睡眠模式在綠能無線網路之建模、設計、與最佳化

(1)

國立臺灣大學電機資訊學院電信工程學研究所博士論文

Graduate Institute of Communication Engineering College of Electrical Engineering and Computer Science

National Taiwan University Doctoral Dissertation

睡眠模式在綠能無線網路之建模、設計、與最佳化 Modeling, Design, and Optimization for Green Wireless

Networks with Sleep Mode Operations

張正義

Chen-Yi Chang

指導教授：許大山博士共同指導教授：廖婉君博士

中華民國 103 年 1 月

(2)

(3)

(4)

(5)

誌謝

在經過這麼多年的歲月，終於來到了即將畢業這一刻，在漫長的博士班生涯裡，經歷了許多挫折以及孤獨的時光，但也在這樣的磨鍊與思考的過程當中，讓我有所成長領悟，而在這當中，培養出的態度與毅力，也對於我未來的人生有著長遠的影響與幫助。感謝一路上大家的支持與鼓勵。

首先，感謝我的指導教授許大山教授多年來的指導，對於我在研究上的啟發以及研究上的品味有著深遠的影響，教授對於一個博士的期許，讓我能夠朝著一個明確的方向邁進。感謝我的共同指導教授廖婉君教授的指導，慷慨的提供在各方面協助，與我分享研究上和生活上的經驗，並且時常鼓勵我，讓我能夠保持研究上的熱忱。感謝謝宏昀教授，花費很多時間與我討論我的研究和論文，並在研究成果的呈現上和論文寫作上給予我很大幫助。感謝我的其他口試委員，蘇炫榮教授、蔡育仁教授、林永松教授、洪樂文教授、高榮鴻教授，接受我的邀請並且給予我的論文諸多寶貴的建議。感謝電信所多位教授的鼓勵，特別是陳光禎教授、

葉丙成教授、吳宗霖教授，時常勉勵我畢業就在不遠的前方。

另外，感謝電信所趙姐、欣梅、志豪、惠元、雅雯，多年來對於我特別的照顧與協助，忍受我的任性與嘮叨，讓我即使平常一個人在實驗室裡也不覺得孤單。

感謝電信所很多的學長與學弟妹們的陪伴，感謝 BL524 實驗室，特別是蔡隆盛學長、宇鵬、存志、冠錡、家榮、孝賢，感謝 BL608 實驗室，特別是張晉嘉學長、孫哥、昆錚、易翰、博瀚、柏翰、昆霖、昱均、秀秀、筑涵，感謝 BL503 實驗室，特別是劉顏慶、郭軒豪、蔡華龍、林典育，感謝當中許多人常常陪我在 118 巷、女九、活大之間做選擇，感謝與我一起做計畫的學弟，家榮、厚昇、炳成、勝順、景凱、易翰、博瀚、柏翰、昆霖、昱均，感謝與我一起做研究的學弟妹，昆霖、裕恆、小晴，以及感謝 EE530 實驗室很多的學弟妹時常陪我打球。

感謝我很多的好朋友時常約我出去散心與聊天，特別在我心情低落的時候，陪著我大吃大喝並且給予我很大的鼓勵。

最後感謝我的父母、我的妹妹們、我的家人們多年來的支持與鼓勵，在我當初決定要攻讀博士時，就對我的選擇無怨無悔的支持。在經過這麼多年的歲月，

(6)

(7)

摘要

最小化耗能是綠色節能無線通訊的一個基本目標，其中最有效且最有潛力的方法之一是根據網路流量來適應性的開關無線電收發器，這樣的想法可以被廣泛的運用在各種無線通訊設備或是裝置上。在本論文中，我們將探討睡眠模式在無線網路中的運作。我們對於一個網路中的節點功能性是可以互相被取代而且願意一起合作來達成整體目標的不自私(altruistic)網路感興趣，為了綠化傳統無時無刻都是主動運作的網路，我們利用睡眠模式所帶來的額外自由度(degrees of freedom)來節省網路的能量，使其成為一個在必要時刻才主動運作的網路。

在行動網路中，根據網路中的流量來開關基地台是一個節省網路能量的有效方式，然而這樣的運作可能會產生網路覆蓋漏洞(coverage holes)。在這樣的問題中，我們嘗試在保持整體網路覆蓋率的情況下，藉由動態的開關基地台來最小化整體網路的耗能。我們推導出在最小化每單位覆蓋面積的耗能下，每個基地台的最佳的覆蓋細胞大小(cell size)。在滿足網路覆蓋率的前提下，我們提出一個多項式時間複雜度的流量察覺(load-aware)的基地台開啟演算法。除此之外，我們展示了為了增加網路中熱點吞吐量(throughput)，藉由佈建的小細胞(small cell)達成的網路密致化(network densification)也可以在低流量的期間改善整體網路耗能。此外，我們探討異質網路(heterogeneous networks)下有可能的網路架構與多模 (multi-mode)基地台運作方式並且建議非對稱的基地台與移動裝置的連線將有潛力在低流量時大幅降低網路耗能。

在無線多重跳躍中繼網路(wireless multi-hop relay network)中，睡眠排程是一個有效的方法來降低網路耗能，但是通常會造成傳送訊息時的額外延遲。為了尋找在設計綠能多重跳躍中繼網路的洞見，我們建立了一個模型來分析和最佳化無線多重跳躍中繼網路在使用睡眠機制時，效能間的權衡(trade-offs)關係。進一步的說明，我們提出了一個隨機醒來的網路(random wakeup network)下使用投機式的路由(opportunistic routing)所建造的框架來分析。我們發現在最佳的參數設定時，整個網路用來偵測訊息的能量消耗占整體網路的^!!!

!!!，其中!是路徑損耗指數 (path loss exponent)，在最佳的參數設定下，我們發現整個網路用來偵測訊息的

(8)

路在最佳參數設定下的最小耗能和傳送速度的^!!!

