CogNet: a cognitive complete knowledge network system

ZORZI LAYOUT

12/3/08

1:58 PM

Page 81

ACCEPTED FROM OPEN CALL

COGNET: A COGNITIVE COMPLETE KNOWLEDGE NETWORK SYSTEM B. S. MANOJ AND RAMESH R. RAO, UNIVERSITY OF CALIFORNIA SAN DIEGO MICHELE ZORZI, UNIVERSITY OF PADOVA AND UNIVERSITY OF CALIFORNIA SAN DIEGO

ABSTRACT Distributed repository

Repository updates/queries

The authors propose CogNet — a cognitive complete knowledge network system — which makes use of a large amount of information that can be gained from the experience of each node to improve the overall network and user performance.

The benefits of using cognitive information at the physical layer, as in cognitive radios, are many. In this article, we propose CogNet — a cognitive complete knowledge network system — which makes use of a large amount of information that can be gained from the experience of each node to improve the overall network and user performance. CogNet gathers, processes, analyzes, and stores information available through a variety of devices and protocols to build an omnipresent, distributed repository that holds the spatiotemporal, network-experience information. The inexpensiveness and plentifulness of storage resources and increasing processing power in handheld devices help accelerate the development of CogNet-like systems. Our contribution in this article is the proposal of the architecture and of the communication elements, as well as a transport layer application of CogNet as a proof of concept for possible application scenarios of such a system. We also present performance evaluation of CogTCP, the CogNetenabled TCP, which exploits the transport layer experience of other nodes for improved performance. From our experiments, we found that the use of cognitive information is very useful for networking.

INTRODUCTION A variety of systems have been deployed in recent years to provide users access to a wide range of communication, information, and entertainment services over a wired or wireless network. For the most part, these systems are individually designed, deployed, and optimized around information that usually is gathered within the elements of the particular system. Although these systems demonstrate the feasibility of extending new innovative services, they fail to systematically gather and retain information useful for the effective use of the larger set of networked systems to which end users seek access. In this article, we describe the Cognitive Complete Knowledge Network (CogNet), a novel approach that identifies gaps in the aware-

IEEE Wireless Communications • December 2008

ness of attributes (e.g., spectrum availability, interference conditions, usage patterns) and the network parameters that limit user satisfaction. It seeks ways to fill these gaps by actively gathering and persistently preserving large amounts of user and network observations, and it develops new inference algorithms to extract useful information from this data set. This is accomplished by resorting to statistical learning techniques that support the fusion of information obtained through direct measurement, prior experience, and experience acquired from peer devices in the network, so as to build statistical models that enable optimal (highly distributed) decisions for all facets of the operation of individual devices in the network. Much of the information archived in CogNet is gathered from the internal state of network entities such as mobile devices, network access points, and base stations. The information could be sampled at the physical, link, network, and transport layers, as well as at the application itself. Development of numerous tools for actively or passively measuring network traffic and flow patterns in the wireless domain is very important to understand the trends and anomalies and to improve the efficiency of resource management. CogNet takes a system-level approach that is made possible by the following current technological trends: • Rapid drop in the cost of storage: gigabytes of memory will become commonplace in cellular devices in the near future, making it possible to use large amounts of storage to assist with data gathering and archiving. • Advances in learning algorithm design for data analysis. • Advances in processing capabilities and intelligence, both in the network and in the portable devices. • Rapid strides in the development of programmable radio interfaces. These features make it possible to exploit the wealth of information that is naturally present in the network to improve network management and resource efficiency. Unlike existing approaches in cognitive radio and networks, which mostly focus on radio properties and lower layer performance, our cross-

1536-1284/08/$25.00 © 2008 IEEE

Authorized licensed use limited to: Univ of Calif San Diego. Downloaded on August 14, 2009 at 00:31 from IEEE Xplore. Restrictions apply.

81

ZORZI LAYOUT

12/3/08

1:58 PM

The primary motivation behind CogNet is the expected benefit that can be derived from using the cognitive capability in a system-wide manner.

Page 82

Data analysis module

CogNet applications: adaptive spectrum management, user-profile-based resource management, etc.

