Which Architectures for the Cognitive Networks of the Future?

Download as PDF

I. INTRODUCTION

The evolution of communication technologies, especially in the wireless domain, introduced a paradigm shift from static to mobile access, from centralized to distributed infrastructure, and from passive to active networking. Technological advances have brought networking a step forward towards the goal of service provision on an "anytime, anywhere" basis, while ensuring instantaneous and secure communications. However, such innovation is bound by the constraints included in the original Internet (and TCP/IP) design. The fundamental reason for performance inefficiency is the difficulty in configuring and managing networks – a task traditionally performed by network operators and technicians [1].

Self-awareness, self-management, and self-healing characteristics have been proposed in order to optimize network operation, reconfiguration, and management, as well as to improve data transfer performance by bringing “intelligence” into the network, thereby creating a new paradigm known as cognitive networking, which is expected to become a key part of 4th generation wireless networks (4G) [2].

The term cognitive is related to the ability of a network to be aware of its operational status and adjust its operational parameters to fulfill specific tasks, such as detecting changes in the environment and user requirements. Cognition requires support from network elements (routers, switches, base stations, etc.), which should host active tasks to perform measurements to reconfigure the network. These characteristics are related to the paradigm of active networks [3], which differ from cognitive networks service in that they do not include a cognitive process that implements adaptation and learning techniques.

The ability of cognitive network to think, to learn and benefit from past experience requires communication between cognitive elements. Cognitive network implementation can be highly distributed or tend towards centralized solutions. Common cognitive network is composed of the set of cognitive engines which may reside inside a certain protocol layer, be implemented between different layers, or be distributed between different nodes in the network.

Each cognitive agent operates locally but it also contributes into global goals by interfacing with other cognitive agents. As a result, efficiency of cognitive network operation depends on the efficiency of communication between the agents. Depending on the scope, inter-layer, intra-layer, or at the network level, different communication technologies are used which put additional constraints in terms of speed and delay of information exchange. These constraints cannot be neglected and should be taken into account during the design of cognitive network architecture and its agents.

II. A taxonomy of Signaling Architectures

Initially, most of the signaling techniques appeared to overcome different limitations of standard TCP/IP protocol reference model. Depending on the scope signaling techniques can be divided into two broad categories: node-level signaling and network-level signaling.

Node-level signaling techniques provide the means for information exchange between different layers of the TCP/IP stack initially designed to be standalone and separated. Development of such techniques is mostly driven by the field of cross-layer design which enables a certain degree of cooperation between different layers to overcome the limitations of TCP/IP reference model.

Network-level signaling was initially designed as a part of TCP/IP and implemented using Internet Control Message Protocol (ICMP) [4] being an integral part of the IP layer. ICMP provides administrative assistance network nodes and help to determine host and route failures and other network inconsistencies.

In the following paragraphs, we briefly overview signaling mechanisms available in the literature describing their key characteristics.

A. Node-level Signaling

Interlayer signaling pipe is one of the first approaches used for implementation of cross-layer signaling [5], to allow the propagation of signaling messages layer-to-layer along the packet data flow. Signaling information, included in an optional portion of packet headers, follows the packet processing path within the protocol stack, either in a top-down or a bottom-up manner. An important property of this signaling method is that signaling information can be associated with a particular packet incoming or outgoing from the protocol stack.

Interlayer signaling pipe can be implemented using encapsulation of signaling information into packet headers, for example into an optional portion of IPv6 header [6], or using packet structures allocated by the protocol stack internally.

Direct Interlayer Communication (DIC), proposed in [5], aims at improvement of Interlayer signaling pipe method through the introduction of “signaling shortcuts” - performed out of band. As a result, DIC allows non-neighboring layers of the protocol stack to exchange messages, skipping processing at every adjacent layer.

Along with reduced processing overhead, DIC avoids insertion of signaling information into packet headers, which makes it suitable for bidirectional communication. Signaling is typically performed using ICMP protocol.

Despite the advantages of direct communication between protocol layers and a standardized way of signaling, the ICMP-based approach involves operation with heavy protocol headers (IP and ICMP), as well as significant protocol processing overhead.

The Central Cognitive Plane, implemented in parallel to the protocol stack, is probably the most widely proposed interlayer signaling architecture. Each protocol layer is extended with a tiny interface allowing exchange of information and configuration commands to/from the layer. These interfaces are interconnected with a cognitive engine using a common bus.

Implementation of this signaling method could be as simple as a shared database accessed by all the layers [7], while more advanced implementations introduce signaling interfaces as each protocol level internally providing an access to the internal protocol layer parameters and functions [8].

Central cognitive plane is used when optimization process involves availability of information and performing actions at more than two layers simultaneously. This way, a central point assesses values of internal protocol layer parameters via interlayer interfaces, makes an optimization decision, and communicates a set of actions that need to be performed back to the layers.