!!!次方成正比。另一方面，我們研究在無線多重跳躍網路下訊息交換(information exchange)與訊息散播

(information dissemination)的耗能最小化。我們提出了一個隨機散播(random gossip)下週期性醒來的分析模型。我們發現在最佳的參數設定下，每個週期醒來廣播的節點數量只和路徑損耗指數和網路維度有關，這說明了不論增加或減少網路的大小都不會影響整個網路的最佳參數設定，我們得到最佳的參數設定為擁有可拓展性(scalability)的性質。另外我們藉由模擬的方式展示我們提出的模型可以被很好的運用在效能的朔模(performance modeling)與最佳參數的預測。除了跨層最佳化之外，我們提出的模組可以在無線多重跳躍中繼網路下，藉由感知 (cognition)與最佳化來巧妙的調整軟體可重組化(software-configurable)的功能用以達成整體網路的目標。

關鍵字：覆蓋漏洞(coverage holes)、異質網路(heterogeneous networks)、低流量(low traffic density)、多重跳躍中繼(multi-hop relay)、訊息偵測(message detection)、投機式路由(opportunistic routing)、耗能與延遲的權衡(energy-delay trade-offs)。

(9)

Abstract

Energy consumption minimization is a fundamental goal for green wireless communications. One of the most effective and promising way to save energy is to power off radio transceivers adaptively according to network traffic conditions. This idea can be utilized in a wide spectrum of wireless communication devices and equip- ments. In this dissertation, we explore the “sleep mode” operations in wireless networks. We are interested in altruistic networks in which nodes are interchange- able in functionalities and are willing to cooperatively achieve network level goals.

To “greenize” traditional networks from always-on networks to necessary-on networks, we exploit the additional degrees of freedom by sleep mode operations for network energy conservation.

In cellular networks, switching on/off base stations (BSs) according to network traffic profile is an effective way to conserve network energy. However, such operations may create coverage holes in the network. We attempt to minimize the total power consumption of the network by switching on and off BSs adaptively while maintaining the network coverage. We derive the optimal cell size for minimizing BS power consumption per unit coverage area and propose a polynomial-time load- aware algorithm for energy-efficient BS activation while avoiding creating coverage holes. Besides, we demonstrate that network densification with small cells for bursting throughput in hot spot areas can also improve network energy savings during the low traffic load period. Furthermore, we explore potential network structures with multi-mode BS operations in heterogeneous networks (HetNets) and suggest that asymmetric BS-MS (mobile stations) connections with uplink by small cells and downlink by macro cells will be energy-efficient during the low traffic load periods for green HetNets.

For wireless multi-hop relay networks, while applying sleep-awake scheduling is an effective way to reduce network energy consumption, it usually comes at a price of additional delay for message delivery. To seek insights for the design of green wireless multi-hop relay networks, we develop a model for the analysis and optimization of performance trade-offs in wireless networks while sleep-awake mechanisms are applied. Specifically, we propose a random wakeup network with opportunistic relay as a framework for our analysis. Under optimal settings, we find that the proportion of power for message detection in the entire network is ^α−1_α+2, where α is the path loss exponent; the energy consumed for message detection should be of the same order as that for actual communications. Moreover, we find that the minimal network power consumption under optimal operations grows at ^α−1_α+2-th order of the delivery speed, meaning that the investment of network power can efficiently reduce delivery delay. Besides, we investigate network energy minimization for information exchange and information dissemination in wireless multi-hop relay networks. We propose

(10)

We find simple rules govern the optimal settings for network-wide information exchange in the random gossip network. One key relationship is that the optimal number of nodes broadcasting messages in a time epoch within the transmission range depends only on the path loss exponent and the network dimensionality. It is shown that neither increasing nor decreasing the physical network size will affect the optimal value of these design parameters. The optimal setting for information exchange is a scalable solution. We show through simulation results that our proposed frameworks are well applicable for performance modeling and parameter optimization in wireless multi-hop relay networks when sleep-awake operations are adopted.

Beyond cross-layer optimizations, our proposed frameworks can facilitate the cognition and optimization for ingeniously adapting software-configurable functionalities to achieve network goals in wireless multi-hop relay networks.

Keywords– coverage holes, heterogeneous networks, low traffic density, multi- hop relay, message detection, opportunistic routing, energy-delay trade-offs.

(11)

List of Figures

1-1 The vision of wireless networks in the future. To increase network capacity and support diverse applications, a large amount of equipment is expected to be deployed in the networks. Extensive network densification (e.g., BSs, relays, antenna) is expected in future wireless networks to support advanced wireless technologies. . . 12

2-1 Illustration of the coverage hole problems due to switching on/off BSs.

The rectangular region (shaded) is the targeted covering region. Five BSs are deployed in the region. The disk centered at a BS is the coverage of the BS. (a) Coverage hole present when a BS is switched off. (b) A BS activation strategy with high power level. The coverage of target area is ensured. (c) A BS activation strategy with low power level. The coverage of target area is ensured. . . 18 2-2 The relative positions between an uncovered point and the active BS

closest to the uncovered point. (a)The regional classification of an uncovered point. (b)The position of an uncovered point in region I.

(c)The position of an uncovered point in the region II. . . 30 2-3 The coverage of activated BSs by COMIC to cover a 10-by-10 planar

network. BSs are deployed in a grid structure from position (0.5,0.5) to (9.5,9.5). The coverage range R = 1.8. Only 22 out of 100 BSs are activated by COMIC to cover the network. . . 36

(16)

2-4 The minimal number of active BSs as a function of coverage range to cover a 30-by-30 planar network. Three BS deployment strategies are examined (hexagonal, grid, and random). . . 37 2-5 The activated BSs (green) by COMIC with a hot spot area at the

center in the network. Only 56 out of 368 BSs (gray) are activated.

In the network, 5000 users (blue) are uniformly distributed in the hot spot region and 8000 users (blue) are uniformly distributed in the other region. . . 47 2-6 The minimal network power consumption as a function of average user

data rate. The static part of network power consumption is also shown.