Intelligent information repository

Distributed data processing, compacting storage, analysis, modeling, and retrieval

User profiles

User preference

WiMAX and other wireless access networks Cellular networks Wireless LANs and mesh networks

Cell-phone-based sensor platforms

Specialized tools (TEMS, End-user SDR CAIT, and Chester) platforms

Network elements such as CalMesh routers

Control channels

n Figure 1. Schematic diagram of the CogNet architecture. layer approach spans all layers of the protocol stack and explicitly addresses higher layer performance and adaptation issues. In fact, the vast majority of existing works in cognitive radio networks focus mainly on spectrum sharing between primary and secondary users. In contrast, even though some of the concepts developed here apply to more general systems, in this article we mostly refer to situations where unlicensed bands are used (e.g., IEEE 802.11 or hybrid WLAN/3G scenarios). In addition, there exist a plethora of potential applications that can be developed using the cognitive capabilities embedded within all the networking elements, a paradigm that goes well beyond the concept of cognitive radio. Through our work, we attempt to shed light on those potential areas of application in addition to dynamic spectrum management.

RELATED WORK Software radios have become an established paradigm since the original concept was introduced by Mitola in the late 1990s [1]. The ultimate software radio would include only wide-band smart antennas and analog-to-digital /digital-to-analog (AD/DA) circuits in hardware, whereas everything else, including physical (PHY) and radio frequency (RF) processing, would be performed by software in powerful processing units. In addition to these features, a cognitive radio [2, 3] also has the ability to learn its own capabilities, user behavior, and the physical and radio environment to execute complex adaptations and configure itself to best suit the situation. Whereas software radios are fairly well understood, cognitive radios are actively under research. In recent years a number of projects have been initiated in these areas that focus mostly on reconfigurable and frequency-agile radio testbeds for dynamic and opportunistic spectrum management. In most cases, except for medium access control (MAC) issues, there is relatively little focus on networking in general and on the overall network system in particular.

82

The concept of cognitive networks was first introduced by Clark et al. in [4]. The three important elements of such networks are cognitive radios, cognition engines, and the layered-network protocol stack. The cognition engines infer knowledge by gathering information through a variety of sources such as radios, sensors, user profiles, and the different layers of the network protocol stack. The higher layers can benefit from the inference made by the cognition engine and utilize the unique features of the programmable lower layers. In addition, cognitive networks must be able to learn from their actions. Recent research on using cognitive information in networking includes: studies on the MAC [5], initial discussion on dynamic spectrum access in multihop networks [6], and discussion on using cognition in the Internet [7]. Despite some studies that recently moved toward a more complete view of a cognitive network architecture [7, 8], a gap still exists in learning the spatiotemporal aspects of the protocol performance at every layer. For example, the temporal and spatial periodicity of higher layer traffic is not utilized in any of the above mentioned work. Also, current work addresses only specific aspects. The primary motivation behind CogNet is the expected benefit that can be derived from using the cognitive capability in a system-wide manner. As a distinguishing feature from the existing approaches, our system concept focuses more directly on the intelligent use of cognitive radio and cognitive networking, coupled with an application quality of service (QoS)/user experience and reconfigurable protocol stacks, to make creative use of information that is naturally present and gathered in wireless systems. To achieve this, it also advocates the use of a cross-layer paradigm, not only within a single device but also across different devices. As a concrete example of this paradigm, we describe and evaluate a cognitive Transmission Control Protocol (TCP) scheme.

IEEE Wireless Communications • December 2008 Authorized licensed use limited to: Univ of Calif San Diego. Downloaded on August 14, 2009 at 00:31 from IEEE Xplore. Restrictions apply.

ZORZI LAYOUT

12/3/08

1:58 PM

Page 83

Measurement system

Observer

Application layer

Cognitive application module

Analyzer

Transport layer

Cognitive transport module

Network layer

Cognitive network module

Programmable MAC and link layer

Cognitive MAC module

Programmable physical layer

Cognitive physical layer module

Detection algorithms

Session management

Resource management

Mobility management

Programmable MAC layer Reconfigurable physical layer (a) Decision flow diagram in CogNet

Decision maker

Reconfigurer

Cognitive cross layer bus

(b) Cognitive cross layer bus

n Figure 2. Communication elements in CogNet.

