Cooperative Power Saving Strategies in Wireless Networks: an Agent-based Model Federico Albiero #1 , Frank H.P. Fitzek ∗2 , Marcos Katz #3 # 1
VTT, Technical Research Centre of Finland
[email protected], ∗
3
[email protected]
Aalborg University, Denmark 2
[email protected]
Abstract— Power consumption of mobile devices is a major subject of concern for the wireless domain. New multimedia services and advanced communication techniques such as upcoming TV-on-mobile will consume in the early future even more power. This may result in a troublesome energy trap: therefore some solutions shall be found. In this paper a novel system architecture is proposed in order to address power saving, which is neither pure cellular neither peer-to-peer (P2P). Our scheme is based on a cooperative framework that relies on the combination between a central and a short-range communication link. In this study, the power saving technique is tested by means of an agentbased simulation model. The system is designed as a Virtual World of wireless mobile terminals. The nodes are conceived as autonomous agents capable of independent decision-making based on strategies, with no imposed central authority. The strategies pursue the goal of making the cooperative clusters to achieve energy savings. The results of the numerical analysis show considerable power saving potentials of this approach.
I. S CENARIO OF I NVESTIGATION : C OOPERATIVE M ULTICAST P OWER S AVING T ECHNIQUE Cooperation is the strategy of a group of entities working together to achieve a common or individual goal. In this investigation, we aim to map cooperative behavior among individual entities into the world of wireless systems. We motivate the introduction of such a technique into the wireless domain because of the widespread success of cooperative behavior in nature and social sciences [1]. Moreover, the problem of cooperation has been recently analyzed in the field of game theory. Particularly, it has been studied by several authors under what conditions cooperation can arise in a world of individual entities (see, for instance, [2]). Although we are not directly dealing with game theory, we will show that our approach is well suited for the emerging application of gametheoretic analysis to wireless communications [3]. The cooperative framework we introduce here relies on a novel network architecture [4]. The key feature is the possibility to exploit the potentials of combined data reception/transmission among the terminals over a central (or cellular, C) and a short-range (SR) communication link. We notice this is a hybrid approach, which is neither pure cellular, neither peer-to-peer. We will consider a specific application of this idea to address the ever increasing power consumption of mobile terminals, a highly relevant topic in current and future wireless networks. Nevertheless, many other applications are
possible. The basic technique is illustrated in the context of multicast transmission, though cooperation is not limited to this example. The considered scenario is given in Figure 1.
Fig. 1. Scenario of investigation for cooperative power saving strategies. In figure are shown some cooperative groups (clusters).
Our analysis is based on the following assumptions: • a number of wireless terminals, interested in the same service, are distributed under the coverage of a central access point (AP). The AP provides a multicast transmission service and transmits according to the principle of Time Division Multiple Access (TDMA); • every terminal supports two air interfaces: a cellular link (C) for the AP reception, and a SR link for exchanging packets locally. In particular, this study is based on WLAN (Wireless Local Area Network) technology for both the cellular and the SR range communication; • we assume rate adaptation depending on channel quality as suggested by standards IEEE802.11a/g [5], [6]. Specifically, the cellular rate RC is set to the lowest specification of 6 Mbit/s, as the remote AP is assumed to be relatively far from the terminals. Instead, the shortrange rates RSR ∈ {54, 48, 36, 24, 18, 12, 9, 6} Mbit/s are dynamically assigned depending on the distance between the terminals; • the power levels P [W] used in the numerical analysis are chosen according to specifications of common presentday WNICs (Wireless Network Interface Cards) [7]. We also assume that the service can be split into substreams
according to a multicast coding scheme such as MDC (Multiple Description Coding, [8]) and focus on a given number J of terminals that are located in close proximity and form a cooperative cluster. We also assume, for simplicity and without loss of generality, that the number of transmitted substreams is also set to J. In order to get the best service quality, each user has to receive all the J substreams. We distinguish two possible operation modes: • Non Cooperative Operation (self-sustaining or autarky). The J terminals do not cooperate and receive all the J substreams with autonomous operation over the downlink channel. • Cooperative Operation (terminals cooperate with each other). Each unit receives only one out of the J substreams and forward it to the other J − 1 cooperative terminals over the short-range communication link (SR exchange). In the cooperative case, the power consumption of wireless terminals can be reduced by switching the receiving devices to low-power mode during idle periods. Furthermore, the communication over the SR link is more efficient, as the energy/bit required for the transmission is lower. An illustration of this technique is provided in Figure 2.