The performance is compared with the algorithm by Zhou et al. . . . 48 2-7 The minimal network power consumption for maintaining network cov-

erage as a function of the number of deployed small cells in the network. 49 2-8 The average number of activated BSs for minimizing network power

consumption while maintaining network coverage as a function of the number of deployed small cells in the network. . . 50 2-9 The network power consumption of grid deployment of 900 BSs for

different power saving strateiges. Both the simulation results and the lower bounds of network power consumption are shown. . . 60 2-10 The network power consumption of hexagonal deployment of 368 BSs

for different power saving strateiges. Both the simulation results and the lower bounds of network power consumption are shown. . . 61 2-11 The network power consumption as a function of the constant oper-

ational power of a small cell. Various network structures in HeNets are examined with 368 hexagonally deployed macro-cell BSs and 4000 randomly deployed small-cell BSs. . . 62 2-12 The network power consumption as a function of the constant opera-

tional power of a small cell. Various network structures in HeNets are examined with 368 hexagonally deployed macro-cell BSs and 40000 randomly deployed small-cell BSs. . . 63

(17)

3-1 An example of a message traversing from a source node to the sink node in the RWN. A message originated from a source node is to be delivered to the sink node by multi-hop transmissions. In epoch k, some nodes (white nodes) wake up to listen to potential incoming traffic. If node rk

currently relays a message to sink node s, the effective neighborhood of node rk,N^rk,s, is the shaded region. Awake nodes (two white nodes) in the effective neighborhood of node rk serve as candidates for the next forwarding node. The wakeup node in the effective neighborhood of node rk closest to the sink node will be selected as the next forwarding node. . . 69 3-2 An illustration of the sleep-awake mechanism in the random wakeup

network. A node without a message to transmit wakes up randomly in a duty-cycle period. If it does not detect any message during the listening interval, it will go back to sleep mode immediately. . . 71 3-3 An example of node’s behavior in a duty-cycle period. A duty-cycle

period can be divided into four intervals: (a) listening interval (b) data transmission-reception interval (c) route negotiation interval (d) sleep interval. These four intervals may be reordered for different scenarios. 72 3-4 An illustration of the effective neighborhood of a transmitter. (a) The

effective neighborhood N^rk,s of transmitter rk in the RWN in which the separation between the transmitter rk and sink node s is Lr_k,s. (b) N^rk,s(x) is the region where a message can advance a distance more than x toward sink node s from rk. (c) An upper bound of N^rk,s. (d) A lower bound of N^rk,s. (e) An approximation for N^rk,s. . . 78 3-5 The expected forwarding distance in a time epoch as a function of the

wakeup density of nodes with transmission range D = 1 and distance toward the sink node Lrk,s = 5. . . 85 3-6 Optimal number of nodes awakened to forward messages in a time

epoch within the transmission range as a function of response-to-transmit energy ratio E_c¹/E_t¹. . . 89

(18)

3-7 The optimal number of nodes awaken to forward messages in a time epoch within the transmission range per dimension as a function of response-to-transmission energy ratio E_c¹/E_t¹. . . 97 3-8 The optimal number of nodes awaken to participate the relay activities

in a time epoch within the given n-dimensional cube Dⁿ as a function of the key engineering parameter β ≡ EtDⁿ/(_E[L^T^m

i,s]T1V Ed+ EcDⁿ). . 102 3-9 Network power consumption and delivery delay as a function of the

wakeup probability of a node. Both the simulation results and analytical results are shown. . . 106 3-10 Optimal wakeup probability for minimizing network power consump-

tion as a function of mean delivery delay time. Simulation results as well as analytical results are shown. . . 107 3-11 Minimal power consumption under the jointly optimal values of the

wakeup probability, transmission range, and sleep period as a function of the mean delivery delay constraint. . . 108 3-12 Optimal value of the wakeup probability and the maximal network

lifetime as a function of mean delivery delay time. Energy parameters Et and Ec are normalized to 1 for transmission range D = 100 m.

Initial battery energy of a node is set to 10⁵. . . 109 3-13 The optimal number of tree nodes as a function of mean delivery speed

in the EVBT network with sleep mechanism adopted. α11 = 12.8 µJ/packet. α12 = 12.8 µJ/packet. α2 = 16nJ/packet/ m². . . 112

4-1 One-dimensional network with unit-length spaced nodes. . . 121 4-2 The optimal number of nodes gossiping within an n-D cube of size D^∗

in a time epoch. Both the optimal values and the lower bounds for 1-D 2-D and 3-D networks are shown. . . 125

(19)

4-3 The optimal gossip probability as a function of the average delay time for information dissemination. The delay time is normalized by the message origination period. Both the simulation results and the corresponding optimal parameter in the analogous RGN are shown. The path loss exponent α = 3 is given. Different average reception probability requirements are examined. . . 129 4-4 The optimal gossip probability as a function of the average delay time

for information exchange. The delay time is normalized by the message origination period. Both the simulation results and the corresponding optimal parameter in the analogous RGN are shown. The path loss exponent α = 4 and Ed/E_t¹ = 10⁴ are given. Different data fusion models are examined. . . 130

(20)

THIS PAGE INTENTIONALLY LEFT BLANK

(21)

Chapter 1 Introduction

The evolution of wireless communication networks brings us fold-by-fold throughput increase last decades. To meet explosive growth of data volume requests in the future, potential directions for bursting capacity in next generation wireless communication networks can be peeked from pioneers. Since spatial frequency reuse serves as a simple and effective design technique for bursting network capacity [1], small-size cells are expected to densely deployed to further drain the capacity from frequency reuse.

Recently, massive MIMO is proposed to exploit the fertile channel diversification in wireless environments [2, 3] by deploying a large excess of antenna at BS side. By utilizing the property that channel vectors of different users will tend to be more orthogonal as the increase of antenna, massive MIMO is expected to scale up cell capacity 10 times or more. Besides, it is known that the network capacity will increase but per node capacity will decrease as the increase of nodes in the network [4]. Putting dumb nodes in the network with multi-hop relaying can also enjoy extra capacity gain.

In conclusion, no matter what kind of approaches will be adopted, extensive network densification (e.g., BSs, relays, antenna) is expected in the future. An illustration of our vision for wireless networks in the future is given in Figure 1-1.

It is recently pointed out that carbon emission and energy consumption of wireless networks (10-20% compounded annual growth rate (CAGR)) reach a level which can not be ignored [5, 6]. If we are not aware for what is paid for the development of advanced wireless technologies, the potentially exponential increase of electricity

(22)

Figure 1-1: The vision of wireless networks in the future. To increase network capacity and support diverse applications, a large amount of equipment is expected to be deployed in the networks. Extensive network densification (e.g., BSs, relays, antenna) is expected in future wireless networks to support advanced wireless technologies.

consumption will exacerbate the energy crisis and global warming. “Greenizing” now becomes a key design objective to mitigate energy crisis and suppress carbon emissions in wireless networks. From economic perspectives, the growth of data volume does not directly reflect to operator’s revenue. The data rate demands will grow with 92% CAGR in 2012-2015 (estimated by Cisco [7]) but revenue of network operators is growing only at 13.4% in 2013 (estimated by ABI Research). On the other hand, reducing energy cost will directly contribute to the reduction of network operational cost. In view of the issues discussed above, we believe that how to greenize the ever dense wireless networks will be important and challenging in the near future.