THE COGNET ARCHITECTURE The overall system schematic diagram for CogNet is shown in Fig. 1. In this architecture, cognitive agents gather information from their associated network element and also control the behavior of that element by setting parameter values or modes of operation. A network element can be a protocol (e.g., TCP), a device (e.g., a radio), or an environmental sensor. The use of these agents simplifies the representation of salient events and of the mechanisms of control for any particular network element. Example parameters of interest at the physical layer include noise level, interference temperature, path loss, delay spread, coherence bandwidth, physical layer data rate, spectrum occupancy parameters, and the coding scheme used. Examples of parameters at the higher layers include the end-to-end delay, jitter, throughput, throughput fairness, call-acceptance rate, call-dropping rate, and perceived user quality. All such parameters are spatiotemporally tagged and stored at a common intelligent repository, which is a key element of the proposed architecture. Our architecture is fully distributed and requires a cognition module at every layer, which serves as a local sensor, controller, and actor, gathering information and controlling the protocol parameters within that layer. This is very much required as far as applying cognitive networking is concerned, due to the following reasons: • Only this architecture maintains the layered abstraction of the networking protocol stack, which is one of the primary factors behind the successful evolution of the computer networks of today. • This architecture can simplify the complexity of cognitive processes that otherwise may become unmanageable. In addition, the semantic interpretation of network events, the behavior of protocol parameters, and the actions taken at every layer can be more efficiently handled if each layer has a cognitive module of its own. A cognitive plane helps coordination of the cognitive modules and of the information and data exchange through an internal cognitive bus, as well as of communication

with other CogNet-enabled nodes and the common repository. In the following, we describe in more detail the main components of CogNet. Cross-Layer CogNet Bus — The cognitive agents exchange information with each other and the local storage over a cross-layer bus, the CogNet bus. The flat structure of a bus allows for new hierarchies to develop in the sharing and processing of information. The cognitive modules at each layer of the networking protocol stack provide a possibility for joint optimization and scheduling of resources, as well as an opportunity for dynamic resource allocation and management with the help of the past history of the user, device, and network. It is important to take the cognition aspect in the vertical integration of the network layers. The CogNet bus will provide a means to publish or exchange cognitive information across various layers to achieve a joint resource optimization and scheduling framework for an improved user experience. The CogNet bus overrides the intermediate layers in the interaction between a pair of non-adjacent layers, for example, the application layer and the physical layer. There are two major challenges in designing the cross-layer bus architecture. The first is the requirement for a lightweight design, and the second is the definition of a format for exchanging CogNet-specific information across all layers. The impact of information format, bus design, and ease of implementation on a variety of heterogeneous systems, as well as the performance achieved, must be thoroughly investigated and are part of our ongoing work. Cognitive Plane — The Cognitive Plane (CogPlane) is responsible for translating the end goals to the responsibilities or action items required for each layer. The cognitive modules at each layer of the networking protocol stack report their observations, which will be collected by the CogPlane and stored in a local repository. The cognitive executive function (CogExec) within the CogPlane encapsulates the highest level supervisory functions in the CogNet system. CogExec builds an interaction model for the net-

IEEE Wireless Communications • December 2008 Authorized licensed use limited to: Univ of Calif San Diego. Downloaded on August 14, 2009 at 00:31 from IEEE Xplore. Restrictions apply.

83

ZORZI LAYOUT

12/3/08

1:58 PM

To exchange the information learned by the short-term learning process, the CogNet bus is used. Upon the user application request, the CogPlane executes algorithms for joint optimization and scheduling of resources.

