frame retransmission from transmitter to receiver
[13]. HARQ, in general, is a combination of FEC and ARQ.
Particularly, it corrects transmission errors; however, if the
channel quality is not at a good level, the receiver performs
error detection before requesting retransmission. HARQ is
recently deployed in 3G wireless systems and, in conjunction
with adaptive modulation and coding (AMC), improves the
performance of TCP over 3G wireless systems [14]. Nonethe-
less, even though these mechanisms improve the wireless
channel’s reliability, TCP will still treat all losses as conges-
tion-related.
In wired networks, when congestion occurs, it is dealt with
by the Adaptive Queue Management (AQM) mechanism
offered by network routers. Consequently, it prevents the
potential delays due to duplicate acknowldegments (ACKs)
and packet retransmission. More specifically, router networks
support Random Early Detection (RED) algorithms based on
the average queue length exceeding a threshold. Explicit Con-
gestion Notification (ECN) notifies the receiver of congestion
in the end-to-end communication path. The congestion is
indicated using a 2-bit-long ECN field in the IP header. On
the other hand, ECN-capable TCP contains two additional
fields in the TCP header for TCP-endpoint to TCP-endpoint
signaling [15]. If the sending TCP entity is informed of con-
gestion-related losses, it will avoid redundant retransmissions
and thus facilitate the proper operation of congestion control
at the TCP layer. However, as mentioned, the ECN mecha-
nism was designed for a wired network and does not indicate
the transmission error on the wireless link [16].
Indicating Transmission Errors — The Explicit Loss Notifi-
cation (ELN) scheme notifies the TCP sender of packet losses
caused by reasons unrelated to network congestion [19]. A
snoop agent located at the BS treats a packet loss as a corrup-
tion in the wireless medium; however, this agent does not pro-
vide local recovery through packet retransmissions as does the
agent in the snoop protocol [16]. More specifically, it retains a
sequence block of ACKs indicating the successful transfer of
packets from the sender to the receiver. It compares the pre-
vious with the newly arrived ACK value. If there is a gap in
the sequence of received ACKs due to a packet loss, it sets
the ELN bit. The ELN bit is contained in the TCP checksum
since there is currently no specific bit in the TCP header for
ELN. Whenever an ACK is received successfully, the agent
cleans up the old block and retains a new one. The notifica-
tion is passed to the sender by using either TCP header
options in the packet header or Internet Control Message
Protocol (ICMP) messages.
ECN and ELN are technology-independent and do not
require any particular wireless network architecture or radio
technology. There are, however, cases where the susceptible
performance of TCP can be improved by exploiting the char-
acteristic features of the underlying radio technology.
Exploiting Properties of Underlying Technologies — Due
to the nonorthogonal nature of signals in a CDMA system,
the interference among a user’s substreams degrades the TCP
performance of each user. When TCP users demand substan-
tial throughput, interference among the associated radio sig-
nals downgrades the TCP transmission capacity [20]. Hence,
in CDMA networks (e.g., multicarrier CDMA) where the
available capacity is interference-limited, the objective is to
improve TCP’s performance with minimum impact on inter-
ference. However, such a solution requires cooperation
between link layer resource allocation and TCP. Therefore,
TCP should exploit the wireless link layer properties in order
to achieve the target TCP throughput with the minimum pos-
sible amount of resources. In [21] the required resource
amount depends on a resource vector denoted (M, Γ). The M
value denotes the packets that can be scheduled for transmis-
sion in a slot, and the Γ value denotes the required bit-energy-
to-interference-plus-noise density ratio (E
b
/I
0
) for all M
packets. In the transport layer in particular, M
i
is the target
number of scheduled packets for TCP flow i, with a Γ
i
value
for the signal-to-interference-plus-noise ratio (SINR) level of
the link layer unit (i.e., frame). The resource vector (M
i
, Γ
i
)
determines the packet loss rate and transmission delay over a
wireless link.
AMC is a widely known technique that is pertinent to
matching the transmission rate to time-varying channel condi-
tions [22]. It has been deployed in both WCDMA and
WiMAX wireless broadband networks, and can realize several
benefits for TCP’s performance over wireless links [23, 24].
Reference [26] advocates a CLD approach that effectively
conflates AMC with TCP in order to maximize TCP through-
put. In particular, while sustaining a prescribed packet error
rate (PER) P
0
, better TCP performance can be achieved in
terms of throughput by maximizing the data rate the AMC is
able to render. By selecting the channel-dependent parame-
ters such as the average of the received signal-to-noise ratio
(SNR) g, the mobility-induced Doppler spread f
d
, the fading
parameter m, and the number of packets K the data link
layer’s queue can serve as well, the TCP throughput is
improved for a prescribed P
0
.
INCREASING THE RATIO AND CAPACITY OF SERVED USERS
Increasing the Ratio of Served Users — In cellular net-
works, multiple access methods have been used for the trans-
port of voice data between BSs and mobile devices. These
include time-/frequency-division multiple access (TDMA/
FDMA) methods. Data transmission and telephony, however,
create bursty traffic. In such a case, resources allocation like
the allocation of a fixed number of time slots by transmitters
in the TDMA method leads to underutilization due to the
nature of traffic. Thus, even if a user is not transmitting any
data (i.e., voice), the transmitters have already assigned the
time slot to him/her; hence, the dedicated voice channel aim-
lessly consumes bandwidth [13]. The same problem is posed in
FDMA-based communication systems. Consequently, the way
to handle bandwidth efficiently is to allocate it in a flexible
manner by using cross-layer techniques.
Asynchronous CDMA uses spectrum more efficiently than
static allocation methods (TDMA/FDMA) when traffic is
bursty in nature. More specifically, flexibility in spectrum allo-
cation is provided by scheduling algorithms that decide which
users are permitted to transmit during a specific time slot.
IEEE Communications Surveys & Tutorials • 1st Quarter 2008
76
The decision depends on channel state information (CSI) fed
back from users to BSs and is known as channel-state-depen-
dent scheduling. In this case mobile users send CSI feedback
to the access points indicating the effective received rate at
which data can be transmitted. Such information is generally
the received SNR from a certain user [20]. Thus, there is no
need to continually reallocate time slots; on the contrary, only
users with a good level of BER consume time slots. The gain
achieved by this mechanism is known as the multiuser diversi-
ty gain [13].
Furthermore, 3G wireless systems have deployed cross-
layer methods that exploit CSI in order to increase the served
users and reduce undesired power radiation. These methods
consist of policy scheduling algorithms that take into account
information related to the user’s channel state. As mentioned
above, [9] has proposed such a scheduling algorithm using
cross-layer information to provide more capacity and smaller
delays for each mobile user. This algorithm gives priorities to
each user based on short-term channel variations instead of
using only long-term variations as was typically done in
CDMA wireless networks. The short-term information
expresses the knowledge of the transmission power variation
at the frame level, while the long-term information implies the
expected future conditions, which include the expected future
value of the required transmission power for the next frame.
The proposed algorithm has been simulated for UMTS down-
link channels and could be applied in the current UMTS
radio resource management (RRM) framework. The selection
of users in terms of transmission power allocation has been
made with respect to the following priority rules:
• Users with better channel conditions (i.e., requiring lower
transmission power in the downlink) should have higher
priority.
• Users with improving channels (i.e., the current needs in
terms of transmission power are lower than those needed
in the last N frames) can be afforded higher priority.
• Users that experience bad channel conditions for a rather
small number of consecutive frames should be provided
with a priority that compensates for their transmission
delays.
Increasing the Capacity of Served Users — Reference [10]
has also applied the multiuser diversity concept to real-time
traffic. Due to the time-varying feature of CDMA channels,
multiuser diversity provides the opportunity to use channel-
aware scheduling methods employing the good/bad threshold
value. More specifically, the data rate that can be achieved in
CDMA channels is related to the SINR of one mobile station
(MS). In CDMA the interference caused is intercell interfer-
ence. In the downlink, if the BS only transmits to the MS with
the highest SINR at time instant t, the maximum system
throughput will be achieved. On the contrary, in the uplink, if
only one MS transmits to the BS, the minimum intercell inter-
ference will be achieved. In a multicell environment, channel-
aware scheduling leads to an increase in total system capacity.