Network traffic variations provide great opportunities to save network energy dynamically. Traditionally, most of network designers and operators focus on how to plan and maintain a network in peak load conditions. However, a network is not always at such high load conditions. Based on the measurement results in wireless networks, the network energy consumption without adapting to network traffic loading causes a lot of energy waste. With more diversification of wireless network applications (e.g., smart metering versus 3D video streaming) in the future, traffic

(23)

demands is expected to vary more dramatically in various domain (e.g., space, time, frequency, and application). While a network shifts from high traffic load conditions to low traffic load conditions, we believe lots of network energy can be saved.

The aggregated power consumption for detecting network traffic may easily sur- pass the power consumption for actual transmissions in a network with low traffic load density. An intuitive approach to conserve energy in low traffic density network is to put “idle” nodes into so called “sleep mode”. The term “sleep” generally refers to the practice where a communication terminal disengages itself for a period of time from the network, during which the transceiver is shut off to greatly reduce the power consumption. With the inclusion of the sleep state operation, in this dissertation, we desire to seek for general design principles and capture key network characteristics for green wireless networks while sleep-awake scheduling is adopted in the networks.

1.1 Motivation

The rapid exacerbation of energy crisis and the dramatic rise of carbon emissions caused by wanton abuse of energy resources gather our attention to green technologies.

In wireless networks, the explosive access demands results in the exponential growth of energy consumption during last decades. Technology trends toward extensive network densification reveal that network energy consumption will continuously expand in the future. Nowadays most of human life is inseparable of wireless networks. The traffic variations resulting from human daily routine and diverse applications in wireless networks motivate us to investigate the intuitive “sleep mode” operation for energy conservation in wireless networks.

1.2 Summary of Contributions

The high level contribution of this dissertation is to provide general design principles of sleep mode operations for green wireless networks. We build analytical frameworks with sleep mode operations for quantifying network performance trade-offs and seek-

(24)

ing key properties in an optimized network. Applying the obtained design insights, we propose energy-efficient algorithms with sleep mode operations and suggest promising network structure for network energy conservation. Furthermore, we demonstrate our results can be applicable for the optimization of software-configurable functionalities in various wireless networks with different applications. Our proposed frameworks are expected to facilitate the cognitive processes in wireless relay networks while satisfying network level goals [8]. The specific contributions of this dissertation are highlighted as follows.

1.2.1 Sleep Mode Operations for Green Cellular Networks

To conserve network power in cellular networks, a promising approach is switching off BSs at low traffic load. However, the switching-off operations may induce network coverage holes. We attempt to minimize the total power consumption of the network by switching on and off BSs adaptively while maintaining the network coverage. We find that BS activation problem for minimal network power consumption with full network coverage preservation is an NP-hard problem. We derive the optimal cell size for minimizing area power consumption and and propose a polynomial-time load-aware algorithm for energy-efficient BS activation in HetNets. The simulation results show that our algorithm can approach the minimal network power consumption. Besides, we demonstrate that network densification by small cells for bursting throughput in hot spot areas can also be beneficial in saving network energy during the low traffic load period. Finally, we explore potential network structures with multi-mode BS operations in HetNets and suggest that uplink by small cells and downlink by macro cells will be promising for network energy saving during low traffic load periods.

1.2.2 Sleep Mode Operations for Green Wireless Multi-Hop Relay Networks

Conservation of network energy is a critical design goal for wireless communications comprised of battery-powered devices. While applying sleep-awake scheduling is an

(25)

effective way to reduce network energy consumption, it usually comes at a price of additional delay for message delivery. For message delivery in wireless relay networks, we develop a model for the analysis and optimization of performance trade-offs in wireless networks while sleep-awake scheduling mechanisms are applied. Specifically, we propose a random wakeup network with opportunistic relay as a framework for our analysis. In the proposed framework, nodes wake up randomly to detect potential incoming messages in a duty-cycle period, and messages are opportunistically relayed by these wakeup nodes. We first derive the closed-form expressions for network power consumption in the random wakeup network, and then we apply the model in an optimization problem. The goal is to minimize the total network energy consumption with a requirement on the end-to-end delay by jointly selecting key design parameters including transmission range, duty-cycle period, and wakeup density. We find that for a network that employs the optimal routing and sleep parameters thus obtained, the proportion of power for message detection in the entire network is ^α−1_α+2, where α is the path loss exponent. That is, under the optimal settings, the energy consumed for message detection should be of the same order as the energy consumed for actual communications. Moreover, we find that the minimal network power consumption under optimal operations grows at a rate of the ^α−1_α+2-th order of the message delivery speed, meaning that the investment of network power can efficiently reduce message delivery delay. We show through simulation results the proposed framework is well applicable for performance modelling and operational parameter optimization in wireless relay networks when sleep mode operation is adopted.

For information exchange and information dissemination in green wireless relay networks, we propose a “random gossip” wireless-relay network (RGN) as a framework for our analysis. In the RGN, we consider that each node initiates a message and wishes to exchange its message with all the other nodes in the network. We will demonstrate that our results in the RGN can indeed predict the optimal value of operational parameters for information exchange as well as information dissemination in actual WSNs. We find key properties for the three design parameters in the optimized RGNs. First, a key relationship is that the optimal number of nodes

(26)

broadcast messages within the transmission range in a sleep period depends only on the path loss exponent and the network dimensionality. It does not depend on factors such as the physical network size, the reception mode energy, and the message origination rate. Next, the existing optimal number of broadcasting nodes in a time epoch within the transmission range shows that either shortest hop transmission or blind flooding is rarely optimal for power consumption minimization. Last but not the least, it is shown that neither increasing nor decreasing the physical network size will affect the optimal value of these design parameters. The optimal setting for network power minimization is a scalable solution.

1.3 Outline

The rest of this thesis is organized as follows. In Chapter 2, we explore the sleep mode operation in cellular networks. We investigate how to switch off BSs in an energy-efficient manner with the consideration of network traffic variation and network coverage preservation. We formulate the green network coverage preservation problems and propose energy-efficient BS activation algorithms in HetNets. Besides, we evaluate potential network structures with multi-mode operations of BSs for network energy saving. In Chapter 3 and Chapter 4, we explore the sleep mode operations in wireless relay networks. We investigate the minimization of network energy consumption while satisfying network level goals. In Chapter 3, we investigate message delivery with sleep mode operation for green wireless relay networks. We propose the random wakeup model with opportunistic routing and demonstrate our derived results can be well applicable for message delivery in many proposed wireless relay networks. In Chapter 4, we investigate information exchange with sleep mode operations in green wireless relay networks. We propose the random gossip network with periodic listening and demonstrate that the derived results can be applicable for both information exchange and information dissemination in wireless relay networks.