Page 84

work elements and designs mechanisms that the relatively rudimentary cognitive agents can effect to respond to more complex global goals. CogExec learning algorithms are applied to glean behavioral models of critical network parameters. There are two levels of learning activity involved in this system: • Short-term learning within individual modules such as physical layer, network layer, and transport layer • Long term learning for the overall system, which is carried out at the distributed intelligent repository To exchange the information learned by the short-term learning process, the CogNet bus is used. Upon the user application request, the CogPlane executes algorithms for joint optimization and scheduling of resources. These optimization algorithms generate the right parameters to be selected at each of the network layers, and the cognitive modules are responsible for reconfiguring the protocols at each individual layer. Thus the CogPlane provides an opportunity for dynamic resource allocation and management with the help of the past history of the user, the device, and the network. In a heterogeneous networking environment, it is important to represent information flexibly and in a way that can be interpreted across different technologies and protocols, without explicit knowledge of the technology-specific details. This makes the design modular and more scalable and makes it easier to add new air interfaces to the system. An example of such a technology-independent representation of crosslayer information based on fuzzy logic is described in [9]. CogNet Repository — CogNet is a large-scale distributed environment with numerous devices of various capabilities, where heterogeneous data is generated, and in some cases also consumed, by the CogNet agents (which can be protocols, network nodes, embedded devices, etc.). According to its scope, this information may be stored within a limited local repository or alternatively, transferred to a (possibly distributed) central repository. Where data is stored also depends on bandwidth availability and the cost of communication. When retrieved from a local or remote repository, this information can be used, for example, by the application layer for making a service quality decision or by the network layer for better routing. The information storage infrastructure must incorporate a variety of requirements, for example, reliability, accuracy, and time constraints. Current database technology can support many of these requirements, although, in some cases, new query types may be required to handle the spatiotemporal dynamic information storage and retrieval. A key challenge in CogNet is to incorporate the CogNet agents into the system architecture and to enable a dynamic information storage mechanism in a resource-efficient way. Due to heterogeneity in the nature and capabilities of CogNet agents, connection capabilities (partially connected, fully connected, or disconnected), state (idle, sleep, active, etc.), and so on, a simple XML-based representation of data is desir-

84

able because it enables the utilization of existing methods and mechanisms for information storage and retrieval. The CogNet agents and the central repository also can execute protocols to provide quality guarantees to the information exchange between producers and users of this data. One very important issue when designing the repository is to preserve the privacy of data producers. For example, tagging the spatiotemporal parameters can help mask the actual identity of the cell phone that generated the information regarding the network status at any geographical location. CogNet Radio System — One of the main components of the CogNet system will be the radio devices. These are responsible for providing effective means to communicate and to implement the physical layer adaptations in response to the environmental awareness of CogNet. A key feature in the CogNet system is the heterogeneity of devices and technologies, which poses significant technical challenges but also opportunities in terms of sharing capabilities (e.g., a more powerful device could make some measurements available to other devices that otherwise would not be able to gather them). In this context, the cognitive radio framework is a useful starting point [2]. All the attributes of a cognitive radio can still be considered, with the important enhancement that the radio be aided by useful and often difficult to obtain information mined from the network data repository. This leads to the important question of the appropriate architecture for such a radio wherein information from environmental sensing is combined with prior knowledge learned from the data repository. Developing flexible transmit waveforms (with adaptive RF bandwidth, modulation type, etc.) and transmit strategies (spatial multiplexing vs. space-time coding, channel state information feedback, etc.) by optimally integrating instantaneous measurements and information archived in a CogNet is a key research challenge. We point out that our cognitive network architectural concepts do not necessarily require the use of cognitive radio nodes so that an interesting alternative is to test them using more “traditional” (and therefore cheaper and more easily available) hardware platforms. This may help popularize cognitive networking concepts within a larger community without imposing large investments or requiring specialized hardware expertise. We are actively pursuing this line of research based on the CalRadio [10] and CalMesh [11] platforms. Reasoning Algorithms — The most challenging part in cognitive networking is the decision-making process that requires efficient data mining, analysis, and inference over the large set of spatiotemporally tagged data. In addition to the use of traditional concepts such as clustering, association, classification, k-means, and time-series analysis, machine learning concepts such as supervised, unsupervised, and semi-supervised learning approaches must be employed. Examples of such artificial intelligence algorithms include Bayesian learning, artificial neural networks, decision trees, the variational Bayes

IEEE Wireless Communications • December 2008 Authorized licensed use limited to: Univ of Calif San Diego. Downloaded on August 14, 2009 at 00:31 from IEEE Xplore. Restrictions apply.

ZORZI LAYOUT

12/3/08

1:58 PM

Page 85

approach, Markov random fields, reinforcement learning, and Q-learning (e.g., see [12]). Reliability of the inference is critical, and new computationally efficient solutions along the above-mentioned algorithmic directions are required to achieve them. In addition, the base data set generated in such cognitive networking is enormously multidimensional in nature, and therefore, extensions of the traditional approaches are required to obtain computationally efficient solutions.