However, in a CDMA system an MS does not transmit only
when it has the best channel quality, but also when its channel
gain is not F db less than the average value. As mentioned
above, the F value implies the good/bad threshold.
Moreover, the capacity of served users is realized through
the spectral efficiency. Reference [27] describes a CLD
between physical and data link layers in order to achieve high-
er spectral efficiency. This combination contains an AMC
scheme and an ARQ mechanism at the data link layer. Given
that the maximum number of retransmissions must be within
delay constraints, an optimal design of the AMC scheme at
the physical layer should be targeted to maximize the spectral
efficiency. If the maximum number of retransmissions allowed
per packet is N
r
max
, the probability of packet loss after N
r
max
retransmissions must be no larger than an upper bound
imposed by the application requirements (e.g., video transmis-
sion). After that, the joint design consists of an AMC that sat-
isfies a PER upper bound at the physical layer and an ARQ
that grants the N
r
max
upper bound at the data link layer. The
experimental results of this approach show that a small num-
ber of retransmissions in conjunction with the chosen modula-
tion-coding pair (mode), the latter determining the SNR level
of the communication channel, can improve spectral efficiency
in terms of bits per transmitted symbol. Of course, an arbi-
trary increase of retransmissions severely downgrades spectral
efficiency.
In much the same concept, [28] combines AMC with an
HARQ scheme against the truncated ARQ used in [27]. More
specifically, [28] uses a type-I HARQ mechanism for packet
retransmission and aims for maximum optimization under a
prescribed delay constraint. The target PER determines the
target BER for a transmission block and, in turn, the carrier-
to-noise ratio (CNR) region of the AMC scheme can be
determined. However, contrary to [27], which used a fixed
packet size, [28] takes into account a variable packet size L. It
is demonstrated that by adjusting the packet size in conjunc-
tion with the CNR region’s level, a high spectral efficiency can
be achieved. Furthermore, [29] combines type-II HARQ with
AMC against the pure ARQ mechanism. Moreover, it pre-
sents a comparison between type-II and type-I HARQ mecha-
nisms by numerical results. Particularly, the authors observed
that type-II HARQ improves spectrum efficiency more than
type-I HARQ does in the high CNR region specified by the
■ Table 2. The application categories on which most cross-layer design solutions focus.
Objectives Interactions Summary
Improving the performance
of TCP over wireless networks
Between data-link layer, network
and transport layers.
Using notification mechanisms, the indication of network conges-
tion and wireless channel errors is accomplished. The exploitation
of parameters (i.e., channel conditions, frame size) from the
lower layers helps improve TCP performance.
Increasing the ratio and the
capacity of served users
Physical and data-link layer co-
design and consolidation.
The ratio of the served users is increased by using the multi-user
diversity and the power control mechanisms respectively in a
cross-layer fashion. The co-design of AMC at the physical layer
and HARQ at the data-link layer provides more capacity to users.
Application and optimization
at the mobile application
layer
Application and physical layer
synchronization. Data -link layer
involvement is possible.
Application-driven cross-layer optimization approach is evaluated
using an application objective function (e.g., PSNR). The lower
layers provide multi-user diversity, power control and knowledge
of the frame’s type. Statistics-based estimations at the data link
layer enhance the optimization process.
IEEE Communications Surveys & Tutorials • 1st Quarter 2008
77
AMC scheme. On the other hand, the smallest packet size L
amplifies spectral efficiency.
Reference [30] also proposed a CLD methodology based
on an AMC scheme at the physical layer in conjunction with a
scheduling mechanism at the MAC function of the data link
layer. It proposes a scheduler that considers the channel state
of the physical layer and the queue state of the data link layer.
The queue is a finite-length buffer that is implemented at the
gateway of the wireless access network. The adopted access
mechanism is the time-division multiplexing (TDM)/TDMA.
For each user i the numbers of time slots that can actually be
scheduled are assigned depending on both the channel state
and queue state. The reserved time slots are conditioned by
the prescribed QoS levels for each user.
ADAPTATION AND OPTIMIZATION AT THE
MOBILE APPLICATION LAYER
Cross-layer optimization for wireless video streaming can off-
set the end-to-end distortion caused in the received video at
the application layer [6]. Such a cross-layer solution is accom-
plished by exchanging information between tje source and
channel coder in the application and physical layers, respec-
tively. This cross-layer information is known as source signifi-
cance information (SSI) [31]. Unequal Error Protection
(UEP) is such a cross-layer approach that protects important
information from impairments caused by channel errors. This
approach is also known as Joint Source/Channel Coding
(JSCC) [32]. In CDMA networks the use of power control
improves the error probability by combating the negative
effects of interference. More specifically, the power level is
determined by the energy-per-bit to multiple access interfer-
ence (MAI) density ratio γ
b
= E
b
/N
0
. Reference [32] presents
an evolutionary approach tp JSCC that proposes joint control
between the source coding and power control in terms of the
source rate R
s
of the video codec and the average γ
b
of the
physical layer, respectively. This approach is known as Joint
Source Coding-Power Control (JSCPC) and attempts to
achieve an end user’s QoS level by adjusting the combination
of R
s
and γ
b
.
However, adaptation at the application layer such as video
streaming could benefit from link adaptation as well. Refer-
ence [33] presents such a combined solution that improves
real-time streaming video quality by adapting the video encod-
ing rate at the application layer according to MAC layer
statistics-based information (e.g., throughput) in conjunction
with the physical layer. Especially in the case of multiuser
access, throughput predictions at the MAC layer give feed-
back to the systems’ combination.
On the other hand, application layer optimization should
be based on a user-perceived video quality factor like the
quantitative parameter PSNR that indicates video distortion.
To this end, in JSCPC methodology the optimal operational
distortion rate for a single user with effective bandwidth
requirement W
eq
in hertz is given as
PSNR(W
eq
) = max PSNR(R
s
, γ
b
). (1)
Reference [6] seeks to maximize the expected user-per-
ceived video quality. This goal can be achieved by selecting
the optimal parameter values for each group of pictures
(GOP). A GOP is a sequence of groups of consecutive frames
of the video stream. The expected user-perceived video quali-
ty implies that the video reconstruction quality at the user’s
side is
D = D
s
+ D
L
. (2)
The D
s
term is a source rate function and defines the
reconstruction quality in the error-free case, called source dis-
tortion. The D
L
term is a packet loss rate function and repre-
sents the distortion derived from the transmission. The D
L
is
called loss distortion. By combining the so-called distortion
profile information (i.e., the rate vector and distortion matrix)
plus the transmission probabilities from the two-state Markov
packet burst loss model, the radio link parameters can be cho-
sen. Using this approach, the expected quality can be achieved
for a particular application, and radio link layer parameters
can be set with respect to the desired objective function (i.e.,
the PSNR) depending on the value of D.
Moreover, JSCC in conjunction with the multiuser diversity
methodology can improve the QoS in time-varying CDMA
channels. Reference [10] proposes such a cross-layer approach
called multiuser adaptation, which exploits information in the
application and physical layers. In previous sections we dis-
cussed the LL units named batches as well as the class and
arrival batch size that represent the crucial cross-layer infor-
mation. In the transmission queue within the BS, each batch
LL unit is managed according to its priority for the corre-
sponding session. In order to exploit multiuser diversity in
cross-layer design, each batch i is determined as in bad chan-
nel state when the channel fades fast, as mentioned above.
Whether a batch is in a bad channel state depends on the
aforementioned good/bad threshold F. Comparing the thresh-
old values F = 10 dB and F = 5 dB, it is inferred that in the
case of F = 5 dB the service degradation in terms of LL unit
losses is greater than in the case of F = 10 dB. This is because
the threshold F = 5 dB sustains the bad channel state proba-
bility of one MS, and the corresponding batch is considered
idle for a longer period of time.