Finally, concluding remarks are given in Chapter 5.

(27)

Chapter 2 Sleep Mode Operations for Green Cellular Networks

Energy consumption minimization is a fundamental goal for green communications [6].

It is estimated that the cost spent in electricity globally for mobile networks is more than $10 billion dollars annually, and among all the network components, radio base stations (BSs) may consume over 80% of the energy in the network [9]. Therefore, it is very important for mobile operators to design energy-efficient mechanisms for BSs.

One effective way for energy saving in cellular access network is to turn on and off BSs adaptively. The daily traffic profile in [9] shows that the long-term behavior of BSs is predictable (e.g., day versus night), and for almost half of a day, the traffic load of a BS is low. However, a BS with low or no traffic load still consumes more than 90% of its peak power. As a result, switching off the cells which are under low traffic condition is one of the most effective ways for energy saving. A similar approach is also taken in the standards of next generation wireless networks (e.g., 3GPP LTE-Advanced), in which BSs are switched off to conserve energy [10].

Energy conservation by switching off BSs may encounter several challenges. An important problem, identified in many studies [9,11–15], is the traffic sharing problem induced by the switched off BSs. Users in the switched off cells shall be handed over to other nearby cells. The work in [9] points out that energy can be saved by adaptively switching on and off BSs with the consideration of daily traffic fluctuations. Based on

(28)

(a) (b) (c)

Figure 2-1: Illustration of the coverage hole problems due to switching on/off BSs. The rectangular region (shaded) is the targeted covering region. Five BSs are deployed in the region. The disk centered at a BS is the coverage of the BS. (a) Coverage hole present when a BS is switched off. (b) A BS activation strategy with high power level. The coverage of target area is ensured. (c) A BS activation strategy with low power level. The coverage of target area is ensured.

reassociating users among BSs with the available location information, the authors in [13] try to find a minimum set of active BSs with which users can be associated.

In [14], the authors propose centralized and decentralized energy-efficient BS on-off strategies according to network traffic variations and show the trade-off between outage performance and energy efficiency. Taking the dynamics of user population into consideration, the authors in [15] propose an energy-efficient BS operation mechanism with user association policies for the trade-off between network energy saving and flow-level performance. According to the variations of daily traffic demand and inhomogeneous user population, these studies show that load balancing and user association by switching on/off BSs can significantly reduce the network energy consumption. However, these studies all assume that the network coverage will not be affected by switching off BSs, which may not always be the case as indicated in [10]

(see Fig. 1 (a) for example). The work in [16] shows that the coverage maintenance for mobile cellular networks is a key challenge in energy-saving operations of BSs. By turning off a BS in the low traffic condition to save energy, the cells in the neighborhood of shut-down cells may need to extend their coverage (by increasing their power levels) so as to avoid creating coverage holes (see Figs. 1 (b) and (c) for example).

However, in the literature, this coverage hole problem induced by the power-off BSs is rarely investigated.

(29)

The coverage problem has been treated in great details in wireless sensor networks [17, 18]. Many research studies (e.g., [19]) show that to find the minimum number of nodes to cover an area is often an NP-hard problem. In such networks, the key design objective is to maximize the lifetime of the network by uniformly distribut- ing the residual energy in the whole network. A decentralized node scheduling scheme with limited local information is the typical technique used for practical applications (e.g., Optimal Geographical Density Control (OGDC) in [20]). However, in wireless sensor networks, the coverage range may not be adjustable and full network coverage may not be ensured (e.g., only coverage for nodes without mobility). In cellular networks, we must avoid the signal loss of users who turn on at uncovered region or who move through coverage holes. Therefore, existing solutions for wireless sensor networks may not be directly applicable for green cellular networks. Although the coverage and traffic rate requirements can be satisfied in traditional network planning problems, the goal of most network planning problems is to reduce the network deployment cost at peak load conditions instead of saving network power while network traffic varies. Most recent works are focused on energy-efficient network planning for coverage extension via BS cooperations (e.g., in [21]). However, the potential benefits of dynamic power control for cell breathing are still not explored. On the other hand, to maintain network coverage when traffic varies, pre-configured BS on/off patterns for hexagonal BS topologies are analyzed to show the potential energy saving capa- bilities by joint power control and cell activation in [22]. However, these analytical results may not be easily applicable to heterogeneous networks (HetNets).

HetNets with high capacity small cells [23, 24] (e.g., picocells or femtocells) are promising for bursting throughput at hotspot area in next generation cellular systems. To meet the future demands on data access rates (e.g., Cisco [7] estimates that Compounded Annual Growth Rate (CAGR) of network traffic reaches 92% in 2010-2015), Nokia Siemens Networks estimates that the number of BSs will increase by 10-fold in 2020 [25]. Thus, we can reasonably expect that a large number of small cells will be deployed to cooperatively provide service with macro cells in the future.

In addition to increasing network capacity, these small cells may also improve the

(30)

network power savings during the low load periods by maintaining the coverage of certain small areas.

The diversification of BSs in HetNets (e.g., macro BSs, femtocells, small cells, etc.) will bring new opportunities and challenges in the traditional only macro cell scenario.

The idea of providing exactly the same functionality of various BSs all the time may need to be re-examined. Recently, many proposed ideas shift the design paradigms such as coverage extension for mitigating asymmetric transmit power between macro cells and small cells [26], dual connectivity between macro cells and small cell to utilize the inherent asymmetric uplink and downlink coverage in HetNets [27], and umbrella cells to alleviate handover of mobility users [24]. Furthermore, control signalling reduction via the functionality separation of BSs is proposed in [28] to increase the network energy efficiency. The network diversification in conjunction with ”flexible BS operations” will be promising in the future.