COGTCP: A COGNET EXAMPLE AT THE TRANSPORT LAYER CogTCP is a version of TCP that includes a cognitive transport module (CTM) with intra-layer, inter-layer, and inter-node cognitive capability. In a situation where a busy Internet server accepts thousands of TCP connections every second from a large number of networks, currently every new TCP connection must undergo the same transport layer behavior, for example, the slow-start phase, congestion management phase, and transmission window behavior. In CogNet, the mapping of the TCP behavior models, current and past, to the host addresses and network addresses in a spatiotemporal manner helps the wireless clients optimize the protocol parameters, thereby improving the performance. The transport-layer behavior mapping can be either coarse grained (mapping with the network part of the host address) or fine grained (mapping with the destination host address). The intra-layer cognitive capability refers to the ability of the module to learn from various internal transport-layer functional modules (e.g., socket structures and transmission control blocks). The inter-layer learning capability refers to the ability of the CTM to interact with other layers through the cross-layer cognition bus. The inter-node cognitive capability helps a node to obtain cognitive information from other nodes. An example of inter-node learning that involves the transport layers of two different devices is provided in Fig. 3. When a client wireless host connects to a server in the Internet through an access point or a base station, the access point may have accumulated past experiences from TCP connections to the same server or to the same destination network. In some cases, there might be multiple on-going TCP connections from which information is gathered. Even if the server does not have multiple connections at any given time, the spatiotemporal behavior registered by prior connections and/or prior nodes can be used for helping the ongoing or start-up connections. Therefore, a new TCP connection in a wireless client can query the existing connections, clients, or the past history information to avoid slow start and unnecessary throughput degradation related to congestion control. Although many adaptation techniques were proposed in the past to improve TCP performance in wireless systems, none of them has ever considered exchanging information across devices. In the system shown in Fig. 3, Node A initiates a connection with an external server and

FTP server Distributed repository Data traffic sessions

Network Repository updates/queries

Node A

Node B

n Figure 3. Schematic diagram of a simple CogNet system. uploads a description of its network experience in terms of the information regarding TCP connection parameters to the central repository. At this point, we do not consider explicitly the cost of updating the parameters to the central repository, which can be done in an asynchronous or synchronous way. The asynchronous method exploits the huge and inexpensive memory storage resources to accumulate the network experiences and upload them to the central repository when time and bandwidth resources permit. On the other hand, the synchronous repository update provides online updates to the repository, as and when the network parameters are monitored. The initial parameters collected with a time stamp are current congestion window, congestion window averaged since the beginning of the connection, receiver advertised window, unacknowledged sequence number, current slow-start threshold, average slow-start threshold, round-trip time, smoothed round-trip time, and run-time throughput. After these parameters are collected, time stamped, and stored in the repository, they can be used by Node B when it initiates a connection to the external server. Node B queries for parameters such as average congestion window and slow-start threshold and uses them as the initial values for the congestion window and slow-start threshold, respectively, for its own new TCP connection. Although these parameters are sensitive to spatiotemporal dimensions and to network dynamics, the network traffic pattern can be expected to be stationary over a sufficiently long period of time.

SIMULATION SET UP Some initial results obtained through simulation are briefly illustrated here as an example. We used Glomosim with distance vector routing and IEEE 802.11 Distributed Coordination

IEEE Wireless Communications • December 2008 Authorized licensed use limited to: Univ of Calif San Diego. Downloaded on August 14, 2009 at 00:31 from IEEE Xplore. Restrictions apply.

85

ZORZI LAYOUT

12/3/08

1:58 PM

Page 86

Function (DCF) MAC at 2 Mb/s. The network topology chosen (Fig. 4) was a grid topology with grid dimension set to 300 meters and a transmission power of about 15 dBm, which gives an approximate transmission range of 375 meters when simulated with a two-ray propagation model. We used a 25-node network with a File Transfer Protocol (FTP) server running on node 24 and nodes A and B running FTP clients. Initially, when node A runs an FTP session, the CTM module within the transport layer keeps track of the parameters and updates them in the common repository, which can be centralized or distributed, and is not shown explicitly in Fig. 4. When node B tries to use a

375 meters 20

21

22

23

FTP server 24

15

16

17

18

19

10

11

12

13

14

5

6

7

8

9 300 meters

0

1

Node A

Node B

2

3

4

300 meters

n Figure 4. The network topology used for simulation experiments.

