PEILIANG CHANG
Doctoral Thesis in Information and Communication Technology School of Electrical Engineering and Computer Science
KTH Royal Institute of Technology Stockholm, Sweden 2018
TRITA-EECS-AVL-2018:80 ISBN 978-91-7729-989-9
SE-164 40 Kista SWEDEN Akademisk avhandling som med tillstånd av Kungl Tekniska högskolan framlägges till of- fentlig granskning för avläggande av doktorsexamen i informations- och kommunikations- teknik fredagen den 30 november 2018 klockan 13:00 i Sal C (Sal Sven-Olof Öhrvik), Electrum, Kungl Tekniska högskolan, Kistagången 16, Kista.
© Peiliang Chang, November, 2018 Tryck: Universitetsservice US AB
Abstract
To assure the sustainable development of mobile networks, it is crucial to improve their energy efficiency. This thesis is devoted to the design of energy-efficient mobile networks. A cross-layer design approach is adopted. The resource management at the MAC layer, the network layer as well as the service layer are optimized to improve the energy efficiency of mobile networks.
The problem of optimizing the MAC-layer resource allocation of the downlink transmission in multi-carrier NOMA systems to maximize the system energy efficiency while satisfying users’ QoS requirements is firstly considered. The optimal power al- location across sub-carriers and across users sharing one sub-carrier are proposed. Fur- thermore, exploiting the structure of the optimal power allocation across users sharing one sub-carrier, a sub-optimal solution for sub-carrier assignment, which greedily min- imizes the required power to serve all users with required QoS, is developed. Besides optimizing the channel assignment and power allocation within a single cell, the link scheduling in the multi-cell scenario to deal with inter-cell interference is also studied.
A scalable distributed link scheduling solution is proposed to orchestrate the trans- mission and DTX micro-sleep of multiple base stations such that both the inter-cell interference and the energy consumption are reduced.
At the network layer, the operation of base station sleeping is optimized to improve the energy efficiency of mobile networks without deteriorating users’ QoS. The spectral and energy efficiency of mobile networks, where base stations are enabled with DTX, under different traffic load is firstly studied. It shows that as the networks are more loaded, the link spectral efficiency reduces while the network spectral efficiency in- creases. Regarding the network energy efficiency, it will either firstly increase and then decrease or always increase when the network load gets higher. The optimal network load to maximize the network energy efficiency depends on the power consumption of base stations in DTX sleep mode. Based on the findings of the above study, the joint optimization of cell DTX and deep sleep to maximize the network energy efficiency is investigated. A scaling law of transmit power, which assures that the distribution of the received power remains unchanged as more base stations are switched into deep sleep, is proposed. Then the average resource utilization and overload probability of non-deep-sleep base stations are derived. Based on these results, the feasible range of the percentage of deep-sleep base stations is obtained. Finally, the optimal percentage of deep-sleep base stations to maximize the network energy efficiency while satisfying users’ QoS requirements is derived.
Lastly, the service-layer resource provision of edge computing in mobile networks is optimized to improve the energy efficiency. With this work, the trade-offs on ser- vice latency and energy consumption between the computation and the communication subsystems are studied. It is shown that the load of the communication subsystem and that of the computation subsystem should be balanced. Increasing the resource of the highly loaded subsystem can significantly reduce the required resource of the other subsystem. An algorithm is proposed to find out the optimal processing speed and the optimal number of active base stations that minimizes the overall energy consumption while assuring the requirements on the mean service latency.
Keywords: green mobile networks, energy efficiency, base station sleeping, re- source allocation, mobile edge computing, interference coordination
Sammanfattning
För att säkerställa en hållbar utveckling av framtidens mobilnät är det avgörande att förbättra energieffektiviteten i dem. Denna avhandling ägnas därför åt utformningen av energieffektiva mobilnät. En designmetod över lagren antas, där resurshanteringen i MAC-lagret, nätverkslagret samt servicelagret optimeras för att förbättra energieffek- tiviteten.
Problemet att optimera MAC-lagrets resursallokering i nedlänk i NOMA-system med flera bärare för att maximera systemets energieffektivitet samtidigt som användar- nas QoS-krav uppfylls betraktas först. Den optimala effektfördelningen över delbärare och över användare som delar en delbärare föreslås. Genom att utnyttja lösningsstruk- turen för den optimala effektallokeringen mellan användare som delar en delbärare, ut- vecklas en suboptimal lösning för delbärartilldelning, vilket gynnsamt minimerar den behövda effekten för att serva alla användare med erforderlig QoS. Förutom att opti- mera kanaltilldelningen och effektfördelningen i en enda cell, studeras även länksche- maläggningen i ett flercellsscenario för att hantera mellancellsstörningar. En skalbar och distribuerad lösning för länkschemaläggning föreslås för att orkestrera sändning och DTX-mikrosömn av flera basstationer så att både mellancellsstörningar och ener- giförbrukning minskas.
I nätverkslagret optimeras driften av basstationens sovande för att förbättra mobil- nätets energieffektivitet utan att för den delen försämra användarnas QoS. Spektral- och energieffektiviteten i mobilnät där basstationer är aktiverade med DTX studeras först under olika trafikbelastningar. Det visar sig att när nätverksbelastningen ökar, så mins- kar länkspektraleffektiviteten medan nätverksspektraleffektiviteten ökar. När det gäller nätverksenergieffektiviteten så kommer den antingen att först öka och sedan minska, eller alltid öka i takt med att nätverksbelastningen ökar. Den optimala nätverksbelast- ningen för att maximera nätverksenergieffektiviteten beror på effektförbrukningen hos basstationer i DTX-viloläge. Baserat på resultaten från ovanstående studie undersöks sedan den kombinerade optimeringen av cell-DTX och djupsömn för att maximera nät- verksenergieffektiviteten. En skalningslag för sändningseffekt föreslås som säkerstäl- ler att fördelningen av den mottagna effekten förblir oförändrad när fler basstationer kopplas om till djupsömn. Genomsnittliga resursutnyttjandet och överbelastningssan- nolikheten för basstationer som ej är i djupsömnläge härleds också. Baserat på dessa resultat erhålls ett möjligt intervall på andelen basstationer i djupsömnläge. Slutligen härleds den optimala andelen basstationer i djupsömnläge för att maximera nätverkse- nergieffektiviteten samtidigt som användarnas QoS-krav uppfylls.
Till sist optimeras resurstilldelningen i tjänstelagret för kantnodsberäkning (eng.
edge computing), i syfte att förbättra energieffektiviteten i mobilnäten. Vi studerar av- vägningen mellan servicefördröjning och energiförbrukning i beräknings- och kommu- nikationsdelsystemen, och visar att belastningen i delsystemen bör balanseras. Att öka resurserna hos det högt belastade delsystemet kan avsevärt minska resurserna för andra delsystem. En algoritm föreslås för att ta reda på den optimala beräkningshastigheten och optimala antalet aktiva basstationer som minimerar den totala energiförbrukningen samtidigt som kraven på genomsnittlig servicefördröjning säkerställs.