In view of the issues discussed above, in this chapter, we will explore the green network coverage problem in HetNets, in which we attempt to minimize the total energy consumption in cellular systems by joint BS activation and power control given the requirements that the network coverage and traffic load must be ensured. To follow the shifted design paradigms which break the traditional one-to-one association between users and BSs, we propose the ”reception mode” design and demonstrate the potential gain of asymmetric uplink and downlink connections with the support of multi-mode operation of BSs for green HetNets. The major contributions are summarized as follows.

• We formulate the mathematical green network coverage problem in cellular networks. The switching on/off operation and power control of BSs are explored.

We decompose the green network coverage problem into two sub-problems:

BS operational power optimization problem and minimum-power BS activation problem.

• We derive the optimal power and optimal cell size of each active BS for minimizing BS power consumption per unit coverage area. We also derive a lower

(31)

bound for the minimal power consumption of the entire network. This bound can be jointly used with any BS deployment strategy.

• We show that the minimum-power BS activation problem is NP-hard and propose a polynomial-time load-aware algorithm for the activation of BSs while avoiding creating coverage holes in the network. We demonstrate that the performance of the proposed mechanism is indeed a very efficient power saving mechanism in various BS deployment strategies via extensive simulations. Com- pared with the proposed algorithm in [14], 44.5% more power will be consumed by their approach during the low traffic load period.

• We propose the ”reception mode” design of BSs in conjunction with asymmetric MS-BS uplink and downlink connections to further explore the opportunities of network energy saving. We suggest that uplink by small cells and downlink by macro cells is promising for network energy saving during the low traffic load period for green HetNets. Compared with traditional symmetric connection by macro cells, this approach can save network power more than 6 dB in HetNets by simulations.

2.1 System Model and Problem Description

2.1.1 BS Operation Model

We consider a set of BSs, denoted by B = {B1, B2, . . . , BM}, in a cellular network.

We assume that the set of BSs is deployed in such a way that the transmission rate requirements induced by the service level agreement (SLA) to mobile stations (MSs) are satisfied. This assumption is reasonable as mobile operators deploy their BSs to cover the service area so as to satisfy the SLAs of their MSs.

In the network, each BS can be switched on/off according to the network condition.

A BS which serves users in the network is referred to as an active BS. We assume that the long-term behavior of traffic in the network is predictable, and the daily traffic profiles in different BSs are highly correlated (e.g., BSs tend to have low traffic

(32)

load at midnight). Therefore, the focus of this chapter will be on the optimization of energy saving during off-peak hours.

Note that the switching-on/off operation of a BS we consider here is based on the average traffic load (e.g., in each hour). It is different from the sleep mode operation of radio transceivers which may be turned on and off more dynamically according to the load of instantaneous traffic.

2.1.2 BS Power Consumption Model

The studies in [29, 30] show that the power consumption in a cellular BS (e.g., GSM, WCDMA, LTE, etc.) can be modelled by a simple but surprisingly accurate model:

Pm = η_m⁻¹Pt,m+ Pc,m+ η_m⁻¹X

k

Pk,m, (2.1)

where Pm represents the total power consumption of Bm, Pt,m denotes the maximal transmit power of Bm, ηm and Pc,m are the parameters determined by the system hardware and radio access techniques, and Pk,m stands for the link level power consumption from Bm to MS k. Specifically, ηm depends highly on the efficiency of the power amplifier, Pc,m models the power consumption which is independent of the radiation power such as the power consumed in the cooling system, and P

kPk,m

models the aggregated MS-BS link power consumption which depends on the number of active MSs and traffic loads. In this chapter, we call the parameter ηm as the transmit power efficiency and the parameter Pc,m as the constant operational power of a BS. In general, the constant operational power is a large constant which is non- negligible [11, 12]. Based on the measurement results of deployed BSs, it is showed that the power consumption of a BS with low load is almost the same as the power consumption of a BS with peak load [29] (e.g., the power consumption varies about 3% in a UMTS BS and about 2% in a GSM BS during a period of several days). For simplification, we omit the link level power consumption Pk,m in our analysis. We will show that this simplification is reasonable via simulations in Section 2.3.4.

(33)

2.1.3 Channel Model

The propagation channel depends on the physical environment. Here we consider a channel model with path loss and shadowing effects [31]. Specifically, the receive power of an MS, denoted by Pr, is modelled by

Pr = KΨd^−αPtx, (2.2)

where Ptx represents the transmit power, d is the separation between the BS and the MS, α is the path loss exponent, K is used to model the impacts from antenna heights, carrier frequency, and antenna patterns [32], and Ψ models the shadowing effects in the environment. Note that Ψ is a random variable with a log normal distribution.

Therefore, the random variable 10 log₁₀(Ψ) follows the normal distribution with zero mean and standard deviation σΨ. Here distance d is a normalized value with respect to a reference distance [31].

2.1.4 BS Signal Coverage Model

The coverage of a BS, denoted by C^m, shall satisfy the received power requirements for both downlink and uplink direction as follows.

P Pr^d(x, y) < P_min^d ≤ p^dout,

∀

^{(x, y)}^{∈ C}^m^, ^(2.3)

P (Prû(x, y) < P_minû )≤ pûout,

∀

^{(x, y)}^{∈ C}^m^, ^(2.4)

where P(·) represents the probability function, Pr^d(x, y) denotes the received power of an MS at location (x, y) from Bm, P_min^d is the minimum required received power of an MS determined by receiver sensitivity [33], and p^d_out is the outage probability of the received signal power of an MS. On the other hand, P_rû(x, y) denotes the received power of Bm from an MS at location (x, y), P_minû is the minimum required received power of a BS, and pû_out is the outage probability of the received signal power of a BS.

(34)

Note that the area of C^m is denoted by ACm. For example, P_min^d can be computed by

P_min^d (dBm) = N0B(dBm) + SINR(dB) + NF(dB) + IM(dB), (2.5)

where N0 is the thermal noise power density, B is the specified noise bandwidth, SINR is the requirement on signal to interference plus noise ratio for a specific coding and modulation scheme, NF is a measure of loss of SINR caused by RF components, IM is the implementation margin to account the SINR differences between practice and theory. To handle the potential interference caused by the switching-on/off operations, we set an interference margin to mitigate the interference from neighboring cells in the SINR requirement [31,33,34]. In LTE, the thermal noise density is specified as -174dBm/Hz, NF is specified as 9dB, SINR and IM depends on specific modulation and coding schemes [33].