11000 10000

Node A’s CWND Node B’s CWND with CogNet Node B’s CWND without CogNet

TCP congestion window

9000 8000 7000

CONGESTION WINDOW SIZE We tracked the time evolution of the congestion window for both node A and node B. At time 20 s, node A begins an FTP session with the FTP server for transferring a short file of length 10 KB (Fig. 5). At time 20.8 s, node A’s connection completes, and the connection parameters, such as the current congestion window, average congestion window, and slow-start threshold, are uploaded to the repository. At time 21 s, node B initiates a new FTP connection with the FTP server and obtains the protocol parameters that are stored in the repository, namely, the average congestion window (4196 bytes) and the last value of the slow-start threshold (16 KB) experienced by node A. In this case, node B uses the average congestion window from the repository as the initial congestion window, that is, its initial congestion window is set to 4196 bytes. The time evolution of the congestion window for the new connection originated by node B, with and without CogTCP, is also shown in Fig. 5. We noted that the average congestion window at the end of node B’s connection improved from approximately 4000 bytes to about 5376 bytes when node B used CogTCP. Therefore, node B’s transport layer connection could benefit from the experience gained by node A and made available through the CogNet repository. Figure 6 shows the congestion window as a function of time for long file transfer sessions, where we also noted improvement in the average congestion window and the transfer time. When node B used CogTCP, the file transfer ended at approximately 115 seconds in comparison to the time taken by normal TCP, which was approximately 130 seconds. The file transfer time for node B’s file transfer session was reduced by approximately 15 seconds when node B used CogTCP.

THROUGHPUT PERFORMANCE

6000 5000 4000 3000 2000 1000 20000 20200 20400 20600 20800 21000 21200 21400 21600 21800 Time (ms)

n Figure 5. Congestion window vs time for 10 kbyte file transfer.

86

TCP connection with the FTP server, it queries the repository about the right protocol parameters observed by previous nodes (in this case, node A) within the spatiotemporal domain. Hence, node B receives the parameters for average congestion window and slow-start threshold. Node B uses the average congestion window and the slow-start threshold from node A as its own initial values and begins its data transfer session with the FTP server. To estimate the advantages of CogTCP, we focused on the behavior of key TCP performance indicators such as the congestion window size and the throughput.

We ran a 100-seed simulation campaign for estimating throughput improvements in CogTCP due to the exploitation of the information available from the CogNet repository. As before, we used short- and long-file transfer sessions for these experiments. For the short-file transfer experiments, we set up FTP sessions with file size fixed at 10 KB. Initially we considered data transfer sessions without background traffic. Some results for this case are presented in Table 1. The throughput performance for short files shows an improvement for CogTCP compared to traditional TCP, which does not exploit cognitive

IEEE Wireless Communications • December 2008 Authorized licensed use limited to: Univ of Calif San Diego. Downloaded on August 14, 2009 at 00:31 from IEEE Xplore. Restrictions apply.

12/3/08

1:58 PM

Page 87

information. Table 1 also shows the throughput performance over traditional TCP when CogTCP was used for transferring longer files of a length of 1 MB. Here again, we noticed a similar throughput improvement for CogTCP when compared to traditional TCP. The throughput performance in the presence of background traffic generated by constant bit rate (CBR) sources is also shown in Table 1. The background CBR connections were originated from random source nodes and all terminated at node 24. As before, in this experiment, we used short-file transfer sessions (10 KB) and long-file transfer sessions of 1 MB and 10 MB for the FTP sessions from nodes A and B to the FTP server. When we used 10 CBR sessions for creating the background traffic, we noticed that the average throughput achieved by both TCP and CogTCP is reduced significantly when compared to the throughput achieved in the absence of background traffic. However, the relative throughput gain for CogTCP compared to traditional TCP remained almost the same as that of the experiments without background traffic.

18,000 Node A’s CWND (TCP) Node B’s CWND (CogTCP) Node B’s CWND (TCP)

16,000 14,000 Congestion window

ZORZI LAYOUT

12,000 10,000 8000 6000 4000 2000 0 20,000

40,000

60,000

80,000 Time (ms)

100,000

120,000

140,000

n Figure 6. Congestion window vs time for 1 Mbyte file transfer.