Thus far we have classified and discussed a representative
list of CLD proposals in terms of their objectives. Table 2 pre-
sents the interactions between layers that are accomplished by
each of these solutions.
CROSS-LAYER SIGNALING
Given that there are no particular restrictions in the way
cross-layer signaling takes place, various approaches have
been adopted by the implementations thus far [34]:
• An additional packet header carries forward or indicates
cross-layer information (CLI).
• Header options report changes in lower layers.
• Profiles and labels contain and indicate the meaningful
CLI, respectively.
• A network service collects and distributes related CLI.
Inevitably, cross-layer signaling will incur some overhead;
[34] points out the evaluation criteria of cross-layer signaling
methods and presents a representative comparison. All the
same, research in cross-layer signaling and notification mecha-
nisms must strive to find efficient answers to a set of intrigu-
ing questions: What is the best way of indicating CLI, a
header option or an additional header? Should cross-layer sig-
naling be an in-band or an out-of-band one? In what format is
CLI stored and conveyed? In which part of the network
should an entity (i.e., server, router, manager etc.) be located
for collecting and managing CLI? What is the overhead intro-
duced by signaling during the setup phase or throughout the
session?
To this end, next we extend the presentation of [34], point-
ing out the different types of cross-layer signaling mechanisms
such as in-band and out-of-band signaling, the employed pro-
tocols, the format of transport messages/files, and the distribu-
tion of introduced overhead. We focus on the signaling
IEEE Communications Surveys & Tutorials • 1st Quarter 2008
78
protocol, the distribution pattern on network nodes (servers,
routers, gateways, etc.) involved in it, and the potential over-
heads it incurs. As the detailed investigation of CLD signaling
options is still in progress [35], a detailed exhaustive presenta-
tion is beyond the scope of this article.
SIGNALING MECHANISMS AND PROTOCOLS
Reference [41] presents the design concept of interlayer sig-
naling and notification termed the hints and notifications
(HANs) concept. Hints are messages traveling in the down-
ward direction (i.e., from higher toward lower layers of the
protocol stack) and are merely additional information on
packet internals. Hints can be either an additional header or
an in-band protocol (e.g., IPv6) [34]. On the other hand, noti-
fications are messages rising from the lower protocol layers to
notify the upper protocol layers regarding the current opera-
tional state (i.e., channel state, network congestion). To this
end, notifications could be header options of in-band proto-
cols (e.g., IP and TCP header options for ECNs) [35]. Howev-
er, separate out-of-band protocols such as ICMP and RTCP
are also candidates for notifying upper layers [31, 37].
In case of local profiles, two mechanisms are required, as
discussed in the following. One mechanism is a transport
mechanism that, in respect to the local adaptation (LA) con-
cept presented in [38], could be performed by an XML-based
mechanism or, alternatively, using an additional packet head-
er. In this way a collection of parameters called layer-indepen-
dent descriptors (LIDs) are transferred in order to be
organized and stored in profiles locally. Each packet carrying
audiovisual data is associated with an LID profile that indi-
cates the adaptation capabilities of the content to lower proto-
col layers. This indication is a label in the IP packet header
that associates the current packet with a specific profile,
resulting in a so-called LID-label binding established in net-
work routers. The Resource Reservation Protocol (RSVP), an
out-of-band signaling protocol, is used to distribute the bind-
ing between LIDs and labels across routers [35].
On the other hand, [39] facilitates the transmission of CLI
by means of a network service. The network service is realized
through a so-called wireless channel interface (WCI). The
WCI plays the role of mediator between network operators
and mobile clients. The WCI employs an abstraction engine
that first treats the parameters gathered and then distills them
in an abstraction format that is meaningful to the mobile
clients. A WCI server gathers parameters from the radio BSs
and other network elements, and makes them available or for-
wards them to its clients through a proxy server that also
implements the abstraction engine. Mobile clients receive or
retrieve these parameters from the network server in the form
of XML descriptors. However, a mechanism is required to
select parameters from the network. SNMP is used for moni-
toring and selecting parameters that can be deployed as either
an in-band or out-of-band mechanism [40].
SIGNALING ASPECTS AND EVALUATION
Signaling messages derived from lower layers are directed in
most cases to the transport and application layers. The propa-
gation of messages to the TCP layer, whether by an in-band
(e.g., ECN) or out-of-band (e.g., ICMP, RTCP) approach,
ends up in a specific socket. On the other hand, the propaga-
tion of notifications to the application layer ends up in a par-
ticular application. Frequent notification propagations pose
internal overhead problems when those signals are delivered
to application and transport layers [35]. Therefore, the higher
the number of sockets and applications, the greater the
incurred internal overhead. Moreover, the message type for-
mat (i.e., header options, additional header) affects the associ-
ated host’s performance [36].
However, considering signals transferred from one side
(the sender) to the other side (the receiver), the focus for
localizing the possible incurred overhead should be addition-
ally on the signaling mechanism (i.e., in-band or out-of-band).
It is generally argued that in-band messages create less inter-
nal overhead than a separate message (i.e., out-of-band) does
[36]. On the other hand, out-of-band messages preserve the
normal protocol flow from additional external overhead [35].
All the same, in such a case the processing (i.e., internal at
end systems and routers) overhead due to additional packet
flow can be considerable [34]. Furthermore, the additional
control path maintained for the out-of-band messages also
incurs external overhead.
Overheads are incurred during session setup or raised
through a session. In the case of local profiles, there is no
need for transferring CLI throughout the session; hence, the
external overhead of cross-layer signaling is limited to the ses-
sion setup phase [38]. In network service signaling, either a
mobile client or a WCI server initiates an application session.
In this phase negotiations regarding cross-layer parameters
and formats take place between mobile clients and the WCI
server through the proxy [39]. Thus, the majority of external
overhead is limited in the negotiation phase in this CLD sig-
naling method. On the other hand, internal overhead will
depend on the complexity of the abstraction engine.
Figure 3 illustrates the different ways of conveying CLI
between and across protocol layers. Signaling aspects such as
transport and signaling mechanisms in cross-layer communica-
tion are summarized in Table 3.
ELABORATING ON
OPEN TECHNICAL CHALLENGES
As the area of cross-layer design is further developed, con-
cerns are being voiced regarding its architectural repercus-
sions, calling for a more cautionary approach in its use as a
design artifact [3]. Some of the open challenges that lie on the
path towards an efficient resolution of these concerns are [4]:
Standardization of Interfaces/Mechanisms for CLD — As
previously discussed, the CLD architecture should provide the
functionality for its own modules. To this end, an important
question concerns the potential interfaces between these mod-
ules. The need for information exchange and sharing between
nonadjacent protocol layers will determine these interfaces.
Moreover, the layer parameters will indicate the flow of infor-
mation between layers and therefore the direction of informa-
tion exchange inside the protocol stack [3]. To this end, layer
abstraction can expose the mechanisms and parameters of
each layer [6].
Standardization can provide a unique vehicle for smoothly
deploying various cross-layer design solutions in next-genera-
tion mobile communication networks. However, the investiga-
tion, specification, development, and, ultimately,
standardization of cross-layer entities, interfaces, and algo-
rithms to meet the need for cross-layer optimizations and
dynamic interaction patterns between the protocol layers
remain an open technical challenge.
The Coexistence of Different CLD Solutions — The main
consideration for this issue is whether CLD solutions that
intend to solve the same problem could be independently
deployed in a transparent manner. For example, how can a set
IEEE Communications Surveys & Tutorials • 1st Quarter 2008
79
of different cross-layer scheduling algorithms based on AMC
be deployed at different times without changing the physical
layer with regard to the set of exploited parameters? Are
there common mechanisms different CLD approaches may
use? If so, the dynamic deployment of different algorithms
could be achieved with minimal impact on each individual
layer’s implementation. However, that does not preclude the
case of adopting a custom mechanism for each particular cir-
cumstance.