Nyckelord: gröna mobilnätverk, energieffektivitet, basstationssömn, resursalloke- ring, mobil kantnodsberäkning, störningssamordning
Acknowledgements
During this four-year pursuit of PhD, numerous helps have been received and so many people to thank. First of all, I am most grateful to my supervisor Professor Guowang Miao for his encouragement, guidance and patience. He provided me enough freedom to explore various topics in the area of green wireless communication, and gave me rigorous comments and constructive suggestions on accomplishing the research work. I have learnt a lot from him on how to do scientific research. My heartfelt gratitude also goes to my co- supervisor Professor Jens Zender for providing me this valuable study opportunity at the Radio Systems Lab and, together with Professor Marina Petrova, guaranteeing a nice and friendly working environment. Also thanks for his insightful comments and suggestions on my research work.
Working at Radio Systems Lab in KTH is a wonderful experience. All colleges at RS Lab are friendly and helpful. My special thanks go to Göran Andersson for his helps on utilizing Mathematica. I would like to thank Professor Slimane Ben Slimane for reviewing my PhD thesis and answering my questions on wireless networks. My sincere thanks to Professor Marina Petrova and Docent Ki Won Sung for the sharp and constructive com- ments on my research work. I would like to thank Yanpeng Yang, Amin Azari, Haris Celic, Meysam Masoudi, Mohammad Istiak Hossain and all other current and former RS Lab colleges for the inspiring sharings and discussions on diverse topics.
My PhD study has been financially supported by the China Scholar Council. I thank them for providing me the PhD scholarship (No.201408610062). I am sincerely grateful to Professor Mats Danielsson and Professor Cicek Cavdar for providing me supplementary financial supports in the beginning and the ending periods of my PhD study.
This acknowledgement would be incomplete without thanking my family. My parents Xihai Chang and Zhen’e Ma have provided me endless love ever since my birth. Without their support, this thesis could not be possible. My brother Peiyu Chang and my sister-in- law Na Zhu also deserve my sincere thanks for taking care of the whole family during my absence. I would also like to express my sincere gratitude to my parents in-law Weihua Liu and Xia Li for taking care my daughter so that I could devote more efforts to this thesis.
Last but not least, my heartfelt thank goes to my daughter Guangying Chang for bring me inner-peace and joy, and my wife Xuejin Liu for the endless support and understanding.
This thesis is dedicated to you.
Peiliang Chang,
November 2018 in Stockholm
Contents xi
List of Figures xiii
List of Tables xv
List of Acronyms xvii
I Thesis Overview 1
1 Introduction 3
1.1 Background and Scope . . . 4
1.1.1 Evolution of Cellular Networks and 5G . . . 4
1.1.2 Huge Energy Consumption of Mobile Networks . . . 7
1.1.3 Challenges for Green Mobile Networks . . . 8
1.1.4 Thesis Scope . . . 13
1.2 Literature Survey . . . 13
1.2.1 MAC-Layer Energy-Efficient Resource Allocation and Link Schedul- ing . . . 14
1.2.2 Control of Base Station Sleeping . . . 16
1.2.3 Energy Efficient Mobile Edge Computing at the Service Layer . . 17
1.3 Research Problems . . . 18
1.3.1 Energy-Efficient Link Scheduling and Power Allocation . . . 18
1.3.2 Traffic-Aware Control of Base Station Sleep . . . 19
1.3.3 Energy-Efficient Mobile Edge Computing . . . 20
1.4 Research Methodology . . . 20
1.4.1 Analytical Approach . . . 21
1.4.2 Monte-Carlo Simulation Approach . . . 22
1.5 Thesis Contributions . . . 22
1.6 Thesis Organization . . . 24 2 MAC-layer Resource Allocation for Energy-Efficient Mobile Networks 25
xi
2.1 Energy-Efficient Resource Allocation in Multi-Carrier NOMA Systems . 26
2.1.1 System Model and Problem Formulation . . . 26
2.1.2 Optimal Resource Allocation in Downlink Multi-Carrier NOMA Systems . . . 27
2.1.3 Numerical Results . . . 28
2.2 Interference-Aware Scheduling of Cell DTX . . . 29
2.2.1 System Model and Problem Formulation . . . 31
2.2.2 Proposed Scheme . . . 31
2.2.3 Numerical Results . . . 32
2.3 Summary . . . 33
3 Joint Optimization of Base Station Deep Sleep and Discontinuous Trans- mission 35 3.1 Performance Analysis of Mobile Networks with Cell DTX . . . 36
3.1.1 System Model . . . 36
3.1.2 Impact of Traffic Load on Network Spectral and Energy Efficiency 37 3.2 Optimal Operation of Base Station Sleeping . . . 42
3.2.1 System Model . . . 42
3.2.2 Main Results . . . 43
3.3 Summary . . . 48
4 Resource Provision for Energy-Efficient Edge Computing 49 4.1 System Model . . . 49
4.1.1 Network Model . . . 49
4.1.2 Power Model . . . 50
4.1.3 Problem Formulation . . . 50
4.2 Optimal Resource Provision for Energy Efficient Edge Computing . . . . 51
4.2.1 Lower Bound of Service Delay . . . 51
4.2.2 Communication-Computation Trade-Off . . . 51
4.2.3 Optimal Resource Provision . . . 52
4.3 Numerical Results . . . 53
4.4 Summary . . . 53
5 Conclusion and Future Work 55 5.1 Concluding Remarks . . . 55
5.1.1 Mac-Layer Resource Allocation . . . 55
5.1.2 Network-Layer Base Station Operation . . . 56
5.1.3 Service-Layer Resource Provision . . . 57
5.2 Discussions and Future Work . . . 58
5.2.1 Energy Consumption of Communication-Related Computing . . . 58
5.2.2 Applying Data Analytic for Base Station Sleep Control . . . 58
Bibliography 61
1.1 Evolution of mobile cellular networks . . . 4
1.2 Three main service scenarios of 5G cellular networks . . . 6
1.3 Comparison of performance requirements between 4G and 5G cellular net- works, from [1] . . . 7
1.4 Resource allocation at different layers in mobile networks . . . 9
1.5 Long-term traffic variation of cellular networks, from [2] . . . 11
1.6 Short-term traffic dynamics of base stations . . . 11
1.7 Framework of research methodology . . . 20
2.1 System performance with different γ0(Cell radius is 500 m) . . . 30
2.2 Traffic service rate with various DTX control algorithms . . . 33
2.3 Network energy efficiency with different cell DTX control algorithms . . . . 34
3.1 CDF of SINR with different inter-site-distances . . . 38
3.2 Link spectral efficiency under different network load . . . 39
3.3 Area spectral efficiency under different network load . . . 40
3.4 Network energy efficiency under different network load (analytical results) . . 41
3.5 Network energy efficiency under different network load (simulation results) . 41 3.6 SNR distribution under different percentages of deep-sleep BSs . . . 43
3.7 Average load of non-deep-sleep BSs under different percentages of deep-sleep BSs . . . 44
3.8 Percentage of overloaded BSs under different percentages of deep-sleep BSs . 45 3.9 Network energy efficiency with power scaling . . . 47
4.1 Number of required base stations to achieve the target mean delay, D0 = 100 ms, λ = 20 tasks/s . . . . 54
4.2 Power consumption under various processing capacity, D0 = 100 ms, λ = 20 tasks/s . . . 54
xiii
2.1 Simulation parameters for the performance evaluation of resource allocation
schemes in downlink NOMA systems . . . 29
2.2 Simulation parameters for inter-cell interference coordination . . . 33
3.1 Maximum value of required ISD for the tolerable loss below 1% . . . 38
3.2 Simulation parameters for optimal control of base station sleep . . . 47
4.1 Simulation parameters for energy-efficient edge computing . . . 53
xv
1G the first generation
2G the second generation
3G the third generation
4G the fourth generation
5G the fifth generation
BS Base station
CSI Channel state information
DTX Discontinuous transmission
EE Energy efficiency
eMBB Enhanced mobile broadband
FFR Fractional frequency reuse
QoS Quality of service
MAC Media access control
ICIC Inter-cell interference coordination
ICT Information and communication technology
IoT Internet of things
LTE Long term evolution
MEC Mobile (Multiple-access) edge computing MIMO Multiple-input-multiple-output
mMTC Massive machine-type communication
MMS Multimedia messaging service
MNO Mobile network operator
NOMA Non-orthogonal multiple access
NP Non-polynomial
OFDM Orthogonal frequency division multiplexing
SC Superposition coding
SMS Short message service
SIC Successive interference cancellation SINR Signal-to-interference-plus-noise ratio
UE User device
URLLC Ultra reliable and low latency communication
xvii
Thesis Overview
1
Introduction
As a main part of the Information and Communication Technology (ICT) infrastructure, mobile networks have been continuously evolving such that their capacities have been re- markably enhanced and plenty of functionalities have been developed. Promoted by the enhanced capacity and functionalities of cellular networks, numerous mobile applications and services have been developed to facilitate our daily life and work and to improve the social well-being. Mobile networks, like other facilities such as water supply infrastruc- ture, have become an indispensable part of our daily life. Accompanying the great success of cellular networks is the significant energy consumption, which brings in both environ- mental and economical problems. The sustainable development of cellular networks calls for solutions to reduce the energy consumption and improve the energy efficiency of cellu- lar networks. It is particularly important to consider energy efficiency as a key performance metric at the stage of designing a new generation of cellular networks.
