Although cross-layer has been thought as one of the most effective and efficient ways for multimedia communications over wireless networks and a plethora of research has been done in this area, there is still lacking of a rigorous mathematical model to gain in-depth understanding of cross-layer design tradeoffs, spanning from application layer to physical layer. As a result, many existing cross-layer designs enhance the performance of certain layers at the price of either introducing side effects to the overall system performance or violating the syntax and semantics of the layered network architecture. Therefore, lacking of a rigorous theoretical study makes existing cross-layer designs rely on heuristic approaches which are unable to guarantee sound results efficiently and consistently. In this paper, we attempt to fill this gap and develop a new methodological foundation for cross-layer design in wireless multimedia communications. We first introduce a delay-distortion-driven cross-layer optimization framework which can be solved as a large-scale dynamic programming problem. Then, we present new approximate dynamic programming based on significance measure and sensitivity analysis for high-dimensional nonlinear cross-layer optimization in support of real-time multimedia applications. The major contribution of this paper is to present the first rigorous theoretical modeling for integrated cross-layer control and optimization in wireless multimedia communications, providing design insights into multimedia communications over current wireless networks and throwing light on design optimization of the next-generation wireless multimedia systems and networks.
In recent years, ubiquitous computing devices such as laptop computers, PDAs, smart phones, automotive computing devices, and wearable computers have been ever growing in popularity and capability, and people have begun more heavily to rely on these ubiquitous computing devices. Therefore, there has been a strong user demand for bringing multimedia streaming to the devices such as iTunes, PPLive, MSN, and YouTube. However, bringing delay-sensitive and loss-tolerant multimedia services based on the current wireless Internet is a very challenging task due to the fact that the original design goal of the Internet is to offer simple delay-insensitive loss-sensitive data services with little QoS consideration. Therefore, this shift of design goal urges us to rethink the current Internet architecture and develop a new design methodology for multimedia communications over the current and future wireless Internet. So far, cross-layer design has been thought as one of the most effective and efficient ways to provide quality of service (QoS) over wireless networks, and it has been receiving many research efforts. The basic idea of cross-layer design is to fully utilize the interactions among design variables (system parameters) residing in different network functional entities (network layers) to achieve the optimal design performance of time-varying wireless networks.
In order to achieve the global optimality of cross-layer design, we need to consider design variables and the interactions among them as much as possible. However, more does not necessarily mean better. The more design variables we consider, the more difficult is orchestrating a large number of design variables to make them work harmonically and synergetically. From the point of view of nonlinear optimization, the number of design variables increases and the size of state space of the objective function will increase exponentially, making the optimization problem unmanageable. To overcome this problem, one often used approach is to reduce the size of the problem at the system modeling phase and then solve the simplified problem by using various optimization algorithms such as gradient-based local search, linear/nonlinear programming, genetic algorithm, exhaustive search, and heuristic-based approach like artificial neural networks.
However, reducing a high-dimensional cross-layer
optimization problem to a low-dimensional problem in the system modeling phase raises a series of questions: how to evaluate
the fidelity of the simplified problem compared with the problem as what it
should be, how to evaluate
the quality of the suboptimal solution to the global optimum, how to evaluate
the robustness of the solution, that is, whether the solution can guarantee the
predictable sound results at all possible circumstances.
Unfortunately,
at the time of this writing, we have no clear answers to all these three questions.
Moreover, reducing the size of the problem in the problem formulation means that only part of the current Internet architecture can be considered, causing a shift of the design goal of multimedia services from the best user experience to some layer-specific performance metrics such as distortion at the application layer, delay at the network layer, and goodput at the MAC/PHY layer. This shift of design goal may cause an “Ellsberg paradox,” where each individual design variable makes good decisions for maximizing the objective function. But the overall outcome violates the expected utility function. In other words, breaking a big problem into several smaller problems in the system modeling phase can only increase the solvability of the original problem but cannot guarantee that it is a good solution. The “Ellsberg paradox” also tells us that the traditional additive measure such as probability measure may no longer hold in the context of cross-layer design due to the possible strong coupling (interdependency) among design variables. At the point of this writing, there have been many researches done on interdependency modeling in the context of cross-layer design, but they are mostly qualitative rather than quantitative approaches, and their applications are still within the scope of local cross-layer optimization.
We argue that all aforementioned difficulties in the area of cross-layer design of wireless multimedia communications are due to lacking of methodological foundation and in-depth understanding of cross-layer behavior. Our goal is to provide a flexible yet scalable theoretical cross-layer framework to accommodate all major design variables of interest, spanning from application layer to physical layer, for delay-bounded multimedia communications over wireless single/multihop networks. We start from proposing an integrated cross-layer framework for the best user experience. Although the engineering side of cross-layer design is not the main focus of this paper, we still briefly discuss how to utilize the methodological foundation to achieve real-time multimedia communications through a fast algorithm for large-scale global cross-layer optimization based on quantitative significance measure and sensitivity analysis.