Thus, an estimation of the power consumption per terminal of a cooperative cluster PCoop can be easily obtained by replacing in (1) the power consumed for the reception Prx , transmission Ptx and idle Pidle by network devices. The power levels considered in this paper are motivated by measurements and reports of Atheros Communications about several products and technologies for WLAN [9], and listed in Table I. As it TABLE I P OWER L EVELS AND DATA R ATES FOR WLAN T ECHNOLOGY. Link C SR
Description Receiving power from central AP Power for central rx while idle Receiving power over short-range Transmitting power over short-range Power for short-range while idle
Name Prx,C Pi,C Prx,SR Ptx,SR Pi,SR
Value 900 40 900 2000 40
Unit mW mW mW mW mW
can be seen, a very low power is consumed by the devices during idle times, in comparison to the levels required for transmitting/receiving data. Newer technologies are expected to be even more energy efficient in this perspective. Assuming that the low-power mode can be exploited during idle times, the cooperative technique has been proved to be significantly beneficial, involving high power saving gains respect to the normal non cooperative operation (for further details, refer to [1]). Notice we are considering a framework where all the participants gain out of cooperation. II. M ODELING C OOPERATIVE S TRATEGIES : AGENT- BASED N UMERICAL A NALYSIS
Fig. 2.
Technique of cooperative reception.
Our target is to derive the power consumption of the cooperative scenario PCoop when the same service quality as in the non cooperative case is provided. This is given by: 1 1 PCoop (Z, J) = Prx,C + 1 − Pi,C + J J J −1 1 Ptx,SR + Prx,SR + + J J ·Z · Z 1 + 1− (1) Pi,SR Z where: Z is the ratio between the data rates used in the short≥ 1; range link and the cellular link: Z = RRSR C • J is the number of terminals in the cooperative cluster; • the power levels Prx , Ptx , Pidle are defined in Table I. •
In this section we describe how the power saving technique discussed so far has been tested by modeling cooperative strategies in a wireless network of mobile terminals. The strategies have been developed by using NetLogo multi-agent programming language [10]. An introduction to NetLogo for wireless communications, written by the author of this paper, is given in [11]. The language was originally conceived in order to provide a simple but powerful tool for modeling and exploring emergent phenomena, namely large-scale patterns that arise out of the complex interactions of numerous interdependent micro-”agents” at a lower level scale. Although it was not designed specifically for telecommunications, in principle a countless number of potential applications can be found, especially for modeling decentralized networks. The nodes of ad hoc networks, in fact, can be seen as collections of autonomous agents making decisions about transmit power, backoff time, packet forwarding and so on. In this case we focus on power consumption. In our NetLogo model, a number of wireless terminals are initially placed at random positions in a Virtual World, consisting of a simple square box and representing the coverage of the central AP. We rely on the assumptions given in Section I. Each terminal is assigned a label representing its instantaneous energy consumption (normalized to the unit in case of non cooperative operation). As the simulation is started, the nodes set the SR communication links whereas possible and start to cooperate. Then, they move around randomly one step and the
procedure is repeated. In this investigation, the movement of the units is performed by a simple random model; however, more advanced mobility models might be studied. The results of the numerical analysis can be monitored in real time by several charts and indicators. The full program code and a demo version are available on the web [12]. The cooperative interaction among wireless terminals is designed as a process of independent decision-making based on individual strategies. We motivate the introduction of strategies as the network topology is dynamically changing over time. Since cooperation is not beneficial in every situation, the actual payoff must be checked case by case. When the units do not gain anymore, they will stop to cooperate right away. For this purpose, each node is given a full knowledge about the power saving potentials of the cooperative technique described in Section I, by means of a payoff matrix. Such matrix, calculated from (1), reports the normalized energy consumed by one cooperating terminal, PCoop (Z, J), for all possible settings of Z and J. The values of the data rate ratio Z are derived in the assumptions of Section I, while J is restricted in the range {1, 2, 3, 4, 5, 6, 7, 8, 9, 10}. Each terminal is conceived as an autonomous agent having the main target in minimizing its power consumption. In this study, we model cooperative and reliable terminals. As the nodes are aware of the potentials of cooperation, we do not consider individuals that would not cooperate even though energy saving can be achieved. Furthermore, the analysis of malicious terminals (free-riders) that may try to harm the network operation is beyond the scope of this paper. Finally, in order to explore the potentials of cooperation, we also assume that the terminals have the capability to find all connected groups of units in current range (in a real system, of course this would require some signaling). In this way, each node is able to evaluate its best opportunity in terms of power saving gain of cooperation, by searching in the payoff matrix. The goal of our numerical analysis is to find out the best individual cooperative strategy: i.e., the one leading to minimize the power consumption of the overall network. The process of independent decision-making between nodes is based on a mechanism of negotiation. In turn, every node looks for the best cooperative cluster from his current position, according to his strategy. If an opportunity is found, it sends an invitation to the potential partners, by means of the cooperative payoff that could be achieved by cooperating in the new cluster. This data is derived from the payoff matrix discussed above, given the SR data rate and the number of cooperating terminals. Then, the new group is established if all the considered units give a positive response about the current opportunity, depending on individual strategies. In this way, the cooperative clusters are dynamically formed and dismissed, leading to system optimization. To implement the cooperative interaction among wireless nodes three different strategies have been developed, briefly summarized in Table II. The strategies are both rational and self-regarding. The primary goals are: to establish cooperative clusters in order to achieve energy saving (i.e., when