In this thesis, we deal with the issue of energy consumption of mobile networks with a cross-layer approach. Challenges and problems at different layers of mobile networks, i.e., media access control (MAC) layer, network layer and service layer are examined and identified. With the identified problems, resource management solutions at these layers are proposed to improve the energy efficiency of mobile networks. At the MAC layer, we op- timize the power control and channel assignment in multi-carrier non-orthogonal multiple access (NOMA) systems to improve the energy efficiency of a single cell and propose an interference-aware distributed link scheduling solution to avoid inter-cell interference and improve the energy efficiency of multiple cells. At the network layer, we investigate the spectral and energy efficiency of cellular networks with cell discontinuous transmission (DTX) under various traffic load and optimize the operation of base station deep sleep and DTX. At service layer, we consider mobile edge computing (MEC) as a new functional- ity of future mobile networks and investigate the provision of both communication and computation resources in MEC systems to improve their energy efficiency. The proposed solutions are evaluated with numerical experiments and the experiments results well prove the efficiency of the proposed solutions.
The remaining part of this chapter is structured as follows: we first introduce the back- 3
Figure 1.1: Evolution of mobile cellular networks
ground and scope of the thesis in Section 1.1; Then a survey on the related works is pre- sented in Section 1.2; Following that we elaborate our research problems and our adopted research methodologies in Section 1.3 and 1.4 respectively; Then the contributions of this thesis is summarized in Section 1.5. In the end, the organization of this thesis is provided in Section 1.6.
1.1 Background and Scope
1.1.1 Evolution of Cellular Networks and 5G
It has been around 40 years since the first automated commercial cellular network was launched in 1979 by Nippon Telegraph and Telephone Corporation (NTT) [3]. During these 40 years, mobile networks have been continuously evolving in a pace of one gener- ation per decade (see Figure 1.1). With this continuous evolution, the capacity of cellular networks has been significantly improved and many new functionalities have been devel- oped to support diverse services and applications.
1.1.1.1 From 1G to 4G: a brief history of cellular networks
The first generation (1G) of cellular networks was initially developed and deployed in the late 1970s and early 1980s, and they mainly employed analogue transmission techniques.
The main service application of 1G cellular networks was voice communication. The system capacity was quite limited and the equipment cost was high. All these factors blocked the ubiquitous deployment of the first generation of cellular networks.
The second generation (2G) of cellular networks emerged in the early 1990s. Boosted by the advancement of digital signal processing and integrated circuit technology, the 2G cellular networks have been transferred to digital transmission. Although the main service application was still voice communication, low-rate data services have been introduced in 2G cellular networks, such as short message service (SMS) and multimedia messaging service (MMS). The digitalization of cellular networks not only significantly improved the system capacity, but also reduced the equipment cost and made the mobile voice communi- cation affordable for the masses. Furthermore, thanks to the digital encryption techniques, the security performance of 2G cellular networks has been improved compared to 1G sys- tems. However, the data rate in 2G cellular networks is still quite limited.
As more and more data-consuming applications emerged, the third generation (3G) of cellular networks have been developed to provide mobile broadband access in early 2000s.
3G cellular networks employed larger bandwidth of radio resource and the supported data rate has been elevated to hundreds of kbps, compared to that of several kpbs in 2G systems.
Besides voice communication, other mobile services supported by 3G cellular networks are web browsing and mobile music, which consume moderate data rate.
The revolutionary invention of smart phones, such as Apple phone, has promoted the proliferation of mobile applications. Various mobile applications and services, ranging from social networking to mobile payment, from video to online gaming, have been devel- oped. All these mobile applications, particularly the content-rich applications such as video and live broadcast, rely on Internet access with much higher data rate. The fourth gener- ation (4G) of cellular networks, represented by the long-term-evolution (LTE) systems, was developed there to support these applications in the 2010s. The mult-carrier trans- mission techniques is employed in LTE systems to achieve higher spectral efficiency. The adopted radio bandwidth is further increased, compared to that of 3G systems. Besides, multiple antenna techniques are also utilized in LTE systems. All these factors contribute to achieve higher data rate, in the order of mega bps. Furthermore, different from the previous generations, LTE systems adopt an "All-in-IP" approach and replace the previous circuit-switching technology with packet switching solutions. All services, including voice communication, are realized with Internet-Protocol (IP) based packet switching technolo- gies.
1.1.1.2 The 5G cellular networks
The evolution of mobile networks does not stop. The future society is expected to become more connected [4] and the next generation mobile networks, the fifth generation (5G) mo- bile networks, are expected to support not only data-intensive applications, such as high definition video and virtual reality, but also other diverse applications which require more stringent latency and reliability performance. As a pre-standardization research project on 5G, METIS project has identified three categories of service scenarios of 5G cellular net- works [1, 5]. As illustrated in Figure 1.2, these three categories of service scenarios are the enhanced mobile broadband (eMBB), the ultra reliable and low latency communica- tion (URLLC) and the massive machine-type communication (mMTC). The eMBB ser- vice category covers the services that require high data rate, such as high definition video
Figure 1.2: Three main service scenarios of 5G cellular networks
services and augmented reality services. The URLLC service category mainly includes the services that require high reliability and stringent communication latency, such as the vehicular-to-vehicular communication and the automatic control of industrial plants. The mMTC service class mainly consists of the services where a huge number of devices need to be connected via cellular networks, such as the smart metering of water and electricity.