The Role of the Physical Layer in CLD — In wireless net-
works the physical layer plays an important role. Advanced
signal processing at the physical layer provides valuable func-
tions such as rate adaptation, channel-aware scheduling, and
subcarrier allocation. Nonetheless, the inherent variability of
the wireless medium may impact the function of network layer
protocols, thus affecting end-to-end performance. CLD main-
ly relies on the unique features of the physical layer to achieve
better QoS over multicell wireless networks such as CDMA
and OFDM.
IDENTIYING CLD MECHANISMS AND INTERFACES
Notification-Based CLD — With regard to the class of notifi-
cation-based CLD approaches, we have identified the need
for the following categories of mechanisms.
Congestion and Loss Recognition at the Link Layer — In
wired networks congestion is addressed through an adaptive
queue management (AQM) mechanism in network routers
that support RED algorithms based on the average queue
length exceeding a specific threshold [15]. However, in
CDMA wireless networks mechanisms such as RED cannot
be provided because the uplink does not provide any shared
buffer [42]. In this case the estimation of load is used as a
proxy measure of congestion. As the uplink in CDMA net-
works is interference-limited, this depends on the intracell
interference experienced by one mobile user due to the trans-
mission of other mobile users in the same cell. In the down-
link case the load estimation is estimated based on the power
transmitted from the BS since the downlink is power-limited.
On the other hand, ELN uses an agent running at the BS
that inspects the loss due to corruption in the wireless medi-
um [16]. More specifically, the agent retains a sequence block
of ACKs indicating the successful transfer of packets from the
sender to the receiver. By comparing the previous with the
newly arrived ACK value, a packet loss can be detected and
the ELN bit set accordingly.
Congestion and Loss Indication to the Network and
Transport Layer — As mentioned previously, notifications
are conveyed by IP header options or ICMP messages at the
network layer. For example, ECN uses a 2-bit-long ECN field
in the IP header. These fields are the ECN-capable transport
(ECT), which indicates the ECN capability of the end node,
and the congestion experienced (CE) field used by routers to
indicate the congestion on the end-to-end path [15]. ELN uses
in-band ICMP signaling at the IP layer [16]. Another explicit
notification, Explicit Bad Station Notification (EBSN), is
implemented as a type of ICMP message [43]. In this approach
the BS sends ICMP messages to the TCP sender to request
either the postponement of timeouts or retransmission of
packets [44].
On the other hand, to convey cross-layer information
between a pair of TCP endpoints, a TCP receiver adds the
ECN-Echo (ECE) flag in the TCP header to inform the TCP
sender that a CE packet has been received. The TCP sender
sets the congestion window reduced (CWR) flag in the TCP
header to inform the data receiver that the cwnd parameter
has been reduced [15]. Another approach to explicit notifica-
tions, Explicit Wireless Loss Notification (EWLN), uses the
sequence number of the ACK at the TCP header in order to
indicate the specific packet loss sent from the TCP receiver to
the sender over the wireless link in the EWLN [45]. Alterna-
tively, the TCP checksum is used to determine the ELN bit
[16].
■ Figure 3. The ways of exchanging and indicating cross-layer information.
Header
XML
descriptors
HTTP
Physical
Label
N
LID
N
In-band
message
In-band
message
In-band
message
PayloadHeader
CLI
(a) Header options (c) Local profiles
(d) Network service
Label
2
LID
2
LID
Label
1
LID
1
Payload
(b) Additional header
Header
CLI
Out-of-band
message
Payload
PayloadHeader
Data link
Network
Abstraction
engine
Client
Physical
Data link
Network
Abstraction
engine
Client
IEEE Communications Surveys & Tutorials • 1st Quarter 2008
80
TCP Sender Reactions upon Receiving Notifications —
In the ECN-capable TCP, cwnd reduction occurs upon
detection of the CE field. However, the TCP sender should
not reduce its cwnd value more than once per window of
data and must reset the retransmit timer when the cwnd is 1
MSS (maximum segment size). If the sending TCP continues
sending ECE messages while the cwnd is 1 MSS, the retrans-
mission throughput drops to one packet per round-trip time
(RTT) [15].
Contrary to ECN, an ELN-capable TCP sender is fully
aware of which segment is lost and retransmits it without
reducing cwnd or any kind of congestion control action. The
TCP sender identifies the lost segment by the corrupted
counter field received in the TCP option field in conjunction
with the checksum field (ELN bit) in the received ACK [16].
In the same way as [44, 45], an ICMP-RETRANSMIT mes-
sage and an EWLN bit request retransmission of the segment
by the TCP sender when the BS has failed to retransmit that
packet and when the receiver detects errors in a TCP seg-
ment, respectively. In EWLN, the ACK contains the sequence
number of the corrupted packet that has been lost.
It is known that, in addition to the retransmissions at the
transport layer, the BS can initiate retransmissions at the link
layer. In such a case the BS must notify the TCP sender,
which must undertake the following actions: hold the retrans-
mit timeout (RTO) of the corresponding packet the link layer
tries to retransmit, and update the source timer based on the
estimation of the RTT in order to prevent packet retransmis-
sions [43][44].
User-Driven Notification — In certain cases, users may
want to control the download rate on their devices. To this
end, a priority parameter indicates the user’s requirements in
terms of bandwidth for each running application. Based on
this parameter, the TCP receiver can adjust its window for
each application [7]. The calculated receiver window is passed
to the TCP sender through an ACK as the conventional TCP
does. Thus, for each application, the TCP sender will not
exceed a particular amount of sent data before it waits for a
new ACK.
Cross-Layer Optimization and Scheduling — We dis-
cussed notification-based mechanisms that improve TCP’s
performance by mitigating the undesired effects of TCP
congestion control. There are also CLD solutions that
improve TCP’s performance or offer better QoS to upper
protocol layers by means of cross-layer optimization and
scheduling mechanisms. In cross-layer optimization several
alternative combinations of protocol layer parameters can
be pursued [25]. To this end, such an optimization problem
can be formulated by combining several parameters from
different layers across the protocol stack. The aim of the
following paragraphs is to identify and/or clarify some com-
mon parameters to be combined in cross-layer optimization
procedures.
Error Correction at the Data Link Layer — References
[12, 46] propose a mechanism at the link layer dedicated to
correcting packet errors instead of retransmitting them. They
rely on adaptive FEC by which the sender is able to select the
appropriate FEC taking into account an estimated PER calcu-
lated at the link layer [12]. A link layer agent is placed at the
mobile node and the mobile host. At the mobile node the link
layer agent estimates the PER and RTT of the TCP session
[46].
Adaptive Modulation and Coding Scheme at the Physical
Layer — This approach deploys AMC at the physical layer to
enhance TCP’s throughput performance [26]. The AMC
scheme serves a finite-length queue of K packets at the data
link layer. The PER estimation at the link layer depends on
mode n of the AMC scheme and the received γ. The channel
fading is adjusted by the so-called Nakagami parameter m,
and the channel state transition probabilities are modeled by
a finite state Markov chain model. The latter is affected by
the mobility-induced Doppler spread. By changing the K, γ,
and m parameter values, TCP’s throughput is adjusted to
match the target PER.
Queuing Control at the Data Link Layer — In [21] TCP’s
throughput is enhanced by controlling the queue performance
and the packet loss event of TCP in the wireless link. In the
proposed cross-layer optimization concept, link layer efficien-
cy is determined by the number m of LL units to be transmit-
ted by multicarrier CDMA (MC-CDMA) in a slot and the
SINR Γ value for all m LL units. Taking into account the
received power for an LL unit of flow i in a slot, the amount
of resources in a slot is defined as a function of the resource
vector (m, Γ). In the link layer a RED-like buffer is used for
retransmission of LL units. By calculating the mean queuing
delay for TCP packets in the wireless link using the RED
queuing mechanism, the target TCP throughput can be real-
ized, keeping the resource vector (m, Γ) in optimal values.