Throughput without background traffic (kb/s)

CONCLUSION Cognitive radios are able to exploit the information gathered at the physical layer and can achieve many benefits. In this article, we attempted to expand the scope of the cognitive radio approach, using cognitive information across the networking layers and across devices in the network as well. We proposed CogNet, a cognitive complete knowledge network, which is a system that actively acquires, processes, and extrapolates information from a multitude of sources to maintain and disseminate awareness of the context in which the networked users interact. Much of the information archived in CogNet will be gathered by mobile devices from their own internal state. In addition, specialized devices that are designed to enable the data acquisition capabilities can also be used. As an example of the application of these concepts, we studied the performance of CogTCP, a CogNetenabled version of TCP, in which the information gathered from the previous connections of other nodes is used to improve the transport layer performance of a given node. We studied the benefit of using CogTCP for long- and shortfile transfer sessions. We found improvements in terms of average congestion window size of CogTCP sessions and throughput achieved when compared to traditional TCP. This simple example of CogTCP was intended to show how the use of cognitive information in a very simple way can provide performance gains. More sophisticated techniques to do so, including learning algorithms, are expected to achieve more significant gains and are the subject of future studies.

ACKNOWLEDGMENT This work was supported by NSF within the project “CogNet: Cognitive Complete Knowledge Network” (award number 0650048) and by the European Commission within the ARAGORN project (FP7-ICT-2007-1, grant agreement no. 216856).

TCP

CogTCP

Short file transfer sessions (10 KB)

127

139

Long file transfer sessions (1 MB)

62

66

Throughput with background traffic (kb/s) TCP

CogTCP

Short file transfer sessions (10 KB)

120

135

Long file transfer sessions (1 MB)

55

57

Long file transfer sessions (10 MB)

44

50

n Table 1. CogTCP throughput performance with and without background traffic.

REFERENCES [1] J. Mitola, “The Software Radio Architecture,” IEEE Commun. Mag., vol. 33, no. 5, May 1995, pp. 26–38. [2] J. Mitola and G. Q. Maguire, “Cognitive Radio: Making Software Radios More Personal,” IEEE Pers. Commun., Aug. 1999, pp. 13–18. [3] S. Haykin, “Cognitive Radio: Brain-Empowered Wireless Communications,” IEEE JSAC, vol. 23, no. 2, Feb. 2005, pp. 201–20. [4] D. D. Clark et al., “A Knowledge Plane for the Internet,” Proc. ACM SIGCOMM ’03, Aug. 2003. [5] C. Doerr et al., “MultiMAC — An Adaptive MAC Framework for Dynamic Radio Networking,” Proc. IEEE DySPAN 2005, Nov. 2005, pp. 548–55. [6] P. Kyasanur and N. H. Vaidya, “Protocol Design for Multihop Dynamic Spectrum Access Networks,” Proc. IEEE DySPAN 2005, Nov. 2005, pp. 645–48. [7] D. Raychaudhuri et al., “CogNet: An Architectural Foundation for Experimental Cognitive Radio Networks within the Future Internet,” Proc. ACM/IEEE MobiArch ’06, San Francisco, CA, Dec. 2006, pp. 11–16. [8] R. W. Thomas et al., “Cognitive Networks: Adaptation and Learning to Achieve End-to-End Performance Objectives,” IEEE Commun. Mag., vol. 44, no. 12, Dec. 2006, pp. 51–57. [9] N. Baldo and M. Zorzi, “Fuzzy Logic for Cross-Layer Optimization in Cognitive Radio Networks,” IEEE Commun. Mag., vol. 46, no. 4, Apr. 2008, pp. 64–71.

IEEE Wireless Communications • December 2008 Authorized licensed use limited to: Univ of Calif San Diego. Downloaded on August 14, 2009 at 00:31 from IEEE Xplore. Restrictions apply.

87

ZORZI LAYOUT

12/3/08

1:58 PM

Page 88

[10] CalRadio; http://calradio.calit2.net [11] CalMesh; http://calmesh.calit2.net [12] C. M. Bishop, Pattern Recognition and Machine Learning, Springer, 2006.

ADDITIONAL READING [1] B. S. Manoj, R. R. Rao, and M. Zorzi, “Architectures and Protocols for Next Generation Cognitive Networking,” in Cognitive Wireless Networks: Concepts, Methodologies and Visions, F. H. P. Fitzek and M. D. Katz, Eds., Springer, 2007.