These different service scenarios have different performance requirements on the 5G mobile cellular networks. In order to support these heterogeneous services, the 5G cellular networks needs to not only surpass the legacy systems in term of supported data rate but also achieve largely reduced end-to-end service latency and enhanced reliability. Mean- while, 5G cellular networks need to provide connectivity to devices with much higher density. A comparison on key performance metrics between 5G cellular networks and LTE networks is illustrated in Figure 1.3. Compared to LTE networks, it is expected 5G cellular networks will achieve the following performance improvements [1]:
• 1000 timers higher mobile data volume per area
• 10 to 100 times higher number of connected devices
• 10 time longer battery lifetime for low power massive machine type communication
• 5 times reduced end-to-end latency
Last but not least, these performance requirements should be achieved in an economically affordable way and the overall energy dissipation of 5G networks should be at the same level as that of LTE networks.
It relies on diverse technologies to achieve the targeted performance requirements in 5G mobile cellular networks. For example, in order to further improve data rate, we can deploy
Figure 1.3: Comparison of performance requirements between 4G and 5G cellular net- works, from [1]
more access points [6,7], employ the massive MIMO technology [8,9] and utilize larger ra- dio resources at millimetre band [10]. In order to provide connectivity to devices with much higher density, we can employ the non-orthogonal-multiple-access (NOMA) technology to multiplex multiple links on the same block of time-frequency resource [11]. In order to re- duce the end-to-end latency, we can utilize the mobile edge computing (MEC) technology and provide computing service at the proximity of mobile devices [12]. Although specific techniques can be adopted to achieve certain targeted performance requirement, it is more challenging to address the issue of energy consumption as the energy consumption of cel- lular networks is related to many aspects it requires a comprehensive approach to improve the energy efficiency of future cellular networks.
1.1.2 Huge Energy Consumption of Mobile Networks
There is no such thing as a free lunch. One fact behind the great success of cellular net- works is their huge energy consumption. The capacity enhancement of cellular networks is mainly based on three pillars [13, 14]: the improvement of spectral efficiency via ad- vanced signal processing and resource management, the extension of radio band by ex- ploring more frequency bands and the densification of network by deploying more base stations. All these three pillars contribute to increasing the energy consumption of mobile
networks: although advanced waveforms, like OFDM, can help to achieve higher spectral efficiency, the higher peak-to-average-power ratio leads to low efficiency of power ampli- fier and high energy consumption [15]; the extension of radio band mainly comes from the exploration of higher frequency radio band, which generally suffers higher path loss and requires higher transmit power to cover the same area; the network densification requires deploying more base stations, which increases the number of energy consumption sources.
The fact that cellular networks consumes huge amount of energy has been pointed out by many sources [16–18]. In [17], it is predicted that mobile networks would con- sume 0.5% of overall global energy consumption, which is around 200 Million tons of CO2 emission. Besides the environmental concern, the huge amount of energy consump- tion also increases the operation cost of mobile networks. Research shows that the elec- tricity bills contribute substantially to the mobile network operator (MNO)’s operational cost [19, 20] and the increasing energy consumption continues to reduce the profit margin of MNO. Meanwhile, the capacity of the 5G networks is expected to be 1000-fold of that of 4G networks. These would leads to unsupportable energy consumption unless the energy efficiency of 5G networks is dramatically increased. Therefore, it is of paramount impor- tance to improve the energy efficiency of mobile networks for the sustainable development of cellular networks.
1.1.3 Challenges for Green Mobile Networks
The breakdown of the power consumption of mobile networks shows that the main energy consumer of mobile networks are base stations; Power amplifier, among all the compo- nents, contributes the most to the energy consumption of a base station [21–23]. The energy efficiency of mobile networks can be improved from different aspects [21, 24]. For example, we can improve the energy efficiency of hardware with advanced design and sig- nal processing [25, 26]. More efficient cooling solutions can be implemented to reduce the energy consumption of cooling systems in base stations. Furthermore, the deployment of base stations in cellular networks can be improved to save energy [27].
Besides these static solutions, the dynamic network operation and resource manage- ment, including both communication and computation resources, can be improved to fur- ther enhance their energy efficiency [28–33]. As illustrated in Figure 1.4, the resource management in mobile networks can be optimized at multiple layers. At the bottom of the system, the physical layer deals with channel coding/decoding, modulation/demodulation and other signal processing. The performance of the physical layer functionalities depends on the MAC layer power allocation and user scheduling. Thus, the PHY-MAC cross-layer optimization of power allocation and user scheduling can be applied to improve the energy efficiency of a single base station. Also link scheduling of multiple base stations to can be coordinated to avoid inter-cell interference, which can help to reduce energy consumption and improve network throughput. At the network layer, base station sleeping at differ- ent depths can be employed to adapt the operation modes of base stations to the temporal variation of traffic such that the energy consumption during idle period can be reduced.
At the service layer, particularly for the scenario of MEC, the computation resource, e.g., the operation frequency of CPU, and the communication resource can be jointly optimized
Figure 1.4: Resource allocation at different layers in mobile networks
to minimize the overall system energy dissipation while satisfying the quality of service (QoS) requirements. A detailed description on the challenges of optimizing resource allo- cation at these three layers is presented in the following section.
1.1.3.1 MAC Layer: Energy-Efficient Resource Allocation and Link Scheduling As introduced in Section 1.1.1, one performance requirement of 5G mobile networks is to provide connectivity to a huge number of IoT devices. Multiple measures can be adopted to fulfil this performance requirement. We can employ more radio bandwidth at higher frequency band, exploit more spatial freedoms with the MIMO technology or apply the NOMA technology [11, 34]. Multiplexing the data of multiple users on one time- frequency resource block, NOMA can not only enable more simultaneous connections but also achieve higher spectral efficiency, compared to the orthogonal-multiple-access (OMA) systems [11].
the MAC-layer resource allocation in NOMA systems, including power allocation and sub-carrier allocation, is important as system performances such as user fairness, energy efficiency and spectral efficiency are highly affected by these factors. As multiple users share the same time-frequency resource, there unavoidably exists inter-user interference, whose impact on link quality depends on the adopted power allocation and user scheduling schemes. Different resource allocation solutions lead to different system performance on the system throughput, energy consumption, and other user QoS. Therefore, the MAC- layer resource allocation in NOMA systems should be carefully designed to maximize the
system energy efficiency while satisfying the requirements on other performance metrics.
The MAC-layer link scheduling also plays an important role in handling the inter-cell interference coordination (ICIC) in multi-cell systems. In order to further improve the capacity of mobile networks and attain the target of 1000x data rate enhancement, we can deploy more base stations [6, 7] and more aggressively reuse the radio resources. The channel gain of radio links between base stations and user devices improves due to the shortened distance when more base stations are deployed. Therefore, user devices are more likely to be interfered by their neighbouring base stations [35], especially when the network load is high. The situation gets even worse when the deployment of base stations is unplanned. Strong inter-cell interference can not only deteriorate user throughputs, but also waste energy as transmitting the interfering signal consumes energy.