suitable conditions are met) and to preserve the gain of the individuals in situations of interdependent decision-making (interaction with the same/other strategies). For instance, the TABLE II S TRATEGIES FOR W IRELESS C OOPERATION . Selfish
Cooperate in a new cluster if the new individual gain is higher than the current.
Wise
Cooperate in a new cluster if the new average gain of the local subnetwork is higher than the current.
Safe-Wise
Play wise to wise units and selfish to selfish ones.
simplest arguable solution always cooperate is discarded, as the units would cooperate also in non profitable conditions. Thus, the first effective strategy in this scenario is called selfish cooperation. The strategy expresses the basic attitude of rational, self-regarding individuals who seek to maximize their own gain (i.e., minimizing their energy consumption). In our model, this is implemented by a simple comparison between the current gain of the unit and the potential gain that might be achieved by cooperating in a new cluster. As we will show later, this simple strategic profile is capable to accomplish the two main goals described above. However,
Fig. 3. Terminal T1 is looking for clusters and finds neighbors T2 and T3. What is the best strategy for cooperation?
the scenario under investigation is rather complex, due the terminal mobility and the heterogeneity of data rates used in the SR communication link. The best solution in terms of energy savings from the units point of view might not be the best solution from the system as a whole, as shown in Figure 3. Therefore we developed a more sophisticated strategy, called wise cooperation. In this strategy, we assume that the units are given a further knowledge concerning the actual power consumption of every other terminal in current range. In this way, nodes are able to evaluate the average power consumption of the local subnetwork. Then, a comparison is performed
between the power consumption of the current subnetwork and the new potential subnetwork (i.e., the new local network assuming that the cooperative cluster is made). A new cluster is established as far as the latter value is lower. In other words, the focus on minimizing the individual energy consumption (selfish) here is shifted to the attempt of minimizing the overall energy consumption of the local subnetwork. Notice that wise units may sacrifice their best individual profit for this purpose. We motivate this behavior as an attempt to establish a better cooperative framework that is more beneficial to all units in the long run. We will show that wise strategy achieves a better performance than selfish strategy, but has its drawbacks in a higher computational complexity and amount of required information. Furthermore, the wise strategy described above is effective only in the case where all the terminals feature the same strategy. The seek for local optimization fails in a heterogeneous scenario, due to the different behavior of other units (i.e., selfish nodes). Therefore we developed a last strategy called safe-wise cooperation. This strategy simply plays wise to the wise units and selfish to the selfish, thus reciprocating the behavior of the other units. We will show that safe-wise is still capable to achieve the optimum power saving gain of wise strategy also in presence of selfish terminals. Nevertheless, as the strategy relies on the additional cognition of the strategies of the other players, its implementation is even more complex than the one of wise strategy. To summarize, in our agent-based model the aforementioned negotiation process based individual strategies leads to the formation of cooperative clusters. Once the clusters are set, the units exploit the power saving technique of Section I to get energy saving. The cluster negotiation is repeated at every program iteration, as the units move around and the network conditions dynamically change. More details about modeling cooperative strategies can be found in [11]. III. P ERFORMANCE E VALUATION In this section we will discuss the major results of this investigation of cooperative power saving strategies. The output of our NetLogo model is formed by a main graphic window, showing the SR network of mobile terminals, and several monitors and charts displaying the performance figures [12]. The user can observe in real-time how cooperative clusters are formed, preserved and dismissed depending on circumstance. The istantaneous power consumption is shown by the labels of the units. The results are collected at every program iteration and an estimation of the performance figures is provided, in current and average values. Further details about the calculation of this data have been given in [7]. First of all, the selfish and wise strategy discussed in Section II have been tested separately in a homogeneous scenario of 50 mobile terminals. The resulting performance in the long run (> 10000 iterations) is shown by the following Table III. In ¯ = 1 − P¯ stands for the average power saving gain per table, G unit (where P¯ is the average normalized power consumption);
¯ OnlyCoop is the average gain per unit considering at each G step only the units involved in the cooperative exchange. TABLE III P ERCENTAGE POWER SAVING GAIN IN A HOMOGENEOUS SCENARIO .