Retransmissions at the Wireless Link — In [27] a CLD
approach that jointly combines AMC at the physical layer
with truncated ARQ at the data link layer enhances the aver-
age spectral efficiency in terms of transmitted bits per symbol.
The average RTT and packet length N
p
at the data link layer
manifest delay constraints. On the other hand, the probability
of packet loss after the maximum number of retransmissions
specify the PER upper bound. Given the PER upper bound,
the γ regions (i.e., the received SNR bounds) of the AMC
scheme in conjunction with the N
r
max
optimize system perfor-
mance in terms of transmitted bits per symbol. The channel
fading is adjusted by the Nakagami parameter m. On the
other hand, the number of N
r
max
can enhance or downgrade
the PER in these γ regions.
In the same context, the CLD solution in [28] adopts type-I
HARQ in order to minimize buffer size and augment spectral
efficiency. The spectral efficiency optimization results rely on
■ Table 3. Signaling aspects in cross-layer communication.
Features Header options Packet header Local profiles Network service
Signaling mechanism In-band In-band or Out-of-band
In-band (header options
indicate labels)
In-band or out-of-band (depends
on SNMP)
CLI transport
mechanism
No transport
(notification)
Through packet header
XML descriptors or packet
headers
XML descriptors (access network)
SNMP (core network)
Overhead type and
localization
Internal (low)
external (low)
Internal (low) external
(medium)
Throughout the session
set-up (deterministic)
Throughout the negotiation
phase (occasional)
IEEE Communications Surveys & Tutorials • 1st Quarter 2008
81
both N
r
max
and packet length L changes. Additionally, the
work presented in [29] optimizes the average spectral efficien-
cy using type-II HARQ with rate-compatible punctured con-
volutional (RCPC) codes. Changing the rates of RCPC codes,
the spectral efficiency is improved in the range of γ regions.
Moreover, changes in the packet length L improve the spec-
tral efficiency for a specific γ region.
Estimation and Control of Power at the Physical Layer —
Scheduling strategies and priorities in CDMA networks can
be relied on to provide the estimated required power for a
user in the next transmission frame [9]. Moreover, given the
received power strength inferred by fast power control infor-
mation in BSs and the imposed packet delay threshold, pre-
diction of the SINR parameter value for the next frame and,
consequently, the BER and PER, can be calculated [47].
Therefor, based on measured power from the power control
mechanism, predictions of the radio channel’s condition
determine the priorities of packet scheduling.
Video Reconstruction and Adaptation at the Application
Layer — An application-driven cross-layer optimization
approach could be relied on to measure the video quality per-
ceived by users. The PSNR is a quantitative parameter that
implies the reconstruction quality of an image pertinent to the
original image at the receiver.
For video transmission over wireless links, the video recon-
struction quality at the receiver is the sum of the source dis-
tortion D
s
and the expected loss distortion D
L
. The D
s
expresses the error-free image and should be sent along with
the video bitstream. On the other hand, the D
L
is related to
the packet loss rate caused during transmission [6]. The pack-
et loss rate estimated at the data link layer results from infor-
mation about the physical layer such as the modulation
scheme (binary phase shift keying [BPSK], quaternary PSK
[QPSK], etc.), channel coding, channel estimates (i.e., SNR),
and transmit power.
In the same context a PSNR value is expressed as a func-
tion of the video’s source and channel condition characteris-
tics depicting the capability of video transmission over wireless
links when a JSCPC is deployed [32]. The video source and
channel condition characteristics are exposed by the video
source coding rate R
s
and the power level in a CDMA net-
work expressed as the energy-per-bit-to-interference-density
ratio, respectively. The data rate at the application layer is
adapted by changing the quantization step size of the encoder.
Additionally, statistical parameters from the MAC layer (e.g.,
throughput, spectral efficiency) allow smoother video adapta-
tion at the application layer [33].
Multiuser Diversity at the Physical Layer — In [10] the
application’s video frames are encapsulated into LL time-
frames called batches. On each batch a weight is assigned that
indicates the video stream’s priority. The MS creates a trans-
mission queue for each batch of the video frame and assigns a
timer with a timeout value to each batch. The BS knows the
class number, remaining size, and timer value of each batch.
At the link layer, batches are adapted to channel variations
caused by channel fading. The multi-user diversity (the
good/bad threshold) determines the batch’s timer value and,
thus, the probability that an LL frame is kept idle (i.e., the
backoff probability). The faster the channel fades, the larger
the backoff probability and (intuitively) the timer value.
Co-Channel Interference Controller in Multicell Systems
— In OFDM systems a mobile user experiences interference
from BSs using the same subcarrier. Hence, an optimization
problem that should be solved in OFDM systems is the sub-
carrier allocation for each user. The work in [8] employs an
optimization algorithm that considers the maximum achiev-
able rate of the kth user on the jth subcarrier. In addition, a
resource controller at the MAC layer determines the proper
rate for each user based on channel quality feedback informa-
tion such as the average SINR or MCS level. The SINR
parameter is affected by the power received from the serving
BS as well as that from its neighbor counterpart.
Furthermore, in multicell systems users served by the same
subcarrier in adjacent cells (i.e., co-channel users) form a set.
Reference [48] deems a set of users feasible if the BER at all
receivers does not exceed a certain threshold. This feasibility
depends on several factors, like the BS-user link gains, the
modulation scheme that defines the SINR and BER capabili-
ties, and the transmitting power from each BS, which are cru-
cial for controlling the interference at the receivers. By
imposing transmit power constraints at each BS and control-
ling the co-channel interference using a centralized controller,
[48] achieves a large rate in each subcarier. One algorithm
determines the allocation of users to subcarriers in an OFDM
multicell environment considering both the interference
caused by BSs to a new user that enters the network and the
interference caused by the BS serving the new user to previ-
ous users that have already joined the network. Another algo-
rithm assumes a user as preferable when the increase in the
SINR of the other users in the same subcarrier is minimal. In
the last algorithm a subcarrier is allocated to the user with the
larger gain in the subcarrier; subsequently, the total power
budget is allocated to subcarriers according to water filling,
imposing a power constraint to approach the SINR for a set
of subcarriers and remove subcarriers when the constraint is
violated.
Multiuser Scheduler at the MAC Sublayer — In [30] a
CLD solution proposes a scheduling policy for a TDMA sys-
tem served by AMC and a finite-length queue of K packets
per user at the physical and data link layers, respectively. For
any particular user, the maximum number of packets that can
be transmitted at time t depends on the AMC mode n and the
number of time slots b reserved for one user. The key param-
eters of this CLD approach consist of the channel condition
parameters such as Doppler spread f
d
, SNR γ, and Nakagami
parameter m as well as the resource management parameters
such as time slots b, PER P
0
, and queue size K. The objective
of this CLD solution is the minimization of radio resources, b
and K, while guaranteeing the prescribed QoS.
For OFDM systems, [8] provides scheduling for each user
for different modes by exploiting information from the physical
layer (i.e., channel matrix, SINR, MCS level, velocity, and
location) as well as from the MAC layer (i.e., fairness and QoS
in terms of packet delay and packet loss rate). The scheduler
specifies the scheduling of users and the number of packets
that can be scheduled in the current frame. Variable channel-
aware scheduling is applied by using the SINR reported by the
mobile terminals on the uplink control channels.
THE COEXISTENCE OF CLD SOLUTIONS
For different CLD solutions to harmoniously coexist, the com-
monalities between them must be identified, including the
common sets of layer parameters. Although the complete
identification of CLD commonalities is beyond the scope of
this article, in the next paragraphs we discuss the most promi-
nent ones.
For instance, notification-based CLD solutions employ the
aforementioned mechanisms listed in Table 4. To identify
IEEE Communications Surveys & Tutorials • 1st Quarter 2008
82
congestion, a RED-like mechanism that relies on the average
queue length at the link layer should be supported. On the
other hand, to support notifications about packet losses, a link
layer agent that retains an ACK list is an appropriate solution.