The inter-cell interference can be avoided with coordinated link scheduling across mul- tiple cells, i.e., the links that mutually interfere each other should be scheduled on orthogo- nal resources. Meanwhile, with the time domain ICIC, the base stations that are idle during certain slots can be switched into lower power mode, e.g, applying the DTX technique, to save energy. Therefore, interference-aware scheduling of cell DTX is an important way to improve the energy efficiency of cellular networks.
Inter-cell interference coordination (ICIC) can be achieved with centralized link schedul- ing. However, this approach requires both huge signalling overhead to collect channel state information (CSI) and traffic information and significant computation effort to solve the complex scheduling problem. It is desirable to have distributed ICIC solutions that is scalable, takes low signalling overhead and achieves close-to-optimal performance.
1.1.3.2 Network Layer: Traffic-Aware Operation of Mobile Networks
As being recognized by many research efforts [21, 36–38], a clear fact about the traffic load of mobile networks is its temporal variation. The traffic fluctuation occurs at different time scales. At large time scare, due to the mobility and activity patterns of users, the traffic intensity of mobile networks changes from hour to hour and it can be extremely high during certain period while becomes low in other periods. For example, in a central business district area, mobile networks are highly loaded during day time while there is little traffic request during deep night. The traffic profile of mobile networks during 24 hours is illustrated in Figure 1.5. The large time-scale traffic variation is periodic and stable. Traffic prediction techniques can be applied to forecast the upcoming traffic load of mobile networks [39]. In addition to the large-time-scale traffic variation, the traffic request of each base station fluctuates at smaller time scale, e.g., at the level of several milliseconds.
This small-time-scale traffic dynamic is due to the burstiness of traffic request. As shown in Figure 1.6, the traffic request of each base station is not a constant flow. Each base station serves traffic request at certain time while it becomes idle when there is no traffic request.
Different to the large time-scale traffic variation which is easy to forecast in advance, the small-time-scale traffic dynamic is hard to predict.
The temporal traffic variation of mobile networks can be exploited to improve their en- ergy efficiency. Mobile networks are generally dimensioned according to the peak traffic load. At low load periods, network capacity is redundant and energy is wasted if all base
24 hours
Site Traffic Level
1 second
Figure 1.5: Long-term traffic variation of cellular networks, from [2]
BS1
BS2
BS3
t0
t
t
t
TR TR TR TR TR
TR TR TR TR TR
TR TR TR TR TR
BSn
t
TR TR TR TR TR
t0 +∆t
Figure 1.6: Short-term traffic dynamics of base stations
stations are active as the energy consumption of base stations is not proportional to their traffic demand [37]. In order to save energy, some base stations can be switched off (deep sleep) at low load periods and the remaining base stations serve traffic requests. Base sta- tion on/off switch takes considerable delay and it is operated at large time scale. Network operator can pre-configure the on/off work mode of base stations based on the long-term
traffic profile [38]. Besides the long-term traffic variation, the short-term traffic dynamics can also be exploited to reduce energy consumption [2,40]. This is achieved with cell DTX technique. With cell DTX, instead of being entirely switched off, base stations can turn off some components at idle periods and enter micro sleep to reduce energy consumption and activate these components when there is traffic request.
Two different approaches of saving energy are represented by the long-term deep sleep and the short-term DTX micro sleep. The idea of the long-term deep sleep is to reduce the number of active BSs and let the remaining active BSs to carry traffic load while the DTX technique aims to let BSs complete serving traffic requests as soon as possible and then enter the micro sleep mode to save energy. With DTX, the traffic is served by each BS individually and it does not require traffic redistribution. With the same amount of traffic load, the load of active BSs increases when more BSs are switched into deep sleep. Mobile networks then become more congested, which would reduce the spectral efficiency and fur- ther increase the resource utilization of active BSs. Consequently, the active BSs stay less time in the DTX micro-sleep mode and their energy consumption increases [40,41]. When we employ both the long-term deep sleep and the short-term DTX micro sleep , it is unclear whether switching more BSs into deep sleep would reduce the overall energy consumption or not. Furthermore, users’ QoS might get affected by switching BSs into deep sleep. First of all, there might be some coverage hole when BSs are switched into deep sleep. The transmit power of the remaining active BSs need to be increases to avoid coverage hole.
Secondly, the load of the remaining active BSs increases when more BSs are switched into deep sleep. With increased load, on one hand the coverage probability decreases as the inter-cell interference gets stronger when mobile networks are more loaded; on the other hand it is more likely that the remaining active BSs get overloaded and this results in higher blocking probability of users as new traffic arrivals are blocked when BSs are overloaded.
Therefore, we should carefully control the joint operation of base station deep sleep and DTX such that the overall energy consumption is reduced while satisfying the user’s QoS requirements.
1.1.3.3 Service Layer: Energy-Efficient Resource Provision for MEC Systems The last decade has witnessed a magnificent proliferation of smart mobile devices and a great prosperity of mobile applications. Certain newly emerged applications, e.g., aug- mented reality and online gaming are highly computation-demanding. Conducting such computations locally on the mobile devices on one hand brings in significant computation delay as the computation power of mobile devices is limited, and on the other hand drains the battery of mobile devices as the battery capacity does not increase at the same pace as that of the power consumption of both communication and computation functionalities.
Along with the rapid development and application of cloud computing, one solution to reduce the computation burden of mobile devices is to offload their computation intensive tasks to the cloud. One intrinsic drawback of offloading computation tasks to the cloud is the significant communication delay between the mobile devices and the could servers, which is unacceptable for the applications with stringent delay requirements.
Edge computing [12,42] is proposed to surmount this drawback. With edge computing,
the computation resource is pushed to the edge of the Internet, such as the base stations.
The communication delay between the mobile devices and the servers can be tremendously reduced. This enables the offloading of computation-intensive tasks that have stringent delay requirements.
Stories of cloud data centers tell us the computation in data centers consumes tremen- dous energy. This fact makes us believe that with edge computing at base stations, the next generation mobile networks would consume more energy, in addition to the already-high energy consumption of communication functionalities. As a new computation paradigm, it is important to consider the energy performance of edge computing at its early design stage. Although existing solutions on improving energy performance of cloud data centers may be applied, edge computing has its own unique challenges and opportunities to im- prove its energy performance. As the quality of service and energy consumption of edge computing are affected by both communication and computation processes, it would be in- teresting to investigate the management of both communication and computation resources to improve the overall system energy efficiency without sacrificing the quality of service.
1.1.4 Thesis Scope
The goal of this thesis is to improve the energy efficiency of the next generation mobile networks by optimizing the resource management at different layers. This cross-layer op- timization approach mainly consists of three aspects: at the MAC layer, we would like to develop solutions for sub-carrier assignments and power allocation to improve the energy efficiency of NOMA systems while considering all users’ QoS requirements. Meanwhile, we would like to develop scalable link scheduling solutions with low complexity to or- chestrate the transmission and DTX of multiple base stations such that both the inter-cell interference and the energy consumption are reduced; at the network layer, we aim to de- velop traffic-aware control solutions of base station sleep at different levels of depth to save energy while satisfying QoS requirements; at the service layer, we try to investigate the trade-offs between communication and computation in MEC systems and to develop optimal resource provision solutions to improve the energy efficiency of the entire systems.