The rest of the paper is organized as follows. We
briefly introduce the related work in Section
In literature,
topics involving video delivery over multihop
networks such as video coding, multihop routing, QoS provisioning, link
adaptation are separately studied. Therefore, the corresponding video
compression efficiency and the transmission efficiency are also separately
optimized. In prediction mode, selection of video coding, periodic intracoding
of whole frames [
In routing for video delivery in multihop networks, an
application-centric cross-layer approach has been proposed to formulate an
optimal routing problem for multiple description video communications [
Cross-layer optimized wireless video has been studied
from different aspects, such as cross-layer architecture [
In wireless video, optimization has to be done over
multiple source coding units, such as frames and pixel blocks, for the best
reconstructed video quality. There is “
Existing ADP approaches have largely ignored the
interdependencies among control variables, which might lead to
In the protocol stack of multimedia over wireless networks, each layer has one or multiple key system parameters which would significantly impact the overall system performance. At the application layer, tradeoff between rate and distortion is an inherent feature of every lossy compression scheme for video source coding. Prediction mode and quantization level are two critical parameters. At the network layer, routing algorithm is important to find the best delivery path over a single/multihop wireless network. At the data link layer, hybrid automatic repeat request (HARQ), media access control protocols, and packetization are often used to maintain a low packet loss rate. However, the choice of maximum retransmission number is a tradeoff between resultant packet delay and packet loss rate. Note that for real-time multimedia applications, we might not consider HARQ due to strict delay constraints. At the physical layer, adaptive modulation and coding scheme is an important tradeoff between transmission rate and packet loss rate. Furthermore, the end-to-end performance is not completely determined by the parameters of individual layer, but rather by all parameters of all layers. For example, the end-to-end delay consists of propagation delay (determined by the number of hops of the selected path), transmission delay (determined by channel conditions, modulation and channel coding, maximum retransmission number, and source rate), and queueing delay (determined by source rate, transmission rate, and the selected path). Moreover, due to the time-varying nature of wireless channels, each node in the network should be capable of adjusting these parameters quickly to maintain a good instantaneous performance. Clearly, the layer-separated design no longer guarantees an optimal end-to-end performance for multimedia delivery over wireless networks.
We develop a cross-layer framework to optimize
multimedia communications over single/multihop wireless networks. In order to
demonstrate the main idea of the proposed framework as shown in Figure
(a) Performance comparison using sample video clip: global cross-layer optimization
versus existing piecemeal cross-layer optimization. Here, assume that multihop
paths and their link quality can be found by a multihop routing protocol, such
as optimized link state routing protocol (OLSR) [
Let us denote by
We assume that the considered multihop network
consists of
Note that in this work,
Thus, the proposed cross-layer framework for wireless
multimedia communications can be formulated as
Recall that the focus of the proposed framework is to
jointly find the optimal parameter set for each frame
Clearly, in (
Note that the unique feature of (
For the global optimality of system performance, we
need to optimize current control action
So far, there is only one exact method for global
optimization over time with nonlinearities and random disturbances [
Then, the
global optimization problem turns into calculating the cost-to-go function
We have evaluated the performance of the proposed integrated cross-layer framework through extensive simulations based on H.264 JM12.2 codec. In general, we are interested in comparing our integrated cross-layer design with the best possible results of H.264 codec. Our goal here is to illustrate the difference of performance gain between the global optimality achieved by the proposed framework and the superposition of multiple local optimality done separately at different network layer (s). In this paper, the best baseline performance is derived: (1) at the application layer, it uses the rate control scheme of H.264 codec; (2) at the network layer, it always chooses the path with the best average SNR at each hop; (3) at the MAC and PHY layers, it always chooses the AMC scheme for the shortest delay while keeping the predefined PER performance.
From the simulation results, up to 3 dB PSNR gain can
be achieved by using the proposed approach compared with using the existing
piecemeal approach, as shown in Figure
An integrated cross-layer framework of multimedia communications over multihop wireless networks.
An integrated cross-layer framework of multimedia communications over multihop wireless networks.
We have proposed a top-down theoretical cross-layer framework for multimedia over wireless networks, and the correctness of the proposed methodology is based on its rigorous theoretical foundation. Moreover, the proposed methodology is based on dynamic programming, which means that it is very flexible and scalable; any interaction of interest in the system can be easily integrated into the proposed framework. Since we consider all the major interactions of interest spanning from application layer to physical layer, we have overcome the major drawback of existing cross-layer designs where the simplification occurs at the system modeling phase rather than the problem solving phase. Therefore, the proposed methodology provides the true global optimality and a new design guidance to the cross-layer design for multimedia over various wireless networks.
In this
section, we will further discuss how to apply the aforementioned global
optimization framework for real-time multimedia communications as formulated in
(
So far, we have
presented a new theoretical framework for cross-layer design of multimedia
communications over wireless networks, which provides a sound methodological
foundation for us to evaluate cross-layer designs using dynamic programming
(DP) which has been widely adopted to study sequential decision-making problems
(stochastic control). However, the practical applications of dynamic
programming are limited mostly due to the dual
curses of dimensionality and uncertainty, that is,
the large size of underlying state space of the
The most sensible and rational way to deal with the difficulty caused by “dual curses” is to generate a compact parametric representation (compact representation, for brevity) to approximate the cost-to-go function for a significant complexity reduction through mapping the huge state space to a much smaller feature space characterized by a compact representation.
Currently, the selection of a compact representation largely relies on heuristics which somewhat contradicts the nonheuristic aspects of the dynamic programming methodology. Therefore, we propose a new method based on nonadditive measure theory, which can dynamically generate compact representations of the huge state space. Unlike other nonlinear feature-extraction approaches such as artificial neural network, the proposed method is adaptive and nonheuristic in the sense that it allows us to quantitatively characterize the significance or the desirability of state vectors with considerations of interactions among different state variables. Therefore, new feature-based approximate dynamic programming can be developed based on the adaptive feature extraction and compact representation.
We consider a large-scale dynamic programming problem defined
on a finite state space
In the context
of dynamic programming, the cost-to-go vector
In large-scale dynamic programming problems, the size
of state space normally increases exponentially with the number of state
variables, making it extremely difficult to compute and store each component of
the cost-to-go function. Therefore, the most sensible way is to map a huge
state space
In the context of approximate dynamic programming, we
would like to see that when
Formally, a feature
For approximate dynamic programming using
feature-based compact representation, the approximate cost-to-go function is
Feature
extraction requires us to catch the “dominant nonlinearities" in the
optimal cost-to-go function
In our preliminary study [
Once, we determine the
significance measure of state variables
As discussed
earlier, based on the significance measure and sensitivity analysis, we can
derive a new method for feature extraction and compact representation for
approximating the original large-scale dynamic programming. Using the same
problem setting as of Figure
In this simulation, based on the significance measure,
the interaction between QP and AMC has the most significant impact on the
cost-to-go function, meaning that “path" is not as significant as the
other variables. So, it could be excluded from the
optimal search. This way, the cardinality of the approximated state space can
be
In this section, we propose a new method for feature extraction and compact representation of approximate dynamic programming, which is based on the significance measure of each set of design variables. We discuss a novel feature-based approximate dynamic programming approach for solving the large-scale dynamic programming problem in support of real-time multimedia applications. Furthermore, since all the significant measures of a power set of design variables are available, a scalable complexity framework by exploring the tradeoff between the quality of approximation (QoA) and the quality of service (QoS) could be developed in future. Note that the proposed significance measure method and the feature-based approximate dynamic programming approach are fairly generic and are applicable for any large-scale design optimization and real-time control scenarios.
The major
challenges of current cross-layer design for multimedia communications over
wireless networks are (1) lacking of understanding of cross-layer behaviors,
(2) simplifying cross-layer design at the system modeling phase, and (3)
relying on heuristic approaches. We argue that all these challenges are caused
by lacking of a new methodology for cross-layer design of multimedia
communications over wireless networks. This has motivated us to propose a new
methodological foundation for cross-layer design of multimedia communications
over wireless networks, which has made two major contributions to the research
area: (1) the theoretical framework with major design variables spanning from
application layer to physical layer for cross-layer design of multimedia
communications over wireless networks, and (2) the novel feature-based
approximate dynamic programming approach based on a new significance measure
method to understand cross-layer behaviors and speed up large-scale cross-layer
optimization. The proposed methodological foundation is fairly general and can
be applicable to other applications in multimedia communications. However, we
are
Perceptual
video quality comparison based on H.264 codec in a real-time environment
u[ | QP | 179.23 | T | Global | Baseline | ||||
---|---|---|---|---|---|---|---|---|---|
u[ | Path | 31.95 | 0.01312 | 146433 | 164824 | 156336 | 149483 | 194813 | |
u[ | QP,Path | 93.31 | 0.02217 | 88552 | 98207 | 97470 | 95454 | 123607 | |
u[ | AMC | 23.18 | 0.03454 | 56983 | 67341 | 64555 | 60364 | 78341 | |
0.03960 | 49719 | 54891 | 51555 | 50201 | 63891 | ||||
u[ | Path,AMC | 3.22 | 0.06105 | 33702 | 36535 | 35623 | 34564 | 42535 | |
u[ | QP,Path,AMC | 157.31 | 0.09725 | 21573 | 23551 | 23337 | 22811 | 27551 |