At the network layer, ICMP messages and IP header options
represent the out-of band and in-band signaling, respectively.
At the transport layer, the TCP header options and other
fields (e.g., checksum) may be used to indicate network con-
gestion and packet loss. It is realized that upon receiving a
notification about congestion or loss of a segment, the TCP
sender will either reduce the cwnd or retransmit the segment,
respectively. To avoid unnecessary retransmissions performed
by the BS (local recovery) and the TCP sender as well, the lat-
ter should postpone or reset the RTO for the particular lost
packet. Besides, in application-driven notification, the receiver
window is calculated based on user preferences associated
with each user’s open sockets.
To summarize the discussion, a clearly open issue lies in
properly choosing how to identify and signal congestion from
the link layer in wireless networks. Physical layer characteris-
tics (interference level, transmitting power budget, etc.) in
infrastructure networks (i.e., CDMA) as well as in ad hoc net-
works (i.e., wireless LANs [WLANS]) can support this task, as
mentioned previously. However, further investigation is
required for an integrated end-to-end ECN solution in hetero-
geneous networks [42].
Apart from identifying some generic mechanisms used by
notification-based CLD approaches, research should also
investigate the coexistence of cross-layer optimization and
scheduling approaches. One coexistence aspect is the deploy-
ment of several algorithms/techniques that rely on the same
coupling between different layers [4]. For instance, solutions
that combine the AMC at the physical layer and an ARQ-like
protocol (e.g., type-I, type-II HARQ) at the data link layer
consider the same parameters, such as RTT, N
p
at the data
link layer, and the mode n and SNR γ at the physical layer
(Table 5). A cross-layer optimizer that uses these parameters
as input can activate different algorithms at different time
instants. This objective is undertaken by the co-
channel interference controller operating in a
multicell environment where the modulation
scheme, interference level (SINR), transmit
power, and BS-user link gain serve as input to
different subcarier allocation algorithms (Table
5). With regard to the power control estimation
at the physical layer, the received power strength
measurement (SINR) is a crucial parameter for
scheduling strategies at the MAC layer along
with cardinal parameters such as the time slots b
and queue size K at the data link layer (Table 5).
THE ROLE OF THE PHYSICAL LAYER
It is obvious that unique physical layer features,
such as AMC, power control, subcarrier alloca-
tion, and multiuser diversity, feature prominently
in CLD applications. These features depend on
parameters like the transmitted power from the
BS, the modulation and coding scheme, the BER
level induced by additive Gaussian noise and co-
channel, intracell, and intercarrier interferences
according to the deployed multicell network, the
BS-user link gain, as well as the velocity and loca-
tion of the MS.
Similarly, while the estimation of SNR con-
cerns the evaluation of modulated and coded
wireless channels, the SINR is a good metric for
evaluating channel strength in multiple access wireless systems
characterized by fading channels where intercell interference
must be added to channel noise. It has been argued [27, 28]
that in cross-layer optimization and scheduling the SINR
should be related to the PER in evaluating the effectiveness
of a CLD proposal. Such mappings are important in simulat-
ing and evaluating the performance of a CLD proposal at the
physical layer [8, 46].
However, SINR-PER mapping is not an easy task [46] as
the calculation of additive interference is affected by the actu-
al network topology [2], and bit errors are correlated [27].
Consequently, when topologies are different, the only parame-
ter required to evaluate a CLD is the SINR-PER mapping
given that the SINR is already well defined. Further investiga-
tion of this mapping is a key aspect of cross-layer optimization
and scheduling approaches for PHY-MAC co-design and
coordination.
On the other hand, spatial processing via multiple-input
multiple-output (MIMO) techniques is an advanced signal
processing technique that alleviates cross-layer scheduling at
the data link layer [49]. Reference [50] proposes a solution
based on MIMO and AMC at the physical layer to support
QoS guarantees at the data link layer in terms of effective
bandwidth and effective capacity. Such cross-layer modeling
that capitalizes on spatial diversity and spatial multiplexing of
MIMO systems is also an area for further investigation.
DISCUSSION AND LESSONS LEARNED
EVALUATION CRITERIA FOR CLD MODELS
Besides the need for cross-layer architecture, the relation
between such architecture and the resulting performance is
also important. Cross-layer design architectures can lead to
considerable improvements in throughput and delay perfor-
mance. To this end, the following evaluation criteria should
be considered [3]:
• Unintended interactions: A CLD architecture must figure
■ Table 4. Mechanisms and parameters involved in notification-based CLD.
Mechanism Parameters
Application layer User-driven notification Application priority
Transport layer
Congestion and loss indi-
cation
TCP header checksum
TCP header options
Cwnd reduction Congestion window
Retransmissions SN of corrupted packet
Timeout reset or pause RTT estimation
Receiver window control Receiver window
Network layer
Congestion and loss indi-
cation
ICMP message
IP header options
Data link layer
Congestion recognition RED (average queue length)
Loss recognition Agent (ACKs list block)
Physical layer Congestion recognition
Load estimation (intra-cell
interference, BS transmit
power)
IEEE Communications Surveys & Tutorials • 1st Quarter 2008
83
out the effect of interactions caused on a separate part
or layer of the protocol stack. Cross-layer architectures
usually do not take into account the impact of the inter-
actions they introduce between nonadjacent layers.
• Dependency: Many different but possibly interrelated
parameters are involved in CLD architectures. The rela-
tion between these parameters should be considered
using a dependency graph.
• Stability: When a parameter in the dependency graph is
controlled by two different loops implemented on the
same protocol, the stability of the entire scheme must be
carefully evaluated.
Furthermore, some kind of cost-benefit analysis for CLD
architectures that takes into account the following costs will
be necessary:
• The evaluation of an objective function with a large set of
variables may introduce excessive delays.
• The abstraction process for choosing the states and capa-
bilities of different layers also introduces some communi-
cation overhead.
• A cross-layer architecture that does not provide a good
level of modularity is more difficult to manage and opti-
mize.
LESSON LEARNED
Cross-Layer Architectures, Models, and Entities — Evi-
dently, the modularity of the overall architecture is a critical
issue in system design. An unwisely designed architecture will
suffer from excessive delay as a result of suboptimal interac-
tions between its modules. For example, [7] established a high
level of modularity by specifying control interfaces for each
module in the architecture; however, the additional functional
or procedure calls incur internal overhead in the protocol
stack. Cross-layer designers should take these unintended
interactions into account. To this end, designers should vali-
date their architecture to identify code deadlocks (i.e., parallel
control loops) that may seriously affect system stability. The
abstraction process may also introduce processing and com-
munication overheads [6]. An abstraction of a particular layer
can be represented by a small set of selected parameters that
clearly expose the associated layer’s capabilities. Obviously,
the overhead associated with the exchange and processing of
these parameters will be proportional to their cardinality [10].
In addition, the translation of abstracted parameters into
layer-specific parameters and actual modes of operation (and
vice versa) could introduce significant signaling overhead if
the set of selected parameters is excessively large. At protocol
layers residing below the network layer, the unnecessary trans-
mission of control messages due to the retransmission of
packets at the data link layer may also waste precious wireless
channel resources [8]. In such a case the data link layer’s effi-
ciency is degraded, and the achievable throughput drops sub-
stantially. Apparently, when a scheduler refrains from
collecting channel state information (i.e., no SNR feedback),
no external overhead results [9].
Cross-Layer Signaling — Thus far, we have elaborated on
cross-layer signaling mechanisms between the mobile clients
and the radio access network, and considered the way these
mechanisms are realized in an end-to-end networking context.
For such mechanisms, customarily one can adopt one of the
four available options (Fig. 3). It is evident that an important
aspect of CLD is a way to indicate cross-layer signaling
between peers, for which two primary alternatives are avail-
able: reusing a simple option (or options) in the current head-
er format of the employed transport protocol or using an
additional header. In the first case, the options in the current
header are considered mostly as an in-band signaling method
(e.g., ECN, ELN, EBSN). In the latter case, one must also
consider that an additional header will often require a sepa-
rate out-of-band signaling protocol (i.e., ICMP, RTCP,
RSVP). On the other hand, local profiles and a network ser-
vice convey valuable CLI for local or application layer adapta-
tion, respectively. Both these cases will undoubtedly introduce
external overhead when the session starts since all the LIDs
and XML descriptors must be transmitted for the first time.