Selfish Wise
¯ G
¯ OnlyCoop G
34.4 37.3
38.2 39.1
As it can be noticed, the results lead to considerable power saving gains, about 35 − 37% in percentage respect to the normal (autarchic) network operation. Other case scenarios, involving different technologies for the cellular and the SR communication, have shown to be even more promising in terms of energy saving potentials [7]. From the strategy perspective, a first outcome to be highlighted is that the simple selfish strategy provides an effective way to build-up cooperation. A second important result is that the wise strategy achieves the best performance in terms of power saving gain. Other results, which provide more insights on the profile of the two strategies, are beyond the scope of this paper. A second test of our model has been performed to analyze the interaction between selfish and wise strategy. In a heterogeneous scenario of 25 units for each strategy, it has been shown that the selfish can achieve a higher power saving gain, thus reversing the previous results arising out of the homogeneous scenario. The explanation is that the wise strategy was implicitly conceived to get the best cooperative performance in a world of wise individuals. If other terminals with a different behavior are operating, the strategy is not capable to take into account this fact. Hence, wise nodes are somehow exploited by selfish nodes, that take advantage of cooperating with wise units but at the same time make the usual ”elite” clusters among themselves. In order to address this issue, the initial model has been slightly modified by introducing an additional small box in the middle of the screen (two-box model). The purpose was to divide the Virtual World in two separate sides only occasionally interacting. The analysis of this model has suggested the way to improve the optimal strategy, wise, into safe-wise, as described in Section II.
Fig. 4. Comparison between the performance of wise and safe-wise strategy.
The simulation is based on two different scenarios: 1) the model is tested several times with different number of units NI ∈ {2, 4, 6, 8, 10} inside the box. The cooperative payoffs (normalized power) are calculated without any interaction with the external world; 2) the same tests are repeated by placing each time a constant number of selfish units outside the box, NO = 20. The payoffs of the internal/external network are derived. The numerical analysis has been performed in two steps: first by using wise strategy for the internal network, then safe-wise. A comparison between the normalized energy consumption of the two strategies is given Figure 4. The Sc1 solid curve (Pwise/saf e−wise ) represents the identical payoff of wise and safe-wise when no external units are present. In the first case, by using wise strategy, the curve shifts up as the units are exploited by selfish terminals and the Sc2 Sc1 > Pwise ). In the secpower consumption increases (Pwise ond test, by using safe-wise strategy, this gap almost disSc2 Sc1 appears (Psaf e−wise Psaf e−wise ). Moreover, the power consumption even decreases for small numbers of terminals Sc2 Sc1 (PSaf e−wise < Psaf e−wise , left side of the chart) as the units gain from the presence of the external nodes. This provides a numerical example of the meaning of the sentence - ”real egoistic behavior is to cooperate!” [1]. In conclusion, we show that safe-wise strategy is capable to achieve the best power saving gain out of cooperation and, at the same time, to preserve this gain when the units interact with selfish nodes. Notice that this comes at the price of increased complexity. In a real system, the higher amount of information needed by wise and safe-wise strategy would require additional signaling, and the implementation would result in a higher computational complexity at the mobile side. IV. C ONCLUSION In this section we draw some conclusions from our analysis. In this paper a novel cooperative power saving technique was introduced. The potentials of this idea have been explored through an agent-based simulation model of a mobile wireless network, relying on independent decision-making. For this purpose, three individual strategies have been developed. The numerical analysis has led to promising results, shown in Section III, in terms of network power saving gain. The power saving technique of Section I relies on a cooperative architecture, exploiting