B. Network-level Signaling

Most of the existing cross-layer signaling proposals employ signaling between different layers within the protocol stack of a single node. However, as emphasized in [9], true cognitive networking should maintain a network-wide scope - with the cognitive process operating on end-to-end goals. Consequentially, cognitive networks require signaling approaches capable of signaling information delivery between different nodes in the network in an effective way.

Packet headers can be used for propagation of signaling information between different nodes of the network. Nowadays, many protocol headers of TCP/IP family, like TCP or IPv6, are extended with optional fields. Signaling information transmitted in these optional fields propagate along the packet flow and can be assessed at every router as well as end nodes. Such signaling methods keeps overhead at the minimum while allows signaling information be associated with a particular network packet. On the other hand, disadvantage of signaling using packet headers is in the limitation of signaling direction to the packet flow. However, this drawback can be resolved with the use of ICMP messages for signaling.

ICMP messages constitute the default signaling method from the early days in networking. Signaling information, encapsulated into ICMP and IP headers, can be directed and processed by the destination in the way ordinary IP data packets are routed in the network. Moreover, with a few exceptions ICMP messages are processed at the protocol stack kernel level rather than in the user application domain.

Signaling using ICMP messages is desirable when instant communication should be performed out of the regular data flow direction. In order to maintain association of signaling information with a particular packet an explicit reference to this packet should be included.

ICMP messages consume network bandwidth and influence delay resources of other flows corresponding to a heavy overhead solution. Thus, they should be used as a complimentary signaling scheme to packet header.

Explicit Notification schemes, like Explicit Congestion Notification (ECN) presented in [10], is another example of network-level signaling. ECN signaling is performed in-band by letting network routers to mark in-transit TCP data packets with a congestion notification bit. Then, at the received this marking is turned back in TCP acknowledgement directed to the sender node.

The main advantage of explicit notification schemes is a low overhead. The drawbacks are in the limitation of signaling propagation to the data packet paths, requirement for maintaining signaling loop through the receiver, as well as requirement of all network routers to support signaling and traffic generation functionalities. Comparison and Relevance for Cognitive Networks

In this section we compare available signaling approaches by the comparison of their individual characteristics like type of signaling, scope, signaling latency, communication overhead, in-band or out-of-band type of signaling, direction of signaling and whether signaling information can be associated with a particular packet flowing in the network.

It appears there is no optimal choice of signaling scheme performing well both for node- and network-level signaling in all the considered scenarios. For that reason, several signaling methods should be employed in cognitive networks at the same time to ensure efficient functionality of cognitive algorithms.

The following characteristics are relevant when considering the choice of signaling method.

Scope defines the boundaries of signaling method operation. Solutions limiting their operation to a single protocol stack tend to be more flexible in the choice of signaling techniques: they can use internal protocol stack techniques such as packet structures or callback functions, thus avoiding processing related overhead and the need for standardization effort.

Solutions operating at the node are suited for signaling between reconfigurable elements of cognitive network injected inside the protocol layers. In case only several protocol layers are concerned by a cognitive network implementation signaling is typically performed using direct interlayer communication methods. However, in case of many protocol layers concerned, either interlayer signaling pipe or central cognitive plane are the desired solutions.

Type of signaling corresponds to the communication primitives supported by each signaling method. Approaches encapsulating signaling information into packet structures, like interlayer signaling pipe, packet headers, and explicit notification, are limited to indication primitive. While other approaches performing out-of-band signaling transmissions can perform wider range of communication types including request-response actions.

According to the type of signaling the choice of appropriate approaches depends on the actions required to be performed between cognitive agents. At the node level, cognitive engine performing blind monitoring of the environment can be connected with the cognitive engine core using methods supporting indication primitive only. This will allow low-overhead communications. However, in case a cognitive agent should follow setup comments request-response actions become unavoidable requiring the use of heavier signaling approaches.

In-band / Out-of-band parameter shows whether existing data traffic flow is used as carrier for signaling information (in-band) or signaling information is sent on its own (out-of-band). In-band signaling methods do not add any significant overhead in term of network bandwidth and routing resources. However, the main drawback of in-band signaling, like packet headers or explicit notification, is in type of signaling limited to indication primitive only and relatively high latency of message delivery. On the other hand, out-of-band signaling is not constrained in signaling type and allows the fastest information delivery between ends. However, this is done at expense of network resources.

Direction of signaling is an important characteristic which defines the applicability of the signaling approach to the chosen cross-layer optimization scheme. The out-of-band signaling schemes are packet path independent and can provide a faster reaction to an event. This reaction can be preformed also in synchronous way, while packet path dependant signaling provides only asynchronous reaction. The speed and flexibility of path independent signaling comes at the expense of the additional communication resources. Nevertheless, path independence cannot be only considered as an advantage as it does not allow packet association.