However, once a session starts, the offered QoS may be
affected by the need to appropriately manipulate packet head-
ers and the size of header itself as well as the transformation
of the XML files. To summarize, the CLI can be conveyed by
either XML files or an additional packet header, and the way
of indicating CLI could be assigned to a header option (i.e.,
label, ECN).
Therefore, it is evident that the choice of cross-layer indi-
cations or the way of passing CLI may differ according to
application requirements. To this end, there is work in
progress within IETF that elaborates on the different types of
■ Table 5. Mechanisms and parameters involved CLD optimization and scheduling.
Mechanism Parameters
Application layer Video reconstruction and adaptation
Source distortion, expected loss distortion (packet loss rate P)
Video source-coding R
s
(quantization step-size)
Data link layer
Error correction Link layer agent (PER, RTT)
Queuing control with RED Length of K packets, queuing delay
Retransmissions
Number of Retransmissions N
r
max
, average RTT, packet length N
p
MAC layer scheduler Time slots b, queue of K packets per user, MCS level (SINR)
Physical layer
Adaptive modulation and coding
Mode n, SNR γ, Rate code
Power control Received power strength measurement (SINR)
Multiuser diversity Good/Bad threshold (channel fading)
Co-channel interference controller Subcarier rate (BS-user gain, transmit power, SINR)
IEEE Communications Surveys & Tutorials • 1st Quarter 2008
84
cross-layer signaling and notification [35]. In addition, [51]
argues for an in-band mechanism when signaling QoS require-
ments. In most cases, in-band signaling does not introduce a
new header but uses selected options in the current header
such as ECN, which is a well-known in-band solution [35].
Finally, [36] outlines several aspects on cross-layer communi-
cation to help CLD practitioners improve protocol perfor-
mance.
CONCLUDING REMARKS
As the evolution of wireless communication technologies con-
tinues, cross-layer approaches are being increasingly studied
as a design approach for confronting unintended performance
degradations as well as supporting new processes for beyond
3G mobile communication networks. Not surprisingly, over
the last decade cross-layer design has developed into a hot
research topic that is still growing. Currently, several research
efforts are seeking ways to integrate cross-layer design solu-
tions into wireless communication standards for purposes of
allocating resources to mobile users, scheduling access to
shared resources with higher throughput, and achieving better
quality of service for multimedia applications. Many of these
solutions employ functional entities that support cross-layer
processes for mobile devices. We envisage an evolution
toward 4G networks based on these kinds of functional enti-
ties, protocols, and processes that provide optimization and
reconfiguration at both the terminal and the network.
REFERENCES
[1] H. Zimmermann, “OSI Reference Model — The ISO Model of
Architecture for Open System Interconnection,” IEEE J. Trans.
Commun., vol. 28, no. 4, Apr. 1980.
[2] L. Georgiadis, M. J. Neely and L. Tassiulas, Resource Allocation
and Cross-Layer Control in Wireless Networks, Now Publishers
Inc., Apr. 24, 2006.
[3] V. Kawadia and P.R . Kumar, “A Cautionary Perspective on
Cross-Layer Design” IEEE Wireless Commun., vol. 12, no. 1,
Feb. 2005, pp. 3–11.
[4] V. Srivastava and M. Motani, “Cross-Layer Design: A Survey
and The Road Ahead,” IEEE Commun. Mag., vol. 43, no. 12,
Dec. 2005, pp. 112–19.
[5] G. Carneiro, J. Ruela, and M. Ricardo, “Cross-Layer Design in
4G Wireless Terminals,” IEEE Wireless Commun., vol. 11, no. 2,
Apr. 2004, pp. 7–13.
[6] S. Khan et al., “Application-Driven Cross-Layer Optimization
for Video Streaming Over Wireless Networks,” IEEE Commun.
Mag., vol. 44, no. 1, 2006, pp. 122–30.
[7] V.T. Raisinghani and S. Iyer, “Cross-Layer Feedback Architec-
ture for Mobile Device Protocol Stacks,” IEEE Commun. Mag.,
special issue on Cross-Layer Protocol Engineering, Jan 2006.
[8] T. Kwon et al., “Design and Implementation of a Simulator
Based on a Cross-layer Protocol between MAC and PHY Layers
in a WiBro Compatible IEEE 802.16e OFDMA System,” IEEE
Commun. Mag., Dec. 2005, pp. 136–46.
[9] R. Ferrús et al., “Cross-Layer Scheduling Strategy for UMTS
Downlink Enhancement,” IEEE Commun. Mag., vol. 43, no. 6,
June 2005, pp. s24–28.
[10] H. Jiang, W. Zhuang, and X. Shen, “Cross-Layer Design for
Resource Allocation in 3G Wireless Networks and Beyond,”
IEEE Commun. Mag., Dec. 2005.
[11] X. Lin, N. B. Shroff, and R. Srikant, “A Tutorial on Cross-Layer
Optimization in Wireless Networks,” IEEE JSAC, Special Issue
on Non-Linear Optimization of Communication Systems, 2006.
[12] M. Miyoshi, M. Sugano, and M. Murata, “Improving TCP Per-
formance for Wireless Cellular Networks by Adaptive FEC Com-
bined with Explicit Loss Notification,” IEICE Trans. Commun.,
vol. E85-B, no. 10, Oct. 2002.
[13] S. Shakkottai, T.S. Rappaport, and P.C. Karlsson, “Cross-Layer
Design for Wireless Networks,” IEEE Commun. Mag., vol. 41,
no. 10, Oct. 2003, pp. 74–80.
[14] H. Zheng and H. Viswanathan, “Optimizing TCP Performance
with Hybrid ARQ and Scheduler,” Proc. 14th Int’l. Symp. Pers.,
Indoor, Mobile Radio Commun., vol. 2, Sep., 2003, pp. 1785–89.
[15] K. Ramakrishnan, S. Floyd, and D. Black “The Addition of
Explicit Congestion Notification (ECN) to IP,” RFC 3168, Sept.
2001.
[16] H. Balakrishnan et al., “A Comparison of Mechanisms for
Improving TCP Performance over Wireless Links,” IEEE ACM
Trans. Networking, Dec. 1997.
[17] D. Kliazovich and F. Graneill, “A Cross-Layer Scheme for TCP
Performance Improvement in Wireless LANs,” IEEE GLOBECOM,
vol. 2, no. 2, Dec. 2004, pp. 840–44.
[18] T. Goff et al., “Freeze-TCP: A True End-To-End TCP Enhance-
ment Mechanism For Mobile Environments,” Proc. 19th INFO-
COM, Sept. 2000, pp. 1537–45.
[19] H. Balakrishnan and R. H. Katz, “Explicit Loss Notification and
Wireless Web Performance,” Proc. GLOBECOM Internet Mini-
Conf., Nov. 1998.
[20] P. Bender et al., QUALCOMM, Inc., “CDMA/HDR: A Band-
width Efficient High-Speed Wireless Data Service for Nomadic
Users,” IEEE Commun. Mag., July 2000.
[21] H. Jiang and W. Zhuang, “Cross-Layer Resource Allocation for
Integrated Voice/Data Traffic in Wireless Cellular Networks” IEEE
Trans. Wireless Commun., vol. 5, no. 2, Feb. 2006, pp. 457–68.
[22] Goldsmith and S. G. Chua, “Adaptive Coded Modulation for
Fading Channels,” IEEE Trans. Commun., vol. 46, no. 5, May
1998, pp. 595–602.
[23] IEEE 802.16 Working Grou
p, “IEEE Standard for Local and
Metropolitan Area Networks Part 16: Air Interface for Fixed
Broadband Wireless Access Systems,” 2004.