The scope of this thesis is summarized with the following high-level research questions:
• RQ1: How to schedule link and allocate resource at the MAC layer to improve the network energy efficiency?
• RQ2: How to control the work mode of base stations with multiple levels of sleeping to save energy while satisfying QoS requirements?
• RQ3: How to provision both computation resource and communication resources to achieve energy-efficient edge computing?
1.2 Literature Survey
In this section we present an overview of existing works related to the scope of this thesis and clarify the research gap, based on which we formulate our research problems in the
Section 1.3.
1.2.1 MAC-Layer Energy-Efficient Resource Allocation and Link Scheduling
1.2.1.1 Resource Allocation in NOMA Systems
The design of MAC-layer energy efficient resource allocation solutions in NOMA systems has attracted numerous research efforts. In [43], the authors studied the joint power con- trol and sub-carrier assginment in multi-carrier NOMA systems to maximize the weighted sum of all users’ throughput. A globally optimal solution based on monotonic optimiza- tion is developed. To overcome the high complexity of the proposed optimal solution, a sub-optimal scheme based on successive convex approximation was proposed. With the objective to maximize the weighted sum of all users’ throughput, the work [44] studied the joint optimization of sub-carrier assignment and power allocation. In order to address the issue of NP-hard complexity of the original optimization problem, the authors proposed to iteratively solve the sub-carrier assignment problem with the two-sided matching method and the power allocation problem with geometric programming. Consider the proportional fairness as a performance metric to balance between system throughput and user fairness, the authors of [45] studied the user-paring and power allocation in the two-user NOMA systems where the number of co-channel users is limited to two. A systematic study on the tractability and computation complexity of MAC-layer resource allocation problems in NOMA systems under a range of constraints and utility function was presented in [46].
The above works mainly focus on improving the system throughput or user fairness. The energy efficiency of NOMA systems was not considered there.
Some works were devoted to minimizing the transmit power of NOMA systems. The problem of optimal power allocation to minimize the overall transmitted power in a multi- carrier NOMA system was studied in [47] and a sub-optimal algorithm based on "relax- then-adjust" was proposed there. Considering the imperfect CSI, the authors of [48] studied the optimization of power allocation, rate allocation, user scheduling and SIC decoding policy in downlink multi-carrier NOMA systems to minimize the overall transmit power.
Although the transmit power is important to the overall energy consumption, the circuit power was not considered in the above works.
The static circuit power was considered in some other works. Considering a single- carrier system, the authors of [49] investigated the power allocation problem to maximize the system energy efficiency and the sub-carrier assignment problem was not solved there.
The energy efficient resource allocation in millimetre wave massive MIMO systems with NOMA was studied in [50]. In the proposed solution, users are grouped into clusters based on their CSI, and the number of users in each cluster is limited to two. Then a power allocation strategy based on non-linear fractional programming [51] was proposed to max- imize the system energy efficiency. A similar work was presented in [52] where the authors considered the energy efficient resource allocation in general MIMO systems. In this work the users are randomly grouped into clusters with equal size and the number of users in one cluster can be larger than two. The author of [53] studied the optimal sub-carrier as-
signment and power allocation to maximize the energy efficiency for downlink NOMA systems. In the proposed solution, the sub-carriers were assigned to users based on two- sided matching and the number of users on each sub-carrier is limited to two. Regarding the power allocation across sub-carriers, an iterative solution based on difference of con- vex programming is proposed to approximate the optimal power allocation. Although the above works consider both the transmit power and the static circuit power, their studied problems are limited as they only consider the signle-carrier scenario [49] or only power allocation and ignore user grouping [50, 52] or the considered user grouping is limited to two user per group [53] which is not general. Furthermore, in [53], the proposed user grouping is based on preference lists that are purely determined by the static channel gain and the mutual affects among users are not considered.
1.2.1.2 Inter-Cell Interference Coordination
The management of inter-cell interference is an old research topic but with new importance in the context of densified mobile networks. Different approaches can be adopted to deal with inter-cell interference, such as interference avoidance, interference randomization and interference cancellation [54]. Among these three solutions, interference avoidance is the most appealing as it does not require complexed signal processing as the other two do.
Interference avoidance can be realized statically by strategic resource planning or dynam- ically by coordinated resource allocation among neighbouring BSs [55]. For the inter-cell interference that occurs in a stable zone of the networks, orthogonal access in frequency domain or time domain among neighbouring BSs can be planned to avoid inter-cell inter- ference. For example, the fractional frequency reuse (FFR) [56] is used to avoid inter-cell interference in the cell edge of regular hexagonal cellular networks; and the almost-blank slot based inter-cell interference avoidance is employed to deal with the strong interference from macro BSs in the expanded coverage area of small cells [57]. Although this approach can improve the link capacity of cell edge users, it comes at the cost of degraded overall network capacity due to the reduced frequency reuse. Meanwhile, it is also challenging to implement this approach in the dense small cell networks due to the irregular geometry of small cells. Besides the static solutions, inter-cell interference can be avoided by dynam- ically and cooperatively allocating radio resources among neighbouring BSs [55]. Based on the instantaneous traffic information and CSI, resource allocation decisions are made to avoid interference. Despite the superb performance, the dynamic interference avoidance solutions usually requires central controller and huge amount of control overhead, which hinders their applications in large-scale networks. Scalable and low-overhead solutions are needed to manage the inter-cell interference in dense mobile networks.
Consider the non-homogeneous distribution of users in cellular networks, the authors of [58] proposed a dynamic and distributed classification of inner and outer zone users in FFR scheme to balance load and avoid local traffic congestion. In [59], reinforcement learning was proposed to realize the distributed ICIC in heterogeenous networks. In the proposed solution, the low power pico cells autonomously learn their optimal cell range expansion and power allocation to deal with inter-cell interference. The reinforcement learning method was also applied in [60] to improve the operation of cell DTX to simul-
taneously avoid inter-cell interference and reduce energy consumption of base stations.
The work [61] studied the distributed link scheduling in multi-cell systems to maximize the weighted sum of rate. Considering the dominant interference, a polynomial time dis- tributed algorithm is proposed to realize ICIC. Game theory was applied in [62, 63] to develop distributed ICIC schemes. Both the learning-based approach and the game-theory approach require a considerable convergence period to attain the optimal policy and a new convergence processing is needed whenever the interference situation changes, which is due to user mobility and traffic dynamics. Different from these two approaches, we aim to develop quick and simple solutions to orchestrate the transmissions and micro sleep of base stations.