BIOGRAPHIES B. S. M ANOJ [M] ([email protected]) received his Ph.D. degree in computer science and engineering from the Indian Institute of Technology, Madras, in July 2004. He is currently an assistant research scientist and lecturer in the Electrical and Computer Engineering Department, University of California at San Diego (UCSD). He leads the Robust Networking and Information Collection Project, part of the NSF-sponsored project titled RESCUE — Responding to Crises and Unexpected Events. His current research interests include ad hoc wireless networks, next generation wireless architectures, wireless sensor networks, wireless mesh networks, and cognitive networking. He is a recipient of the Indian Science Congress Association Young Scientist Award for the Year 2003 and the IBM Best Ph.D. thesis award for the year 2004. He co-authored the widely used textbook Ad Hoc Wireless Networks: Architectures and Protocols (Prentice Hall PTR). He co-authored publications that were chosen for best paper awards at IEEE CCNC ’08, IEEE/ACM HiPC ’04, and IRISS ’02. He has more than 75 publications, which include 21 international journal publications in the area of wireless networking. He is a reviewer for Elsevier Computer Networks, Elsevier Ad Hoc Networks, Elsevier Computer Communications, IEEE/ACM Transactions on Networking, IEEE Transactions on Wireless Communications, and IEEE Transactions on Mobile Computing. He has also served on the Technical Program Committees of several IEEE/ACM conferences. His recent TPC services, in addition to being a reviewer for IEEE INFOCOM ’08, include IEEE ICC ’08, IEEE GLOBECOM ’08, IEEE HPCC ’08, and ICDCN ’08. R A M E S H R. R A O [SM] ([email protected]) earned his B.E. degree with honors in electronics and communications in 1980 from the University of Madras, Tiruchirapalli, India. He received his M.S. degree in 1982 and Ph.D. degree in

88

1984, both at the University of Maryland, College Park. Since then he has been on the faculty of the Department of Electrical and Computer Engineering at UCSD, where he is currently a professor and director of the San Diego Division of the California Institute of Telecommunications and Information Technology. In April 2004 he was named Qualcomm Endowed Chair in Telecommunications and Information Technology. His research interests include architectures, protocols, and performance analysis of computer and communication networks. He has authored over 100 technical papers, contributed two book chapters, and led many major interdisciplinary and collaborative projects as principal investigator. He was Editor of IEEE Transactions on Communications and a member of the editorial board of ACM/Baltzer Wireless Network Journal, as well as IEEE Network. Twice he has been elected to serve on the Information Theory Society Board of Governors (1997–1999 and 2000–2002). M ICHELE Z ORZI [F] ([email protected]) received a Laurea degree and a Ph.D. in electrical engineering from the University of Padova, Italy, in 1990 and 1994, respectively. During academic year 1992–1993, he was on leave at UCSD, attending graduate courses and doing research on multiple access in mobile radio networks. In 1993 he joined the faculty of the Dipartimento di Elettronica e Informazione, Politecnico di Milano, Italy. After spending three years with the Center for Wireless Communications at UCSD, in 1998 he joined the School of Engineering of the University of Ferrara, Italy. In 2003 he joined the Department of Information Engineering of the University of Padova, where he is currently a professor. His current research interests include performance evaluation in mobile communications systems, random access in mobile radio networks, ad hoc and sensor networks, energy constrained communications protocols, underwater networking, and cognitive networks. He was the Editor-in-Chief of IEEE Wireless Communications from 2003 to 2005 and is currently the Editor-in-Chief of IEEE Transactions on Communications. He serves on the steering committee of IEEE Transactions on Mobile Computing and on the editorial boards of IEEE Transactions on Wireless Communications, Wiley Journal of Wireless Communications and Mobile Computing, and ACM/URSI/Kluwer Journal of Wireless Networks. He was also guest editor for special issues for IEEE Personal Communications (“Energy Management in Personal Communications Systems”) and IEEE Journal on Selected Areas in Communications (“Multimedia Network Radios” and “Underwater Wireless Communications and Networking”).

IEEE Wireless Communications • December 2008 Authorized licensed use limited to: Univ of Calif San Diego. Downloaded on August 14, 2009 at 00:31 from IEEE Xplore. Restrictions apply.