the synergy of different network platforms as envisioned in upcoming 4G. The technique is described for multicast, very suitable for recent video streaming applications, though is not limited to it. A major feature of the considered cooperative framework is the fact that all participants gain out of cooperation. The potentials of this basic technique have been tested by means of an agentbased simulation model of a wireless mobile network. The SR network has been implemented as a collection of nodes with no central authority, seeking to optimize their power consumption. The development of individual cooperative strategies has led to the following results. In the considered framework, the
decision-making process based on individual strategies leads to optimization of the overall systems. High power saving gain is achieved by all strategic profiles discussed in Section II. Concerning the strategies, we underline the existing tradeoff between efficiency and complexity (amount of information and computation). For instance, in concrete applications, the knowledge of the power consumption or the strategies of the other cooperating nodes would require additional SR signaling. The main challenges for the implementation in a real system are the followings. Measurements of the power consumption of a WLAN network interface card [7] have shown that the MAC of IEEE802.11 does not support low power mode during idle times, which is a fundamental assumption in our scheme. Furthermore in the dynamic establishment of the cooperative clusters, several SR issues have to be considered such as synchronization, interference, delay, and so on. Security is also an important matter to be studied. However the introduced technique leads to high power saving gains in the assumptions of Section I. Moreover, this study also underlines that the simple model of selfish cooperation might be preferable for testing the considered power saving technique in a real system. This is encouraged by the promising results achieved so far. However, additional analysis have to be performed taking into account more accurate models (i.e., for instance, considering the impact of the wireless fading channel). Other tasks arising from this study are the development of further models of cooperative power saving strategies based on individual adaptation/learning techniques, as well as models of cooperation targeting the optimization of different performance figures (such as throughput and spectral efficiency). R EFERENCES [1]
F.H.P. Fitzek, M. Katz. Cooperation in Wireless Networks: Principles and Applications, Springer, 2006, ISBN 1-4020-4710-X. [2] Axelrod, R. The Evolution of Cooperation, Basic Books, 1984. [3] A. MacKenzie, L. DaSilva. Game Theory for Wireless Engineers, Morgan & Claypool, 2006, ISBN 1-59829-016-9. [4] Frank H.P. Fitzek, Marcos Katz and Qi Zhang. Cellular Controlled Short-Range Communication for Cooperative P2P Networking, WWRF 17, 2006. [5] IEEE Std 802.11a Wireless LAN Medium Access Control (MAC) and the Physical Layer (PHY) specifications - High-speed Physical Layer in the 5 GHz Band, IEEE Standard for Information Technology, 1999. [6] IEEE Std 802.11g IEEE Std 802.11g-2003, Amendment to IEEE Std 802.11, 1999 Edn. (Reaff 2003) as amended by IEEE Stds 802.11a1999, 802.11b-1999, 802.11b-1999/Cor 1-2001, and 802.11d-2001. [7] Federico Albiero. Power Savings in Cooperative Networks. A Gametheoretic Approach, University of Padova (Italy) and Aalborg University (AAU - Denmark), 2006. [8] Frank H.P. Fitzek, Basak Can, Ramjee Prasad, Marcos Katz, DS Park. Traffic Analysis of Multiple Description Coding of Video Services over IP Networks, Center for TeleInFrastructure (CTIF), Aalborg University, 2004. [9] Atheros Communications Power Consumption and Energy Efficiency Comparison of WLAN products, white paper edition, 2003. [10] Various Authors. NetLogo User Manual, Center for Connected Learning and Computer-Based Modeling, Northwestern University of Evanston (IL), November 2005, http://ccl.northwestern.edu/netlogo/docs/. [11] F.H.P. Fitzek, M. Katz. Cognitive Wireless Networks: Concepts, Methodologies and Visions Inspiring the Age of Enlightenment of Wireless Communications, Springer, 2007, ISBN 978-1-4020-5978-0. [12] Federico Albiero. Wireless-Coop-Mobile, NetLogo Community Models, April 2006, http://ccl.northwestern.edu/netlogo/models/community/ Wireless-Coop-Mobile.