[24] 3GPP TS 25.308 V6.4.0, “High Speed Downlink Packet Access
(HSDPA); Overall Description; Stage 2,” rel. 6, 2007.
[25] M. van der Schaar and S. Shankar, “Cross-Layer Wireless Mul-
timedia Transmission: Challenges, Principles, and New
Paradigms,” IEEE Wireless Commun., vol. 12, no. 4, Aug. 2005,
pp. 50–58.
[26] Q. Liu, S. Zhou, and G. B. Giannakis, “TCP Performance in
Wireless Access With Adaptive Modulation and Coding,” Proc.
IEEE ICC, June 20–24, 2004.
[27] Q. Liu, S. Zhou, and G.B. Giannakis, “Cross-Layer Combining
of Adaptive Modulation and Coding With Truncated ARQ Over
Wireless Links,” IEEE Trans. Wireless Commun., vol. 3, no. 5,
Sept. 2004.
[28] D. Wu and S. Ci, “Cross-Layer Design for Combining Adaptive
Modulation and Coding with Hybrid ARQ,” Proc. Int’l. Conf.
Commun. and Mobile Comp., July 3–6, 2006, pp. 147–52.
[29] D. Wu and S. Ci , “Cross-Layer Combination of Hybrid ARQ
and Adaptive Modulation and Coding for QoS Provisioning in
Wireless Data Networks,” Proc. 3rd Int’l Conf. Quality of Ser-
vice in Heterogeneous Wired/Wireless Networks, Aug. 7–9,
2006, no. 47.
[30] Q. Liu, S. Zhou, and G.B. Giannakis, “Cross-Layer Scheduling
with Prescribed QoS Guarantees in Adaptive Wireless Net-
works,” IEEE JSAC, vol. 23, no. 5, May 2005.
[31] S. Mérigeault and C.Lamy, “ Concepts for Exchanging Extra Infor-
mation Between Protocol Layers Transparently for the Standard
Protocol Stack,” Proc. IEEE ICT ’03, Feb. 23-Mar. 1, 2003.
[32] Y. S. Chan and J. W. Modestino, “A Joint Source Coding-
Power Control Approach for Video Transmission over CDMA
Networks,” IEEE JSAC, vol. 21, no. 10, Dec. 2003.
[33] I. Haratcherev et al., “Optimized Video-Streaming Over
802.11 by Cross-Layer Signaling,” IEEE Commun. Mag. Special
Issue on Cross-layer Protocol Engineering for Wireless Mobile
Networks, vol. 44, no. 1, Jan. 2006.
[34] Q. Wang and M.A. Abu-Rgheff, “Cross-Layer Signalling for Next-
Generation Wireless Systems,” IEEE Wireless Commun. and Net-
working Conf., vol. 2, no. 16–20 Mar. 2003, pp. 1084–89.
[35] P. Sarolahti and S. Floyd, “Cross-Layer Indications for Transport
Protocols,” Internet draft, IETF, Mar. 2007. work in progress.
[36] P. Sarolahti, S. Floyd, and M. Kojo, “Transport-Layer Consid-
erations for Explicit Cross-Layer Indications,” Internet draft,
IETF, Oct. 2006, informational.
[37] P. Sudame and B. R. Badrinath, “On Providing Support for
IEEE Communications Surveys & Tutorials • 1st Quarter 2008
85
Protocol Adaptation in Mobile Wireless Networks,” Mobile
Networks and Apps., vol. 6, no. 1, Jan.–Feb. 2001, pp. 43–55.
[38] A. Takács et al., “The Layer-Independent Descriptor Concept,”
J. Computers, vol. 1, no. 2, May 2006, pp. 23–32.
[39] B. Kim, “A Network Service Providing Wireless Channel Infor-
mation for Adaptive Mobile Applications: Part I: Proposal,”
Proc. IEEE ICC ’01, June 2001.
[40] J. Case et al., “A Simple Network Management Protocol,”
RFC 1157, May 1, 1990.
[41] L. Larzon, U. Bodin, and O. Schelen, “Hints and Notifications,”
Proc. Wireless Commun. and Networking Conf., Mar. 2002.
[42] V. A. Siris, “Resource Control for Elastic Traffic in CDMA Net-
works,” Proc. ACM MOBICOM, Sept. 2002.
[43] B. S. Bakshi et al., “Improving Performance of TCP over Wire-
less Networks,” Proc. IEEE ICDCS ’97, 365–73.
[44] S. Goel and D. Sanghi, “Improving TCP Performance over
Wireless Links,” Proc. TENCON ’98, Dec. 1998, pp. 332–35.
[45] B. Zhang and M.N. Shirazi, “Implementation of Explicit Wire-
less Loss Notification Using MAC-Layer Information,” Proc. IEEE
Wireless Commun. and Networking Conf., vol. 2, Mar. 2003,
pp.1339–43.
[46] B. Liu, D. L. Goeckel, and D. Towsley. TCP-Cognizant Adaptive
Forward Error Correction in Wireless Networks,” Proc. IEEE
GLOBECOM, Nov. 2002.
[47] M. Saied and A. El-Atty, “Efficient Packet Scheduling with
Pre-Defined QoS Using Cross-Layer Technique in Wireless Net-
works,” Proc. IEEE ISCCC ’06.
[48] I. Koutsopoulos and L. Tassiulas, “Cross-Layer Adaptive Tech-
niques for Throughput Enhancement in Wireless OFDM-Based
Networks,” IEEE/ACM Trans. Net., vol. 14, no. 5, Oct. 2006.
[49] C. Antón-Haro et al., “Cross-Layer Scheduling for Multi-user
MIMO Systems,” IEEE Commun. Mag., vol. 44, no. 9, Sept.
2006, pp. 39–45.
[50] J. Tang and X. Zhang, “Cross-Layer-Based Modeling for Quali-
ty of Service Guarantees in Mobile Wireless Networks,” IEEE
Commun. Mag., vol. 44, no. 1, Jan. 2006, pp. 100–06.
[51] J. Harper and P. McGeer, “Requirements for In-Band QoS Sig-
nalling,” Internet draft, Network Working Group, Jan. 2007,
work in progress.
BIOGRAPHIES
FOTIS FOUKALAS (foukalas@di.uoa.gr) received a Dipl.-Ing. degree in
electrical and computer engineering from Aristotle University of
Thessalonica, Greece, in 2001. In 2003 he received an M.Sc.
degree from the National Technical University of Athens. In May
2003 he joined the Communication Networks Laboratory at the
Department of Informatics and Telecommunications of the Univer-
sity of Athens. Since December 2003 he has been working on his
Ph.D. in the same department. His research interests are in the
area of physical and data link layer techniques enabling cross-
layer optimization and reconfigurability.
VANGELIS GAZIS (gazis@di.uoa.gr) holds B.Sc., M.Sc., and Ph.D.
degrees from the Department of Informatics and Telecommunica-
tions at the University of Athens, and an M.B.A. degree from the
Athens University of Economics and Business). He is a senior
researcher at the Department of Informatics and Telecommunica-
tions ,and has participated in national (OTE-DECT, GUNet, GEANT-
2) and European projects (MOBIVAS, ANWIRE). He specializes in
reconfigurable mobile systems and protocol stacks for beyond 3G
mobile, ontology languages for autonomic systems, reflective and
component middleware, adaptable services, and open API frame-
works for telecommunications.
NANCY ALONISTIOTI (nancy@di.uoa.gr) has B.Sc. and Ph.D. degrees
in informatics and telecommunications from the University of
Athens. She specializes in reconfigurable systems and networks
for beyond 3G, adaptable services, pervasive computing, and con-
text awareness. She has participated in national and European
projects, (CTS, SS#7, ACTS RAINBOW, EURESCOM, MOBIVAS,
ANWIRE, E2R, LIAISON) and is co-editor of Software Defined
Radio, Architectures, Systems and Functions (Wiley Series on Soft-
ware Radio).