1.2.2 Control of Base Station Sleeping
In the last decade, lots of research efforts have been devoted to studying the base station sleeping based energy saving approach. In [29], the base station sleeping operation was demonstrated to have a great potential for energy saving. The authors of [38] explored the traffic pattern of cellular networks and developed a profile-based BS sleeping control approach to improve the energy efficiency of 3G cellular networks. In [64], considering both energy saving and flow-level delay, the authors investigated the optimal control of user association and BS sleeping operation. The authors of [65] investigated the impact of BS sleep frequency on the energy saving of mobile networks and it was shown that the daily traffic pattern is important for the operation of BS. In [31], the authors proposed so- lutions for estimating the load of base stations. Based on the estimated load, a distributed BS switch on/off scheme was developed. The work [66] considered relay-assisted cellular networks and explored the joint problem of BS-relay association and BS sleep control. The design of green cellular networks via BS sleeping and deploying small cells was investi- gated in [33] and the trade-offs related to these approaches were studied there. The authors of [67] proposed solutions for the joint control problem of user association and the work mode of femto base stations. In [68], the authors exploited the user mobility information and developed a distributed BS sleep control solution. These works mainly focused on BS long-term sleeping operations that leverage large-scale traffic variation.
On the other hand, energy saving via BS DTX that relies on fast mode switching has received less attention. The authors of [2, 40] examined the feasibility of DTX in LTE networks and showed that cell DTX can contribute to save significant energy. The authors of [41] studied the energy saving by jointly considering cell DTX and network deployment.
The joint optimization of power control, cell DTX and antenna adaptation in multi-user MIMO-OFDM systems was studied in [69]. Existing works on cell DTX mainly focused on the energy saving of cell DTX while other aspects of network performance, such as network spectral efficiency, coverage probability are not considered. Further more, in most of works, cell DTX operations are passively driven by the traffic dynamic. No control solution of cell DTX considering interference-avoidance, which is identified as a main limiting factor of network spectral efficiency and energy efficiency, was proposed in the literature.
Several works have considered the joint optimization of BS deep sleep and DTX micro sleep [70, 71]. They studied the control of BSs with multiple levels of sleep depths and focused on the trade-off between energy consumption and BS wake-up time. They did not consider the impact of BS deep sleep on the energy consumption of the remaining BSs. The energy consumption of the non-deep-sleep BSs is increased by switching BS into deep sleep for two reasons: the first is that we need to scale up the transmit power of the non-deep-sleep BSs to avoid coverage hole; and the second is that the non-deep-sleep BSs are more loaded and the network becomes more congested. As the network becomes more congested, the inter-cell interference would become more significant. This leads to reduced spectral efficiency, which further increases the resource utilization of non-deep- sleep BSs. Thus, the energy saving of non-deep-sleep with cell DTX is reduced. Therefore, it is valuable to find out how many base stations should be switched into deep sleep when both deep sleep and cell DTX are applied.
1.2.3 Energy Efficient Mobile Edge Computing at the Service Layer
Motivated by the great potential of computation offloading in the next generation of mo- bile networks, numerous works have been devoted to the joint management of both com- munication and computation resources in computation offloading and two recent surveys summarizing the most recent works on computation offloading can be found in [72] [73].
In a seminal work [74], general design guidelines for saving the battery of mobile de- vices with computation offloading are provided. In [75], the authors considered the single- user scenario and studied the optimal offloading strategy under stochastic wireless channel to minimize the energy consumption of mobile devices. Both the CPU frequency of lo- cal computation and the transmission rate for offloading the task were optimized and a threshold-based offloading strategy was developed. The work [76] considered partial of- floading of computation task and proposed an heuristic program partitioning algorithm to minimize the computation latency. The partial offloading was also considered in [77] and the joint scheduling and computation offloading for multiple-components applications was studied. Parallel processing of components can be achieved and the energy consumption of mobile devices is reduced. Focusing on service latency, the work [78] investigated the optimal assignment of tasks to local and remote computers considering the application- specific profile, availability of computational resources and link connectivity and proposed a low-complexity approximation scheme with guaranteed performance. Joint computation- communication resource management in the context of multiple users and multiple edge servers to optimize the energy performance of mobile device or the service latency per- formance were studied in [79–88]. Most of the works studies the design of computation offloading from the user side, focusing on the quality of experience of user devices, such as the latency of accomplishing a task and the energy consumption of user devices. Little attention has been paid to the energy consumption of edge servers, which is one of the main focuses of this thesis.
1.3 Research Problems
Based on our research goal in Section 1.1 and the literature surveyed in Section 1.2, we elaborate our research problems in this section.
1.3.1 Energy-Efficient Link Scheduling and Power Allocation
• Research problem 1(RP 1): How to assign sub-carrier and allocate power in down- link multi-carrier NOMA systems to maximize the system efficiency which satisfying users’ QoS requirements?
NOMA is proposed to address the challenge of massive connectivity and achieve higher spectral efficiency in 5G mobile networks. In downlink NOMA systems, superposition coding (SC) [89] is applied at the transmitter side to multiplex signals from different users on the same carrier. Successive Interference Cancellation (SIC) is applied at the receiver side such that the signal with strongest power is decoded first and removed from the origi- nal signal. Then the signal with second strongest power is decoded and removed from the original signal. So on so forth until the desired signal is decoded. The signals of users with stronger channel gain than the considered user are treated as noise. The link qual- ity of a specific user depends on its channel gain, allocated power and the power of other co-channel users.
Different schemes of sub-carrier assignment and power allocation could result in differ- ent data rates for users in the systems, which eventually leads to different QoS and system energy efficiency. Therefore, the first research problem of this thesis is to investigate the optimal solution of sub-carrier assignment and power allocations in single-cell NOMA sys- tem to maximize the system energy efficiency while satisfying all users’ QoS constraints.
This research problem consists of three sub-problems: how to assign sub-carriers to users and form user groups? With a given sub-carrier assignment, how to allocate power across different sub-carriers? With an allocated power of a sub-carrier, how to allocate power among users that share the sub-carrier?
• Research problem 2 (RP 2): How to orchestrate the transmission of multiple base stations to reduce both inter-cell interference and energy consumption in a scalable and simple way?
Without any coordination, the transmissions of multiple neighbouring base stations in mo- bile networks might collide and result in sever inter-cell interference, which not only wastes energy but also degrades users’ QoS. Ideally, the transmissions of neighbouring base sta- tions can be orchestrated with MAC layer link scheduling so that we can not only avoid inter-cell interference but also save energy by switching idle base stations into low power mode, e.g., applying DTX. Considering the fact that centralized link scheduling solutions have high signalling overhead and complexity, we need to develop distributed solutions, which is scalable and requires low communication overhead, for link scheduling and DTX control.
1.3.2 Traffic-Aware Control of Base Station Sleep
• Research problem 3 (RP 3): What is the energy and spectral efficiency of cellular networks with DTX under different traffic load?
Before working on the design of optimal operation of BS sleeping, we would like to find out the energy and spectral efficiency of cellular networks with DTX. Cell DTX may affect network performance from several aspects. With DTX, base stations have two work mode: micro-sleep mode and transmission mode. During the micro-sleep mode, the trans- mitters don’t transmit any signals, which not only reduces power consumption but also decreases inter-cell interference. And during the transmission mode, base stations serve traffic requests. The time that a base station stays in the transmission mode depends on the load of the considered base station. With different loads, the performance, e.g., the cov- erage probability, energy efficiency and spectral efficiency of mobile networks with DTX would be different. Therefore, the third research problem is to find out network energy and spectral efficiency under different traffic load. Specifically, we try to find out the impact of network load on network energy efficiency, spectral efficiency and coverage probability and to understand under what load would network energy efficiency and spectral efficiency be optimal.