2.1.5 Green Network Coverage Problem

The design objective of green network coverage problem is to find the set of active BSs such that the total power consumption of all BSs in the network is minimized while the network-wide coverage is still maintained. Without loss of generality, we assume that each BS is either in an ON (i.e., active) state or in an OFF (i.e., power- off) state during which there is zero power consumption. Denote the state of Bm by indicator variable Im (“ Im = 1” means ON, and “Im = 0” means OFF). To switch off low-load BSs, both downlink coverage and uplink coverage must be ensured by other neighboring active BSs. Also, the network traffic load requirements shall still be satisfied after the BS switching-off processes. An optimization problem for green network coverage problem can be formulated as:

min

M

X

m=1

Im η_m⁻¹Pt,m+ Pc,m

(2.6)

subject to (2.7)− (2.15).

(35)

We specifically explain the requirements and constraints from (2.7) to (2.15) as follows.

A ⊆ [

Bm∈B

C^m. (2.7)

In (2.7), C^m is the region covered by Bm, and A is the target coverage region. With the two requirements, an MS must be covered by at least one active BS in the target coverage area. The coverage of a BS shall satisfy the minimum receive signal requirements below.

P(Pr^d(x, y)Im < P_min^d )≤ p^dout,

∀

^{(x, y)}^{∈ C}^m^,

∀

^m ∈ {1, . . . , M}, (2.8) P(Prû(x, y)Im < P_minû )≤ pûout,

∀

^{(x, y)}^{∈ C}^m^,

∀

^m∈ {1, . . . , M}. (2.9) In (2.8) and (2.9), the received power requirement for the downlink and uplink coverage of Bm at a point (x, y) is given, respectively. Note that no location will be covered by Bm (i.e.,C^m =∅) if B^m is at off state (i.e., Im = 0). To satisfy the downlink traffic requirement in the coverage ofC^m, the bandwidth constraint and the transmit power constraint are given in (2.10) and (2.11), respectively.

X

∀(xk,y_k)∈Cm

w_k,m^d ≤ (1 − δm^d)W_m^d,

∀

^m∈ {1, . . . , M}, (2.10)

X

∀(xk,y_k)∈Cm

Pt,m,k ≤ (1 − δm^p)Pt,m,

∀

^m ∈ {1, . . . , M}. (2.11)

In the two equations, W_m^d is the total available bandwidth of Bm for downlink, w^d_k,m is the required bandwidth of Bm for MS k, Pt,m,k is the required transmit power of Bm for MS k, and (xk, yk) is the position of MS k. Note that δ_m^d and δ_m^u are the fraction of downlink and uplink bandwidth which is not used for data transmission (e.g., control signalling, pilots). Similarly, δ_m^p is the fraction of power not used for

(36)

data transmission.

X

∀(x_k,y_k)∈Cm

w_k,mû ≤ (1 − δmû)W_mû,

∀

^m∈ {1, . . . , M}. (2.12)

Similarly, to satisfy the uplink traffic requirements, the aggregated bandwidth allo- cated for uplink MS k by Bm, denoted as w^u_k,m, is constrained by the total available uplink bandwidth W_m^u in (2.12). Besides, the maximum transmit power constraints and the BS operation state constraints are given as follows.

0≤ P^t,m ≤ I^mP_t,m^max,

∀

^m∈ {1, . . . , M}, (2.13) 0≤ P^t,k ≤ P^t,max,

∀

^k∈ {1, . . . , K}, (2.14) Im ∈ {0, 1},

∀

^m ∈ {1, . . . , M}. (2.15)

In (2.13), P_t,m^max is the constraint on the maximum transmit power of Bm due to the radiation power regulations or hardware limits. Note that the maximal transmit power of Bm is equal to zero if Bm is switched off (i.e., Im = 0). In (2.14), Pt,max

stands for the maximum transmit power constraint of MS k.

Depending on underlying communication systems and network environments, BS coverage may be uplink dominated (e.g., WCDMA networks [33, 35]) or downlink dominated (e.g., small cells in HetNets [23]). In this chapter, we investigate the uplink coverage-limited problem and downlink coverage-limited problem, respectively.

Finally, we consider a joint uplink and downlink network coverage preservation problem in HetNets in which cells can either uplink-dominated or downlink-dominated.

To further explore the opportunities for network energy saving, we propose ”reception mode” design of BSs and then discuss the potential energy-efficient network architectures in HetNets during the low traffic load period.

(37)

2.2 Minimum Power BS Activation Problem for Uplink Coverage-Limited Networks

We first investigate the green network coverage problem in uplink direction. As we know that BS coverage is usually dominated by the downlink traffic load during high load periods. When network traffic load is reduced, most of BS coverage will be uplink limited by the maximum transmit power of MSs. Here we desire to save network energy by powering-off redundant BSs. Here we assume that the maximum transmit power of MSs are identical and BS coverage is restricted by the uplink direction. Then, an optimization for uplink network coverage preservation can be formulated as:

min

M

X

m=1

ImPc,m, (2.16)

subject to

A ⊆

M

[

m=1

C^m, (2.17)

P(Prû(x, y, Pt,k)Im < P_minû )≤ pûout,

∀

^{(x, y)}^{∈ C}^m^,

∀

^m∈ {1, . . . , M}, (2.18) 0 < Pt,k ≤ P^t,max,

∀

^k{1, . . . , K}, (2.19)

Im ={0, 1},

∀

^m∈ {1, . . . , M}. (2.20)

In the problem, an MS shall satisfy the uplink received signal requirement and the total bandwidth constraint for uplink in a BS must be satisfied. Specifically, C^m is the region covered by Bm, and A is the target coverage region. In (2.18), the received power requirement for uplink coverage of Bm at a point (x, y) is given. In (2.19), Pt,k

is the transmit power of MS k and Pt,max stands for the maximum transmit power constraint of MS k. Here we want to determine the BS activation set for minimizing

(38)

network power consumption while avoiding coverage holes in the network. Note that uplink coverage can be extended by TTI (Transmission Time Interval) bundling but only approximately 1 dB gain on SINR (e.g., [36]) can be obtained. Here we focus on the gain obtained by switching off BSs and controlling power of BSs in low traffic conditions. Other potential techniques (e.g., [36]) for uplink coverage extension can be jointly used with our approach.

