2010/12/17 14:18
Title: Cross-Layer Design Considerations
Jiwoong Lee, University of California, Berkeley
ABSTRACT
INTRODUCTION
The prosperity of the Internet successfully produced the wide deployment of the network infrastructure, enormous number of the network users, and still promising vision on the network industry. In addition to this, the desire of cordless communication has recently been propelled so that there exist a huge trend of technical and academical challenge in wireless communications. Meanwhile the networking protocols and systems have been optimized for conventional wired networks, the wireless medium shows somewhat fundamentally different characteristics so that the legacy protocols and systems do not perform in the optimized sense.
RELATED WORK
WHAT IS ARCHITECTURE
A better understanding of the notion of architecture is required to have a firm base to discuss cross-layer design. Three basic model of architecture can be discussed. The first model is called as brick model. Brick model is used most commonly and most robust in a simple design. Next model is called spagetti model. Spagetti model does not have a layered architecture at all inside. Rather, its functionalty is totally mixed. Its implementation can be really messy. However if carefully designed and implemented, this model can have the optimal performance. The final model, which is a graphical understanding of the cross-layer design, is the jigsaw model. Jigsaw model is placed in the middle of brick model and spagetti model. It has dividable components but their interaction is much more involved than brick model, less than the spagetti model. From this graphical representation, we can acquire two important perspectives of the cross-layer design. First perspective is that the adjacent layers have much larger interface area between them than the brick model that has flat interface area. As will be discussed later in this paper, this expanded area implies the change of abstraction model of each layer. In other words, a layer 'understands' its adjacent layers with more verbose parameters. The second interesting perspective on this is the conglomeration of the functionalities which were believed to belong to different layers. This conglomeration of functionalities have not been recognized widespread. However, this must be the essencial understanding as well as the expanded interface area. One good example of this conglomeration is the merged MAC forwarding-Network routing.
There was a research that emphasized the importance of architecture; the architectured design was believed the key to make it survive and thrive. In [Kumar2005] four examples were introduced as successful architectures. They were Von Neumann architecture which initiated the modern computer architecutre and has been unchanged, the OSI network architecture which has been the fundamentals to understand complex network systems although its implementation has never been used, the Source-Channel separation architecture which was introduced by Shannon and is still regarded as one of the most excellent achievement in communication theory, and finally the feedback control system which is necessary in any form of control system. There are no argue point in that those four examples are fundamentally important and have been rigid bases. However one surprising thing is that except feedback control system, other three architectures are not relatively old enough. Von Neumann architecture was introduced in 1963, OSI 7 Layer architecture was finalized as an international standard in 1988, and Source-Channel separation was pulicized in 1949. Therefore we have to acquire a sense that what we have been believing as an unbreakable fundamental architecture were actually young enough to be challenged by a new architecture.
WHAT IS LAYERING
Before introducing a new challenger, we might as well contemplate the concept of layering design. Layering can be understood in following ways. They are essentially the same.
- Layering is isolating conceptually different functions
- Layering is decoupling the complicated problem into smaller ones
- Layering is decomposing the problem space into subspaces
The advantage of layering is it enables modular design. Optimization of the overall complicated system is reduced to the optimization of each simple module. Modular design enables the reusability of the design, implementation, and operation. It even enables the modular understanding of the system.
As a result of layering, abstraction of layers are produced. Interface of the layer handles this abstaction to understand, communicate with the adjacent layers.
Layering is the product of 'divide-and-conquer' philosophy. This was originally the military strategy. What this strategy claims is that when one fight with 10 enemies, it is easier to fight with one by one rather that fight with ten at the same time. This strategy was later incorporated into the algorithm design paradigm. As one can easily imagine, this is a useful and powerful strategy. Now I have a question for you. Does this strategy always work ? The correct answer is no. This strategy works only when the cost of combining is less than the cost of dividing and solving. In other words, if the tackling the whole problem is more efficient than dividing the problem space and solving each divided subspace, then layering approach might not work. This is the philosophical basement of the cross-layer design.
GOAL OF CROSS-LAYER DESIGN
In contrast to the approach of layering design, the aim of the cross-layer design is to couple some functionalities of heterogeneous layers to boos the systm-wide performance by taking advantage of joint optimization. An interpretation of the trend of the cross-layer design is that we have faced the limit of the divide-and-conquier strategy at the emergence of wireless communication, in the sense of the legacy layering model. Therefore to increase the system-wide performance, we need the expanded abstaction of each layer in the interface or more organic model in which every small components can work with others.
A recent research [Iyer2003] described the possible cross-layer interactions through the expanded layer abstaction. For example, transmission power has been conventionally the internal parameter of the physical layer and has not been shared with other layers. In a cross-layer design, however, the transmission power can be in higher layer; network layer can determine its route based on its current trasmission power. Another example is that physical layer uses the number of the retransmission performed in MAC layer in change its modulation scheme which provides more robust detection at the cost of power consumption.
TENSION
It is true that there is a tension in diving into the cross-layer design since it requires, by definition, the break of the conventionally defined layered architecture. Breaking this boundary can break the modularity, compatibility or guaranteed minimum performance as well.
COUNTER-EXAMPLE
This tension is worthy of a few more words since the expected worries are not a total illusion. A good example was analyzed in [another Kumar2005]. When a rate adaptive MAC, wich adjusts its data rates based on the channel quality information delivered by the physical layer, is used with the network layer's minimum-hop routing, the overall throughput of the system was degraded up to 55% in comparison to the case that does not use the rate-adaptive MAC. A rate adaptive MAC and minimum hop routing were developed independently but when they were used cooperatively, the result of the cooperation was worse. This is a good example showing that cross-layer performance optimization requires cross-layer engineering.
LIMIT
Therefore we had better point out the limit of the cross-layer design not to repeat the same cross-purpose optimization. The biggest limit of the cross-layer design is this design paradigm is quite new so that the consequence of the cross-layer design is not easily imagined like a new born baby. When a new design is combined with other layers' protocols, traffic pattern, user behavior, unexpected conseequences will be brought out, while it can be good or bad. The second limit of the cross-layer design is the lack of the compatibility with the legacy protocol stack. The third limit is the possible lack of modularity. The fourth limit is its longevity. The final limit is essentially not a limit; it is a possibility. A recent research showed the performance improvement of the cross-layer design is constant times of that of the layered-design
CONCLUSION