Packet association shows whether signaling information can be associated with a specific packet transmitted through the network. Such property is required by many optimization approaches. For example, at the network level ECN signal sent along with a TCP data packet and echoed back with TCP acknowledgement by the receiver indirectly carries information related to TCP flow for which ECN signal is sent. At the node level information monitored at the physical layer (SNR or BER) is typically required to be associated with a packet it was measured for.

In-band signaling techniques maintain indirect association with between transferred signaling information and the packet used to carry it. On the other hand, if out-of-band signaling is used such association can be inserted explicitly. A good example is when “Time Exceeded” ICMP message sent by a router for a packet dropped due to expired TTL includes the copies of protocol headers of the packet dropped.

III. Concluding Discussion

Signaling is a key issue for the deployment of cognitive networking solutions. In this framework, several works are already available in the literature targeting cross-layer and inter-node signaling that could fit the cognitive networks requirements, even if no standard solution was proposed. Nevertheless, depending on the scenario, a broad range of proposed approaches for which it is difficult to select the one that best fits the specific cognitive network solution.

It should be underlined that specific issues on signaling should be addressed that are related to the cognitive networking domain, including security, problems with non-conformant routers, and processing efficiency.

Secure signaling is required in cognitive networks in order to avoid protocol attacks attempted by non-friendly network nodes, which furnish incorrect cross-layer information in order to trigger specific behavior.

Misbehavior of network routers represents another relevant issue, since it is estimated that in 70% of the cases IP packets with unknown options are dropped in the network or by the receiver protocol stack. In cognitive networks, such information loss would not be acceptable.

Finally, the problem with processing efficiency is related to the additional costs of the routers hardware for cross-layer information processing. While it is not an issue for the low-speed links, it becomes relevant for high speed ones where most of the routers decrement only the TTL field to maintain a high packet processing speed.

Signaling is a key issue for the deployment of cognitive networks. As information exchange is essential to the analysis and reasoning on the network status. However, signaling needs to be optimized since the choice of a specific signaling architecture can represent a performance bottleneck for the whole system.

References

1. D. D. Clark, C. Partridge, J. C. Ramming, and J. T. Wroclawski, A Knowledge Plane for the Internet, ACM SIGCOMM, Karlsruhe, Germany, August 2003.
2. Clash of the Titans – WiMAX and 4G: The Battle for Convergence is joined, Maravedis Market research and analysis, July 2006.
3. D. L. Tennenhouse, J. M. Smith, W. D. Sincoskie, D. J. Wetherall, and G. J. Minden, A survey of active network research, IEEE Communications Magazine, vol. 35, no. 1, pp. 80 – 86, January 1997.
4. J. Postel, “Internet Control Message Protocol,” RFC 792, September 1981.
5. Q. Wang and M. A. Abu-Rgheff, “Cross-layer signaling for next-generation wireless systems,” IEEE Wireless Communications and Networking (WCNC), pp. 1084 – 1089, 2003.
6. S. Deering and R. Hinden, “Internet Protocol, Version 6 (IPv6) Specification,” RFC 2460, 1998.
7. K. Chen, S. H. Shah and K. Nahrstedt, “Cross-Layer Design for Data Accessibility in Mobile Ad hoc Networks,” Wireless Personal Communications, Special Issue on Multimedia Network Protocols and Enabling Radio Technologies, vol. 21, pp. 49-75, 2002.
8. V. T. Raisinghani and S. Iyer, “Cross Layer Feedback Architecture for Mobile Device Protocol Stacks,” IEEE Communications Magazine, vol. 44, no. 1, pp. 85 – 92, 2006.
9. R. W. Thomas, D. H. Friend, L. A. DaSilva, and A. B. MacKenzie, “Cognitive networks: adaptation and learning to achieve end-to-end performance objectives,” IEEE Communications Magazine, vol. 44, no. 12, pp. 51 – 57, December 2006.
10. K. Ramakrishnan, S. Floyd, and D. Black, The Addition of Explicit Congestion Notification (ECN) to IP, RFC 3168, 2001.
11. D. Kliazovich, F. Granelli, and Nelson L. S. Da Fonseca, "Architectures and Cross-Layer Design for Cognitive Networks," World Scientific Publishing Co, Inc. River Edge, Handbook on Sensor Networks, 2009.

About the Authors

Dzmitry Kliazovich

Research Fellow, University of Luxembourg

Fabrizio Granelli

Editor-in-Chief, ICaST - Associate Professor, University of Trento

Nelson Luis Saldanha da Fonseca

Full Professor and Head of Computer Systems Department, University of Campinas