Let Rm denote as the maximum coverage range of Bm. To satisfy the received signal power requirement at BS side in (2.18) for Bm at active state (i.e., Im = 1), the MS transmit power shall follow the inequality as follows:

Pt,k≥ K⁻¹P_minû R^α_m10^σΨ¹⁰^Q⁻¹^(pûôut⁾, (2.21)

where Q⁻¹(·) represents the inverse Q-function. Given the maximum transmit power of an MS, the uplink BS coverage for Im = 1 can be upper bounded by

Rm ≤ Pt,maxK P_minû 10^σΨ¹⁰^Q⁻¹^(pûôut⁾

!_α¹

. (2.22)

The maximum uplink coverage range for satisfying the received signal requirement is given above. Given resource allocation strategies in BSs, we can obtain the maximum uplink coverage of a BS. On the other hand, the coverage of Bm is C^m is an empty set if Im = 0.

Then, we attempt to find an optimal BS set which minimizes the total power consumption while avoiding coverage holes in the network with an arbitrary deployment of BSs from the obtained maximal coverages. Here we first prove that the minimum power BS activation problem is NP-hard and then propose a heuristic algorithm called Cell Overlap Minimization with Intersection Covered (COMIC) to determine the set of active BSs. Later in Section 2.2.4, we will show via simulations that our proposed COMIC algorithm in this section can approach the performance upper bound derived in this section whenever the network size is much larger than the coverage of a BS.

(39)

2.2.1 Complexity Analysis of Minimum Power BS Activation Problem

The minimum power BS activation problem for minimizing total power by activating a set of BSs while preserving coverage can be closely related to the minimum-weight disk cover problem in which the goal is to minimize the total cost to cover a set of points in the plane by selecting a subset of disks [37–39]. It is known to be NP- hard [39]. Recently, a polynomial time approximation scheme (PTAS) is proposed in [37] for a minimum-weight disk cover problem. However, the authors in [38] show that PTAS for set cover does not exist (unless P = NP ) for circles of roughly the same size. To cover a set of points by unit disks, an algorithm with constant approximation ratio is proposed for minimizing the summation of the weight of disks in [39]. Since the design objective is to cover a set of points instead of covering the entire region, the authors in [39] indicate that one hole or many holes will be present with the proposed algorithm. From the above, we can find that the results in the minimum-weight disk cover problem cannot be directly applicable in our green network coverage problem in which coverage holes shall be prevented.

We first examine the complexity of the problem and derive key properties for preserving network coverage which inspire us to design the BS activation algorithm.

To analyze the complexity of this problem, we look at a network in which 1) the entire network can be covered when all BSs are activated, 2) the entire network is a continuous region within which an MS can move from any position to another position, and 3) no BSs will be placed at the same position.

Given a target covering region A and a set of BSs B, we have the following definitions for the coverage problem.

Definition 2.1. A point i∈ A is said to be covered by an active BS B^m if it is within the coverage range of Bm, D(i, Bm) < Rm in which D(·, ·) is the Euclidean distance between two positions. Otherwise, the point i is said to be not covered by Bm.

Definition 2.2. A boundary intersection is a crossing point between the boundary of two sets. The set can be the coverage of a BS or the target coverage region. Formally,

(40)

Region III

Region II

Region I

w

i^∗_m,u

Bm v

Bn

(a)

w

i^∗_m,u

Bm v

Bn

x u

(b)

w

i^∗_m,u

Bm v

Bn

u x

(c)

Figure 2-2: The relative positions between an uncovered point and the active BS closest to the uncovered point. (a)The regional classification of an uncovered point. (b)The position of an uncovered point in region I. (c)The position of an uncovered point in the region II.

the set of all boundary intersections in the target area A, denoted by I^A, can be represented as

IA ≡ {i ∈ IA| [

Bm,Bn∈B,Bm6=Bn

∂C^m∩ ∂Cⁿ [

Bm∈B

∂C^m∩ ∂A},

where the boundary of coverage Cm is denoted by ∂Cm and the boundary of the target region is denoted by ∂A.

We first derive the sufficient conditions for an activated BS set, denoted by B^on, to maintain network coverage and then show the minimum BS activation problem is NP-hard.

Lemma 2.3. If all boundary intersections of an activated BS set B^on in the target area are covered by B^on itself, all positions in the target region will be covered.

Proof. We prove this lemma by contradiction. Specifically, we will prove that if there exists a point in the network not covered by the set of active BSs, there must exist a boundary intersection which is not covered by the set of active BSs. Suppose that point u ∈ A in the network is not covered by any active BSs. That is, {u ∈ A|u /∈ S

Bn∈BonCⁿ}. If B^m ∈ Bôn is the activated BS closest to the uncovered (not covered) point u ∈ A, we will prove that the boundary intersection closest to u at the boundary of C^m will not be covered by any active BSs. Denote i^∗_m,u ∈ A as the boundary intersection between the coverage of Bm and the coverage of other BS in Bôn closest to u. We will prove that i^∗_m,u will not be covered by any BSs in Bôn if u

睡眠模式在綠能無線網路之建模、設計、與最佳化

國立臺灣大學電機資訊學院電信工程學研究所 博士論文

Graduate Institute of Communication Engineering College of Electrical Engineering and Computer Science

National Taiwan University Doctoral Dissertation

睡眠模式在綠能無線網路之建模、設計、與最佳化 Modeling, Design, and Optimization for Green Wireless

Networks with Sleep Mode Operations

張正義

Chen-Yi Chang

指導教授：許大山 博士 共同指導教授：廖婉君 博士

中華民國 103 年 1 月

誌謝

摘要

Abstract

Contents

List of Figures

Chapter 1 Introduction

1.1 Motivation

1.2 Summary of Contributions

1.2.1 Sleep Mode Operations for Green Cellular Networks

1.2.2 Sleep Mode Operations for Green Wireless Multi-Hop Relay Networks

1.3 Outline

Chapter 2

Sleep Mode Operations for Green Cellular Networks

2.1 System Model and Problem Description

2.1.1 BS Operation Model

2.1.2 BS Power Consumption Model

2.1.3 Channel Model

2.1.4 BS Signal Coverage Model

∀

∀

2.1.5 Green Network Coverage Problem

∀

∀

∀

∀

∀

∀

∀

∀

∀

∀

2.2 Minimum Power BS Activation Problem for Uplink Coverage-Limited Networks

∀

∀

∀

∀

2.2.1 Complexity Analysis of Minimum Power BS Activation Problem

國立臺灣大學電機資訊學院電信工程學研究所博士論文

指導教授：許大山博士共同指導教授：廖婉君博士