• Research problem 4 (RP 4): How to optimally operate networks when both base station deep sleep and DTX are applied?
Although switching BSs into deep sleep can save energy, we should keep in mind the impact of BS deep sleep on other aspects of network performance. Switching BSs into deep sleep will affect affect multiple network performance metrics. Firstly, the received power of the users covered by the switched-off BS would decrease due to the enlarged UE- BS distance unless active BSs scale up their transmit power when more BSs are switched into deep sleep. Secondly, for a given traffic intensity, switching more BSs into deep sleep would increase the load of non-deep-sleep BSs, which on one hand increases of the overload probability of active BSs and results in higher blocking probability as new traffic arrivals would get blocked when BSs become overloaded; and on the other hand degrades the link quality as highly loaded BSs are more likely to generate interference. Furthermore, energy saving of active BSs via DTX micro-sleep will be reduced. Therefore, the fourth research problem of this thesis is to optimize the deep sleep operation of BSs such that the network energy efficiency is maximized while the network performance requirements on coverage rate and BS overload probability are met. In order to find out the optimal operation of base station deep sleep, we need to understand how the transmit power of the active base stations should scale up as more base stations are switched into deep-sleep mode, and how the load of the active base stations increases as more base stations are switched into deep sleep.
Figure 1.7: Framework of research methodology
1.3.3 Energy-Efficient Mobile Edge Computing
• Research problem 5 (RP 5): What are the trade-offs between the computation and communication subsystems in a MEC system? How to provision the computation and communication resources of a MEC system to minimize its overall energy con- sumption while satisfying the required QoS?
MEC is proposed to reduce the end-to-end service latency in 5G mobile networks. In order to utilize the computation resource at base stations, user devices need to firstly offload their computation tasks to base stations. Then base stations accomplish the requested computa- tion tasks and send the computation results back to user devices. In this process, both the computation subsystem which is needed to offload the computation task and to send back the computation results and the computation subsystem which accomplishes the compu- tation task, consume energy and impact service latency. There exist trade-offs between these two subsystems. The fifth research problem of this thesis is to characterize the trade- offs between the two subsystems and to develop efficient resource provision solutions to minimize the overall energy consumption networks without degrading users’ QoS.
1.4 Research Methodology
In this section, we introduce the methodologies that we adopted to study the research prob- lems presented in Section 1.3.
As the research object of this thesis is mobile networks which are very hard to set up a system for experiments in the laboratory, instead of real-system based experiments, we mainly rely on two alternative approaches for conducting research: the analytical ap- proach and the Monte-Carlo simulation approach. The framework of our adopted research methodologies is presented in Figure 1.7. For the analytical approach, we start with the research problems. For a given research problem, we choose proper models and make rational assumptions on the network topology, channel gain and traffic arrivals and then
formulate the research problem mathematically. Certain approximations are adopted by the analytical approach for the sake of tractability. These approximations are justified by comparing the analytical results to the simulation results, which are obtained by the simu- lation approach. By solving the formulated problem, we derive closed-form solutions for the research problem. Meanwhile, for each research problem, we design and implement numerical experiments. For the numerical experiments, we implement virtual cellular net- works in Matlab, simulate the operation of real networks repeatedly and then collect net- work performance metrics. Monte-Carlo simulations are conducted and By analysing both the analytical and simulation results, we answer the corresponding research problem and gain insights and guidelines on the design and operation of real systems.
1.4.1 Analytical Approach
1.4.1.1 Evaluation of SINR Distribution with Stochastic Geometry
This thesis mainly focuses on the optimization of resource management at different layers of mobile networks. Toward this end, it is important to characterize the impact of different network parameters, such as the traffic intensity and the number of active BSs, on the performance metrics, such as network spectral efficiency, network spectral efficiency, and network coverage probability.
In order to analytically model the relationships between the key performance metrics and the key network parameters, we need to properly model the radio proposition, the net- work topology, the power consumption, and the traffic arrival etc. For the channel gain of a given link, both the large-scale fading and the small-scale fading are considered. The Rayleigh distribution is used to model the small-scale fading. For the traffic arrival, we mainly user the Poisson arrival process. For the power consumption of BSs, we mainly resort to the affine power model reported in [37], which is widely accepted in the liter- ature. More details of the adopted model for each research problem are provided in the corresponding chapter.
We resort to theories of stochastic geometry to model the geometry of base stations and user devices. Multiple models can be used to model and analyse the performance of cellular networks, such as the Wyner model [90] and the hexagonal grid model [54]. Al- though the Wyner model is tractable, it fails to capture the principal characteristics of the realistic cellular networks [91]. Regarding the grid model, it becomes intractable when the network size increases. Furthermore, the grid model cannot describe the irregularity of the practical cellular networks, which is more and more common due to the increasing deploy- ment of small cells, such as pico cells and femto cells. In this thesis, the cellular network is modelled as a homogeneous Poisson point process (PPP) and the tools of statistical geom- etry [92] are utilized to analyze the performance of cellular networks. This model could provide both accurate and tractable results on the network performance [93–96].
1.4.2 Monte-Carlo Simulation Approach
In order to validate our findings of the analytical approach, we resort to the Monte-Carlo simulation approach. For each research problem, we implement the considered systems in Matlab and test the network performance by repeatedly simulate the network operation.
The performance data of the simulated network is collected, analysed and compared with the analytical approach.
1.5 Thesis Contributions
This section summarizes the main contributions of this thesis towards the design of energy efficient mobile networks.
• Contribution 1: Energy-efficient resource allocation in downlink multi-carrier NOMA systems
The first contribution of this thesis is to address the first research problem and to study the optimal sub-carrier assignment and power allocation in downlink multi-carrier NOMA systems to maximize the system’s energy efficiency, considering the constraints on users’
QoS. In this study, the optimal power allocation among users on one sub-carrier and the optimal power allocation across different sub-carriers are firstly investigated. And an op- timal power allocation algorithm is proposed to maximize the system energy efficiency.
The convergence of the proposed algorithm is proved. Furthermore, based on the findings on the optimal power allocation, a heuristic sub-carrier assignment scheme is proposed to minimize the required power to serve all users in the system. The performance of the proposed solutions is compared to the state-of-the-art solutions in the literature with nu- merical simulations. The results show that our proposed solutions can not only achieve higher energy efficiency but also reduce user blocking rate.
This contribution is reported in the following manuscript.
– Paper 1. Peiliang Chang and Guowang Miao, "Energy-efficient Resource Alloca- tion in Multi-carrier NOMA systems," to be submitted to IEEE Journal of Selected Topics in Signal Processing, special issue on signal processing advances for non- orthogonal multiple access in next generation wireless networks.
• Contribution 2: Interference-aware scheduling of base station transmission and micro- sleep.
The second contribution of this thesis is to develop a distributed solution for the schedul- ing of transmission and micro sleep of multiple base stations to reduce both inter-cell inter- ference and energy consumption. The complexity and the implementation of the proposed scheme are analysed. We evaluate the performance of the proposed scheme with simula- tions. The results show that the proposed scheme achieves performance close to that of the optimal solution based on exhaustive search. A detailed elaboration of this contribution can be found in the following paper:
