著者
李 允成
学位授与機関
Tohoku University
学位授与番号
11301甲第19355号
ΫϥυແઢΞΫηεωοτϫʔΫʹ͓͚Δ লిྗԽʹؔ͢Δݚڀ
A dissertation presented
by
Yunseong Lee
submitted to
Tohoku University
in partial fulfillment of the requirements
for the degree of
Doctor of Philosophy
Supervisor: Professor Nei Kato
Department of Applied Information Sciences
Graduate School of Information Sciences
Tohoku University
In recent years, the proliferation of smart mobile devices and emerging multime-dia applications is rapidly increasing mobile data traffic. Network densification has been proposed to deal with the demand from the higher volume of mobile data traffic. However, the performance accomplished through network densifi-cation is limited in terms of capacity and energy economy in the base stations. Besides that, mobile network operators (MNOs) are preparing to implement the Fifth Generation (5G) of mobile networks. To meet the 5G requirements, MNOs have introduced cloud radio access network (RAN) as network architecture. C-RANs consist of remote radio heads (RRHs), a centralized baseband unit (BBU) pool, and a fronthaul network. In this thesis, we focus on power saving schemes for C-RAN. Power saving schemes in C-RANs are implemented separately in the RRHs or the BBUs in existing studies. However, the separate configuration of RRH and BBU power saving schemes may increase the overall power consumption of C-RAN, which is still not addressed by researchers. Therefore, in this thesis, we investigate the tradeoff between the power consumption of RRHs and BBUs and propose a novel power saving scheme focusing both on the power consump-tion of RRHs and BBUs. To investigate this tradeoff, we formulate a theoretical model of RRHs and BBUs to estimate the number of active devices based on the traffic load. Based on the theoretical model, we propose an optimal traffic thresh-old to reduce the significant power consumption of C-RAN. Numerical analysis and simulation results reveal that our proposed optimal traffic threshold-based approach can reduce the total power consumption of C-RAN. Moreover, we apply the proposed scheme to time-varying traffics and prove through simulations that our proposed scheme can reduce the average power consumption of C-RAN.
Abstract i
1 Introduction 1
1.1 Background . . . 1
1.2 Research Objectives . . . 3
1.3 Thesis Outline . . . 4
2 Power Saving Schemes in Cloud Radio Access Network 5 2.1 Introduction . . . 5
2.2 Overview of Cloud Radio Access Network . . . 6
2.3 Existing Works Related to Power Saving in Cloud Radio Access Network . . . 9
2.4 Problems of Existing Works . . . 11
2.5 Summary . . . 12
3 Proposed Power Saving Scheme in Cloud Radio Access Network 13 3.1 Introduction . . . 13
3.2 System Assumptions and Definitions . . . 14
3.2.1 Network Model and Assumptions . . . 14
3.2.2 RRH Power Consumption Model . . . 18
3.2.3 BBU Power Consumption Model . . . 20
3.2.4 Total Power Consumption of the Network . . . 21
3.3 Formulation of Power Consumption of Cloud Radio Access Network 23 3.3.1 Expected Number of Active RRHs and Traffic Loads . . . 23
3.3.2.2 Second Allocation (SA) . . . 31
3.3.2.3 Total Expected Number of Active BBUs . . . 33
3.3.3 Expected Total Power Consumption . . . 35
3.4 Proposed Optimal Traffic Threshold . . . 36
3.5 Summary . . . 37
4 Performance Evaluation 39 4.1 Introduction . . . 39
4.2 Numerical Analysis . . . 40
4.3 Simulation Analysis . . . 45
4.4 Application to Time-Varying Traffics . . . 59
4.5 Summary . . . 69 5 Conclusion 71 Appendix 73 Copyright Permissions 97 Publications 99 References 103 Acknowledgments 109
2.1 Structure of a conventional BS in traditional RAN. . . 6
2.2 Comparison between traditional RAN and C-RAN. . . 7
2.3 Power saving scheme of RRHs through cell zooming. . . 9
2.4 Power saving scheme of BBUs through BBU aggregation. . . 10
3.1 Illustration of our considered network model. c 2019 IEEE . . . . 14
3.2 Sample configurations of a subarea. When the configuration pa-rameter θ increases, the number of RRHs in a subarea increases. c 2019 IEEE . . . 15
3.3 Two operation modes in a subarea: LC mode and SC mode. c 2019 IEEE . . . 15
3.4 Example of fRRH,LC(x) and fRRH,SC(x). c 2019 IEEE . . . 28
4.1 Log-normal distribution with identical parameter E(X) but differ-ing parameters σ. . . . 41
4.2 Numerical analysis results for E(X) = 60 % and σ = 0.1. c 2019 IEEE . . . 43
4.3 Examples of the results obtained during simulation analysis using E(X) = 60 % and σ = 0.1. c 2019 IEEE . . . 46
4.4 Numerically calculated Topt(dashed line at the x-shaped point) and simulation results (normal line at the circular point) of TTh that minimize the power consumption. Each panel exhibits different σ values. Circular points show the average values of 100 samples in each case, and range bars show the maximum and minimum values. c 2019 IEEE . . . 47
opt
simulation results (normal line at the circular point) of TTh that
minimize the power consumption. Each panel exhibits different
σ values. Circular points show the average values of 100 samples
in each case, and range bars show the maximum and minimum values. c 2019 IEEE . . . 48 4.4 Numerically calculated Topt(dashed line at the x-shaped point) and
simulation results (normal line at the circular point) of TTh that
minimize the power consumption. Each panel exhibits different
σ values. Circular points show the average values of 100 samples
in each case, and range bars show the maximum and minimum values. c 2019 IEEE . . . 49 4.4 Numerically calculated Topt(dashed line at the x-shaped point) and
simulation results (normal line at the circular point) of TTh that
minimize the power consumption. Each panel exhibits different
σ values. Circular points show the average values of 100 samples
in each case, and range bars show the maximum and minimum values. c 2019 IEEE . . . 50
4.5 Total power consumption of the network, simulated for different
mean traffic loads. Range bars present the maximum, average and minimum values of 100 samples. c 2019 IEEE . . . 52
4.5 Total power consumption of the network, simulated for different
mean traffic loads. Range bars present the maximum, average and minimum values of 100 samples. c 2019 IEEE . . . 53
4.5 Total power consumption of the network, simulated for different
mean traffic loads. Range bars present the maximum, average and minimum values of 100 samples. c 2019 IEEE . . . 54
4.5 Total power consumption of the network, simulated for different
mean traffic loads. Range bars present the maximum, average and minimum values of 100 samples. c 2019 IEEE . . . 55
to Topt. Without hysteresis margin (ΔT ), mode change may
oc-cur because of changes to Topt and power consumption increases
through the mode change, while with ΔT , mode change may not occur and power consumption increase is avoided. . . 60 4.8 Hysteresis margin and the probabilities of the modes at t. . . . 62
4.9 The graph of E(X) used in this simulation evaluation. . . . 63
4.10 The average number of active RRHs and mode switch for σ = 0.1. 65
4.11 The average power consumption of active RRHs and BBUs, mode switch, and the average total power consumption for σ = 0.1. . . . 66 4.12 The average total power consumption of the network, simulated
for different σ. Range bars present the maximum, average and minimum values of 100 samples. . . 68 A.1 The probability of the number of mode changes in a subarea
cal-culated from (A.19). . . 78 A.2 The probability of the number of mode changes in a subarea
sim-ulated through Monte Carlo simulation. . . 79 A.3 The number of mode changes in a subarea according simulation
results and numerical expected values. . . 80
A.4 Zoomed in ΔT range from 0 to 10 of Fig. A.3. . . . 81
4.1 Power Consumption Parameters for Performance Evaluation c 2019 IEEE . . . 41
Introduction
1.1
Background
In the past few years, mobile data traffic has grown significantly due to the in-creasing use of mobile devices and emerging multimedia applications, such as high-definition video streaming and mobile cloud computing [1]. The global mo-bile data traffic between 2017 and 2022 will increase three-fold according to a pre-diction from Cisco [2]. To deal with such a rapid growth, small cells have been deployed densely by mobile network operators (MNOs) [3]. The distances be-tween the base station (BS) and the users’ equipment (UEs) are reduced through these densely deployed small cells, which consequently improves the signal-to-noise ratio at the UEs, thus also improving their throughput. Moreover, dense BSs can spatially reuse the radio spectrum. Owing to these properties, small cells in a dense configuration can dramatically increase the network capacity. How-ever, the performance of densified networks is limited and the network capacity is unlikely to be improved in the near future [4]. Additionally, a high emission of greenhouse gases is caused by the power consumption of the network because of the high amount of small cells [5]. Moreover, the BSs consume approximately
To deal with the limitations of dense small cell networks and their high power consumption, MNOs are preparing to implement the Fifth Generation (5G) of mobile networks. 5G systems are supposed to provide a 1000-fold growth in system capacity, 10 Gbps maximum and 100 Mbps average individual user expe-rience, low-latency between 1 and 10 milliseconds, 100 billion devices connectivity, small Capital Expenditure and Operational Expenditure, and a 90 % reduction in network energy usage [7]. In order to meet these 5G requirements, cloud ra-dio access network (C-RAN) is expected to be a promising architecture for 5G realization [8].
In comparison to traditional RAN architecture, C-RAN has a centralized base-band unit (BBU) pool, which is different from conventional BSs that have to per-form both baseband processing and transmission/reception of radio signals [9]. Consequently, BSs in C-RAN only have to transmit/receive radio signals. In this context, they are called remote radio heads (RRHs). It is easier to deploy a large number of RRHs because RRHs have a simpler structure than traditional BSs. Moreover, flexible resource allocation and efficient interference management can be realized in C-RAN because of the centralized BBU pool. However, the higher data traffic demands for mobile networks from recent times necessitate a large number of network devices, such as BBUs in the BBU pool as well as RRHs. Although reducing network energy usage is a key requirement of 5G, increasing network devices eventually escalates the energy usage of networks. Therefore, power saving schemes in C-RAN are important to meet 5G requirements. Unless the devices’ (RRHs’ and BBUs’) power consumption is reduced at a circuit level, power saving schemes through control of the devices, i.e. active/sleep or on/off schemes, are important to reduce the power consumption in C-RAN. Hence, an efficient control scheme to save energy in C-RAN is necessary for future networks.
1.2
Research Objectives
In this thesis, we aim to analyze the power consumption of C-RAN and reduce its power consumption. Especially, we focus on power saving schemes of RRHs and BBUs, which are the main components of the C-RAN. Moreover, we consider power saving schemes of both RRHs and BBUs jointly, which have been consid-ered only in a few studies. To efficiently reduce power consumption in C-RAN, we propose a new power saving scheme with the three following steps.
1. Formulate a theoretical model of power saving schemes for RRHs
and BBUs.
2. Investigate the relationship between the power consumption of
RRHs and BBUs based on the theoretical model.
3. Propose an optimal configuration of power saving schemes for
C-RAN.
At first, we formulate a theoretical model of power saving schemes for RRHs and BBUs to analyze their relationship in this context. Because BBUs are in charge of baseband processing for the RRHs, the power saving scheme for RRHs may affect BBUs’ power consumption. Through active/sleep (or on/off) schemes, not only the number of active RRHs but also the traffic load carried by each RRH changes, which can change the active number of BBUs.
Next, we investigate the relationship between the power consumption of RRHs and BBUs based on the theoretical model. From the theoretical model, we cal-culate the expected number of active RRHs and BBUs numerically, and then we calculate the estimated power consumption of RRHs and BBUs.
Finally, we propose an optimal condition of power saving schemes of C-RAN. Through the findings from the first and second steps, we calculate the optimal
1.3
Thesis Outline
The remainder of this thesis is organized as follows.
The overview of C-RAN and related works on power saving in C-RAN are pre-sented in Chapter 2. Moreover, the problems of the existing works are introduced in this chapter.
Chapter 3 presents system assumptions and definitions of our study. Addi-tionally, we formulate a theoretical model of the system and propose a scheme to reduce the total power consumption of C-RAN.
Chapter 4 presents a numerical analysis and simulation results of our formu-lated theoretical model and proposed scheme. In addition, an application of our proposed scheme to time-varying traffics is introduced.
Power Saving Schemes in Cloud
Radio Access Network
2.1
Introduction
In this chapter, we describe the overview of C-RAN in comparison with traditional RAN systems. We focus on the structure of BSs in RAN, and then present the differences of these structures between traditional RAN and C-RAN. Additionally, we explain power saving schemes in C-RAN and describe the existing related works in this research area. Moreover, we present the problems of existing works in terms of power saving in C-RAN.
Some parts of the content in this chapter are presented in the following paper, which was written by the author of this thesis.
• Yunseong Lee, Keisuke Miyanabe, Hiroki Nishiyama, Nei Kato, and Takashi
Yamada, “Threshold-Based RRH Switching Scheme Considering Baseband Unit Aggregation for Power Saving in a Cloud Radio Access Network,”
2.2
Overview of Cloud Radio Access Network
As mentioned in Section 1.1, C-RAN has a centralized BBU pool, which is a dif-ferent structure from a traditional RAN. In this section, we present the structure of conventional BSs in a traditional RAN, and then present which part of the structure is changed in C-RAN and what are the merits of C-RAN.Fig. 2.1 shows a simple structure of traditional RAN and its BS. The main functions of a BS can be divided into two parts: baseband processing and trans-mission/reception of radio signals. The baseband processing module is called BBU and the transmission/reception module is called RRH. They are integrated inside a BS in a traditional RAN. A BBU has sub-functions of channel cod-ing/decoding, modulation/demodulation, resource-block mapping, fast Fourier transform, etc., while a RRH has sub-functions of digital processing, frequency filtering, and power amplification [9]. Because a BBU and a RRH are inside a BS together, a BBU can be connected to only a RRH in traditional RAN.
h
ZZ,
/ŶƚĞƌŶĞƚ
Ͳ tŝƌĞůĞƐƐƌĞƐŽƵƌĐĞĂůůŽĐĂƚŝŽŶ Ͳ DŽĚƵůĂƚŝŽŶΘĞŵŽĚƵůĂƚŝŽŶ ĂƐĞ^ƚĂƚŝŽŶh
tŝƌĞůĞƐƐ ĐŽŵŵƵŶŝĐĂƚŝŽŶdƌĂĚŝƚŝŽŶĂůZE
h
ZZ,
/ŶƚĞƌŶĞƚ
h
ZZ,
ͲZE
/ŶƚĞƌŶĞƚ
ZZ,
ZZ,
h
h
hƉŽŽů
BS and it is centralized in a BBU pool, as shown in Fig. 2.2. Because the BBUs are centralized in the BBU pool, they can interact with each other. Addition-ally, any BBU can connect to any RRH; also a BBU can manage several RRHs, which is not possible in the traditional RAN. Consequently, a C-RAN achieves flexible resource allocation and efficient interference management of RRHs, such as enhanced intercell interference coordination and multi-point coordination [10]. In addition to the above benefits, C-RAN is suitable for green networks. A traditional RAN utilizes BS energy on the following modules: power amplifier (PA), radio frequency (RF) small-signal transceiver module, baseband processor, DC–DC power supply, active cooling system, and AC–DC unit that connects the RAN to the electrical power grid [11]. However, because the RRH in C-RAN does not have a baseband processor, it has a simpler structure than a traditional BS. Therefore, it generates less heat and requires less energy for air conditioning. As air conditioning consumes 46 % of the cell site power supply [12], C-RAN will likely reduce the power consumption of the cell site. Additionally, a C-RAN requires fewer BBUs than a traditional RAN. In a traditional RAN, each BS has a BBU and a RRH; hence, the number of BBUs is equal to the number of RRHs. However, in a C-RAN, the BBUs and RRHs are placed at different sites, and the BBUs are centralized in a pool. Owing to the statistical multiplexing effect, this configuration reduces the number of required BBUs [13]. Therefore, the BBUs consume less power in total in a C-RAN than in a traditional RAN. Moreover, some BBUs in the BBU pool can enter the sleep state at times of low traffic load, e.g., during the night, further reducing the power consumption of the BBU pool. According to [14], the C-RAN architecture can save 67–80 % of the power consumed by the traditional RAN architecture.
2.3
Existing Works Related to Power Saving in
Cloud Radio Access Network
Existing works on power saving in C-RAN have focused on power conservation in either RRHs or BBUs independently. Most of the proposed power saving strategies, in both RRHs and BBUs, include active/sleep or on/off schemes.
In RRH power saving strategies, a RRH in which some of the modules are turned off is observed to be in the sleep state when the traffic load is light or zero [15]. Another effective power saving scheme is cell zooming, which dynami-cally changes the cell size according to the traffic conditions [16]. Fig. 2.3 shows an example of cell zooming. Before cell zooming, each RRH has a small coverage and there are 7 active RRHs, while after cell zooming, each active RRH has a large coverage and there are only 3 active RRHs. Although a zoomed-out RRH consumes large amounts of energy to guarantee the large coverage, the extra en-ergy consumption is considerably outweighed by the enen-ergy saving of sleeping RRHs. Therefore, this power saving scheme can reduce the power consumption of RRHs.
EƵŵďĞƌŽĨ
ĂĐƚŝǀĞZZ,Ɛ͗ϯ
EƵŵďĞƌŽĨ
ĂĐƚŝǀĞZZ,Ɛ͗ ϳ
^ůĞĞƉ
h
ϯϱй
h
Ϯϱй
h
ϯϬй
h
ϯϬй
h
ϯϱй
h
ϭϬй
h
ϭϬϬй
h
ϲϱй
h
Ϭй
h
Ϭй
h
Ϭй
h
Ϭй
ŐŐƌĞŐĂƚŝŽŶ
^ůĞĞƉ
Figure 2.4: Power saving scheme of BBUs through BBU aggregation.
BBU power saving strategies also implement a sleep state policy to reduce BBU power consumption when there is no traffic load. As the RRH traffic load can be assigned to any BBU centralized in the BBU pool, a BBU with a low load can offload that load to another BBU and be switched to the sleep state. This BBU power saving scheme, called BBU aggregation, improves the resource utilization along with the energy efficiency of the C-RAN [17]. Fig. 2.4 shows an example of BBU aggregation. Before BBU aggregation, each BBU has low utilization and 6 BBUs are active, while after BBU aggregation, each BBU has high utilization and only 2 BBUs are active. Because the number of active BBUs is reduced and the utilization is increased, BBU aggregation can reduce the power usage of BBUs.
2.4
Problems of Existing Works
Only a few studies on power saving in C-RAN have simultaneously considered the power consumption of both RRHs and BBUs. Wang et al. [18] jointly con-sidered the RRH antenna and BBU computational resources. They proposed an energy-efficient resource allocation scheme, but they investigated only the power consumption of the BBU pool and not the power consumption of the RRHs. Ye et al. [19] proposed a new BBU–RRH switching scheme with small delay and high quality of service, but they did not investigate the power consumption of the network operating with the switching scheme. Aqeeli et al. [20] optimized the allocation of computational resources between the RRHs and BBUs to waste less resources and reduce power consumption. However, they did not consider the power saving of RRHs. The power consumptions of both RRHs and BBUs in a C-RAN were analyzed in [21]. The authors investigated a new network traffic model based on queueing theory and showed that C-RAN evidently saves more energy than traditional networks. However, their analysis was related to the power-delay tradeoff and not to the above-mentioned power saving scheme.
In this thesis, we study the power consumption relationship between RRHs and BBUs. For instance, Zhang et al. [17] presented the relationship between power consumption of BBUs and cell size of RRHs. The authors concluded that the power consumed by the BBUs reduces as the size of the cell site decreases because a small cell can be more flexibly packed into a BBU than a large cell. Thus, we consider that this result is related to RRH power saving, particularly to cell zooming. When the number of active RRHs is reduced for power saving, the remaining active RRHs increase their coverage. Consequently, they can be less flexibly packed into a BBU and more active BBUs might be required. In this thesis, we address this potential tradeoff between the power consumption of
2.5
Summary
In this chapter, we explained the overview of C-RAN in comparison with tradi-tional RAN systems. In the C-RAN, a BBU can manage several RRHs, while a BBU manages only a RRH in the traditional RAN. Because of this structure, a C-RAN achieves increased capacity and decreases its power consumption. More-over, we presented existing works on power saving in C-RAN, which have focused on power conservation in either RRHs or BBUs. Additionally, only a few existing works have considered simultaneously the power consumption of both RRHs and BBUs. Therefore, in this thesis, we investigate the relationship of power con-sumption between RRHs and BBUs, and propose a new power saving scheme to reduce the total power consumption of C-RAN.
Proposed Power Saving Scheme
in Cloud Radio Access Network
3.1
Introduction
In this chapter, we present our considered network model and assumptions, the power consumption models of RRHs and BBUs, and the total power consumption of the network. Moreover, we propose the formulation of the expected number of active RRHs and BBUs based on stochastic methods and then calculate the ex-pected total power consumption of the network. Finally, we propose the optimal power saving scheme that reduces the total power consumption of the network.
Some parts of the content in this chapter are presented in the following paper, which was written by the author of this thesis.
• Yunseong Lee, Keisuke Miyanabe, Hiroki Nishiyama, Nei Kato, and Takashi
Yamada, “Threshold-Based RRH Switching Scheme Considering Baseband Unit Aggregation for Power Saving in a Cloud Radio Access Network,”
3.2
System Assumptions and Definitions
3.2.1
Network Model and Assumptions
A schematic of our considered network is shown in Fig. 3.1. We consider a C-RAN system as our considered network, and it is comprised of a BBU pool, a switch, a fronthaul network, and a network service area. There are NBBU BBUs in the BBU
pool and they can communicate with each other. LetB = {b1, b2,· · · , bk,· · · , bNBBU}
be the set of BBUs in the BBU pool, where bk denotes a single BBU in the pool. The network service area is comprised of regular hexagonal grid subareas that are arranged in i rows and j columns. Let A = {a1, a2,· · · , ah,· · · , aNa} be the set
of subareas in the network service area, where ah denotes a single subarea, and
Na denotes the number of subareas, i.e. Na= ij. RRHs are distributed in the network service area and connected to the BBU pool by the high-capacity front-haul network and the switch with no bottlenecks. Moreover, we are not focusing
Switch
Fronthaul
Network Network Service Area
ڭ ڭ ڮ a1 aij a(i-1)j+1 aj+1 aj a2j a(i-1)j+2 a2 aj+2 a(i-1)j+3 a3 aj+3 ڭ ڭ ڮ ڮ BBU Pool BBU BBU BBU BBU ڭ ڭ
Omni-directional RRH (OD RRH) Half-directional RRH (HD RRH) One-third-directional RRH (OTD RRH) 0000000 0000000 ș = 2 00000 00000 00000 00000 00000 00000 ș = 3
ڮ
ș = 4 ș = 5 ș = 6 ș = 7 00000 00000 00000 00000 00000 00000 00000 00000 0000 0000 00000 00000 00000 00000 00000 00000 00000 00000 0000 0000 00000000 0000 0000 00000 00000 0000 0000 00000 00000 0000 0000 0000 0000 0000 0000 0000 0000 0000 000000 00000000 000 0000 000 000 0000 0000 0000 0000 0000 0000 0000 șFigure 3.2: Sample configurations of a subarea. When the configuration param-eter θ increases, the number of RRHs in a subarea increases. c 2019 IEEE
Mode transition
by controlling
transmission power
00000 00000 00000 00000 00000 00000 00000 00000 00000 00000Active
Small coverage (SC)
mode
Large coverage (LC)
mode
00000 00000 00000 00000 00000 00000 00000 00000 00000 00000Sleep
dLC dSCFigure 3.3: Two operation modes in a subarea: LC mode and SC mode. c 2019
IEEE
on delay or energy loss in the frontfhaul network and switch. Hence, a BBU can accommodate several RRHs without the need for additional resources. However, we consider that one RRH cannot be allocated to multiple BBUs.
We introduce an architecture of subareas for RRH power saving with cell zooming. Fig. 3.2 shows sample configurations of the architecture, and Fig. 3.3 shows a schematic of RRH power saving with cell zooming. As shown in Fig. 3.2, there are three types of RRHs:
• Omni-directional RRH (OD RRH): An OD RRH covers a part of the
sub-area omni-directionally.
• Half-directional RRH (HD RRH): An HD RRH covers half the size of the
area that an OD RRH can cover.
• One-third-directional RRH (OTD RRH): An OTD RRH covers one third
of the size of the area that an OD RRH can cover.
Accordingly, we assume that those whole, half, one-third coverage size can be realized by changing design of the antenna of the RRHs [22], [23], horizontal antenna sectorization [24], [25], or beamforming technologies [26]. As shown in Fig. 3.3, a subarea has two modes: LC mode and SC mode. If a subarea is in LC mode, only one OD RRH at the center of the subarea is active and it covers the whole subarea, but the other RRHs are asleep. Let this active OD RRH be
a LC mode RRH and LC cell radius be dLC, which is equal to the circumradius
of the subarea. Conversely, if a subarea is in SC mode, all RRHs in the subarea are active. When the mode is changed from LC mode to SC mode, the LC mode RRH reduces its transmission power so that it reduces its cell radius. Let this SC cell radius be dSC. When changing the mode, the other asleep RRHs will be active
with a cell radius as big as dSC. Let all RRHs in SC mode be SC mode RRHs.
When the mode of the subarea is changed from SC mode to LC mode, all RRHs in the subarea will be asleep except for the OD RRH at the center of the subarea. The LC mode RRH will increase its transmission power and cover the whole subarea. We assume that the decrease in power consumption by putting RRHs in sleep mode is considerably larger than the increase in power consumption by increasing its transmission power. Accordingly, an LC mode subarea has a lower power consumption of RRHs than an SC mode subarea has.
The parameter θ as shown in Fig. 3.2 denotes a configuration parameter of the number of RRHs present in a subarea. It only assumes integer values and
satisfies θ ≥ 2. We consider that all subareas have the same value of θ in a network service area. Here, LC cell radius (dLC), SC cell radius (dSC), and θ have
the following relationship: dLC = θ× dSC. Additionally, let αθ, βθ, and γθ be the
number of OD RRHs, HD RRHs, and OTD RRHs in a subarea, respectively. The above parameters can be calculated by changing θ:
αθ = 1 + 6× θ δ=2 δ 3 , (3.1) βθ = 6× θ + 1 3 − γθ, (3.2) γθ = ⎧ ⎪ ⎨ ⎪ ⎩ 6 (θ mod 3 = 0), 0 (θ mod 3= 0), (3.3)
where x denotes the greatest integer less than or equal to x (floor function), and θ mod 3 denotes the remainder of θ divided by 3 (modulo operation). Let
NLC be the number of RRHs which can be in LC mode RRH in the network
service area, and NSC be the number of RRHs which can be in SC mode RRH in
the network service area. Then,
NLC = Na, (3.4)
NSC = (αθ+ βθ+ γθ)Na. (3.5)
Because the number of RRHs which can be in LC mode RRH is only 1 in a subarea, NLC should be equal to the number of subareas. Similarly, all RRHs in
a subarea can be in SC mode RRH, thus, NSC should be equal to total of RRHs
depending on the traffic loads. To decide the operation mode of the subarea, we introduce traffic threshold (TTh). Let TA ={Ta1, Ta2,· · · , Tah,· · · , TaNa} be the
set of traffic loads in the network service area, where Tahdenotes the traffic load in
ah. When Tah ≤ TTh, the subarea ah will operate in LC mode, while if Tah > TTh,
the subarea ah will operate in SC mode.
3.2.2
RRH Power Consumption Model
Let PRRH(d, Gt) be the power consumption of an active RRH, where d denotes the
cell radius of the RRH and Gt denotes its transmitting antenna gain. The power
consumption of an RRH depends on the antenna output power (Pout(d, Gt)) and
the RF small-signal transceiver module power (PRF) [27]:
PRRH(d, Gt) =
Pout(d, Gt)
η + PRF, (3.6)
where η denotes the PA efficiency. According to the Friis transmission equation,
Pout(d, Gt) is determined from the received transmission power at the cell edge
(PRX), the path loss (P L(d)), and Gt:
Pout(d, Gt) = PRX+ P L(d)− Gt. (3.7)
Note that Pout(d, Gt) in (3.6) is measured in Watts, but Pout(d, Gt) in (3.7) is
measured in dBm, which is a power unit in decibels (dB) with reference to 1 mW. In addition, PRX is measured in dBm, P L(d) is measured in dB, and Gt is
mea-sured in dBi, which is an antenna gain in dB with respect to an isotropic antenna. Here, we assume that the receiving antenna gain is 0 dBi. Hence, the receiving antenna gain is not mentioned in (3.7). For P L(d), we introduce the log-distance path loss model. According to the Shannon–Hartley theorem and [28], the capac-ity is affected by the signal-to-interference-plus-noise ratio (SINR), where SINR
depends on PRX and interference. Recent research works such as [28]–[30] have
already considered how to handle interference to guarantee the capacity of the net-work. Thus, we consider the best approach such as fractional frequency reuse [31] and inter-cell interference coordination [32] to manage the inter-cell interference.
Moreover, if PRX is the same at the cell edge both in LC and SC modes, then
the network can provide stable capacity in both modes. Hence, the optimization of SINR to guarantee the capacity is not considered; rather we investigate the power consumption of RRHs by considering their transmission power.
As the LC mode RRHs have a larger cell radius than the SC mode RRHs, the power consumption of an LC mode RRH is greater than that of an SC mode RRH. However, as mentioned in Section 3.2.1, the sum of power consumption of RRHs in an LC mode subarea should be less than that in an SC mode subarea. Let PRRH,LC be the power consumption of an LC mode RRH, and let PRRHOD ,SC, PHD
RRH,SC, and PRRHOTD,SC be the power consumption of an SC mode OD RRH, an
SC mode HD RRH, and an SC mode OTD RRH, respectively. Then,
PRRH,LC = PRRH(dLC, 0), (3.8)
PRRHOD ,SC = PRRH(dSC, 0), (3.9)
PRRHHD ,SC = PRRH(dSC, 3.01), (3.10)
PRRHOTD,SC = PRRH(dSC, 4.77). (3.11)
Additionally, let PRRH,sleep be the power consumption of an RRH in sleep state.
We assume that all types of RRHs consume the same amount of power in sleep state. According to all of our assumptions, all parameters related to RRH power
consumption should satisfy the following: PRRH,LC+(αθ+βθ+γθ−1)PRRH,sleep < αθPRRHOD,SC+βθP HD RRH,SC+γθP OTD RRH,SC. (3.12)
3.2.3
BBU Power Consumption Model
The power consumption of bk, denoted by Pbk, can be modeled as static or dy-namic depending on the baseband processing load [33], which is well approximated by a linear function of the traffic load [34]. Thus,
Pbk = ⎧ ⎪ ⎨ ⎪ ⎩ Pstatic+ PloadΔlbk (0 < Δlbk ≤ 1), PBBU,sleep (Δlbk = 0), (3.13)
where Pstatic is the static power consumption of a BBU, Pload is the coefficient
of dynamic power consumption, Δlbk is the traffic load processed by bk, and
PBBU,sleep is the power consumption of the BBU in the sleep state. When there
is no traffic load in bk, i.e., Δlbk = 0, bk will be in sleep state. Note that
PBBU,sleep < Pstatic. (3.14)
Whether bk is asleep or active is indicated by a binary variable zbk:
zbk = ⎧ ⎪ ⎨ ⎪ ⎩
0 (BBU bk is in sleep state), 1 (BBU bk is in active state).
(3.15)
In terms of zbk, Pbk can be revised as follows:
Pbk = (Pstatic+ PloadΔlbk)zbk + PBBU,sleep(1− zbk)
= PBBU,sleep+ (Pstatic− PBBU,sleep)zbk+ PloadΔlbkzbk.
3.2.4
Total Power Consumption of the Network
To calculate the total power consumption of the network, the number of active RRHs and BBUs must be known.
First, we calculate the total power consumption of the RRHs. The operation mode (LC mode or SC mode) in ah is indicated by a binary variable zah:
zah = ⎧ ⎪ ⎨ ⎪ ⎩ 0 (Tah ≤ TTh, i.e. LC mode), 1 (Tah > TTh, i.e. SC mode). (3.17)
Then, the power consumption of the RRHs in ah (Pah) is given as follows:
Pah = PRRH,LC+ (αθ+ βθ+ γθ− 1)PRRH,sleep
(1− zah) + αθPRRHOD ,SC+ βθPRRHHD ,SC+ γθPRRHOTD,SC zah.
(3.18)
The total power consumption of RRHs in the network service area (PRRH,total) is
given as follows: PRRH,total= Na h=1 Pah = PRRH,LC+ (αθ+ βθ+ γθ− 1)PRRH,sleep Na− Na h=1 zah + αθPRRHOD,SC+ βθPRRHHD ,SC+ γθPRRHOTD,SC Na h=1 zah. (3.19)
We note that Nh=1a zah is affected by the value of TTh.
The power consumption of the BBU pool (PBBU,total) is calculatd by adding the
as follows:
PBBU,total= NBBU
k=1 Pbk
= NBBUPBBU,sleep+ (Pstatic− PBBU,sleep) NBBU k=1 zbk + Pload NBBU k=1 Δlbkzbk. (3.20) Here, NBBU
k=1 zbk is the number of active BBUs and
NBBU
k=1 Δlbkzbk denotes the
total traffic load in the network service area. The power consumption of the BBU pool is mainly affected by these two parameters. Although the BBU aggregation affects NBBU
k=1 zbk, it does not affect
NBBU
k=1 Δlbkzbk because the aggregation does
not alter the total traffic load.
Therefore, the total power consumption of the network (Ptotal) is the sum of
(3.19) and (3.20):
Ptotal = PRRH,total+ PBBU,total. (3.21)
Note that we do not consider the power consumption of the fronthaul network and the switch.
3.3
Formulation of Power Consumption of Cloud
Radio Access Network
This section presents a theoretical formulation of the expected total power con-sumption of C-RAN and proposes an optimal traffic threshold that reduces the total power consumption of the network. First, we calculate the expected number of active LC mode RRHs and SC mode RRHs. Then, the traffic load transported by the active RRHs are calculated. Thereafter, we theoretically formulate the ex-pected number of active BBUs after aggregation and calculate the exex-pected total power consumption of C-RAN. Finally, this section proposes the optimal traffic threshold that feasibly minimizes the total power consumption of the network.
3.3.1
Expected Number of Active RRHs and Traffic Loads
In the theoretical formulation, the set TA follows a probability density function (PDF) f (x), where x denotes the traffic load and f (x) means the probability density of traffic load x within the considered subarea. We assume that the traffic load shared by SC mode RRHs have specific rates according to the RRH types. For example, let ahbe a subarea in SC mode, and let TOD
ah , T
HD
ah , and T
OTD ah
be the transported traffic load by an OD RRH, HD RRH, and OTD RRH in ah,
respectively, and let SOD, SHD, and SOTD be the coverage area of a SC mode OD
TaOD h = SOD αθSOD+ βθSHD+ γθSOTD Tah, (3.22) TaHD h = SHD αθSOD+ βθSHD+ γθSOTD Tah, (3.23) TaOTD h = SOTD αθSOD+ βθSHD+ γθSOTD Tah, (3.24)
where αθSOD+ βθSHD+ γθSOTD is the area of a subarea. Because SHD is equal to SOD/2 and SOTD is equal to SOD/3 as mentioned in Section 3.2.1, (3.22)–(3.24)
can be revised as follows:
TaOD h = rθTah, (3.25) TaHD h = rθ 2Tah, (3.26) TaOTD h = rθ 3Tah, (3.27) where rθ = 1 αθ+ βθ/2 + γθ/3. (3.28)
Additionally, we assume TMax ≤ TBBU/rθ, where TMax denotes the maximum
traffic load of TA and TBBU denotes the maximum traffic load capacity that can
be processed by a BBU. According to this assumption, a RRH transports traffic load less than or equal to TBBU to prevent traffic overload of the BBU processing
The expected number of active RRHs depends on two PDFs, fRRH,LC(x) and fRRH,SC(x), representing the probability density of the traffic load transported
by the individual LC mode RRHs and SC mode RRHs, respectively. As each
subarea accommodates one LC mode RRH, fRRH,LC(x) can be expressed:
fRRH,LC(x) = ⎧ ⎪ ⎨ ⎪ ⎩ f (x) (0≤ x ≤ TTh), 0 (x > TTh). (3.29)
However, as one subarea has αθ + βθ + γθ SC mode RRHs and they transport
different rates of traffic load, we need to represent the probability density of traffic
load transported by each type of SC mode RRHs to formulate fRRH,SC(x). Let
fOD
RRH,SC(x), fRRHHD ,SC(x), and fRRHOTD,SC(x) be the probability density of the traffic
load transported by an individual SC mode OD RRHs, SC mode HD RRHs, and SC mode OTD RRHs, respectively. Then,
fRRH,SC(x) = fRRHOD ,SC(x) + f HD
RRH,SC(x) + f OTD
f (x) within the range of TTh–TMax can be transformed as: fRRHOD ,SC(x) = ⎧ ⎪ ⎪ ⎪ ⎨ ⎪ ⎪ ⎪ ⎩ CODf x rθ (rθTTh < x≤ rθTMax), 0 (otherwise), (3.31) fRRHHD ,SC(x) = ⎧ ⎪ ⎪ ⎪ ⎨ ⎪ ⎪ ⎪ ⎩ CHDf 2x rθ rθ 2TTh < x≤ rθ 2TMax , 0 (otherwise), (3.32) fRRHOTD,SC(x) = ⎧ ⎪ ⎪ ⎪ ⎨ ⎪ ⎪ ⎪ ⎩ COTDf 3x rθ rθ 3TTh < x≤ rθ 3TMax , 0 (otherwise), (3.33)
where COD, CHD, and COTD denotes a coefficient. Because the traffic load in SC
mode subareas are shared by SC mode RRHs, (3.31)–(3.33) should satisfy the following equation: TMax TTh f (x)dx = rθTMax rθTTh fRRHOD ,SC(x)dx + rθTMax/2 rθTTh/2 fRRHHD ,SC(x)dx + rθTMax/3 rθTTh/3 fRRHOTD,SC(x)dx. (3.34)
(3.34) can be revised using (3.31)–(3.33) as follows: TMax TTh f (x)dx = COD rθTMax rθTTh f x rθ dx + CHD rθTMax/2 rθTTh/2 f 2x rθ dx + COTD rθTMax/3 rθTTh/3 f 3x rθ dx = COD TMax TTh f (x)rθdx + CHD TMax TTh f (x)rθ 2dx + COTD TMax TTh f (x)rθ 3dx = CODrθ+ CHD rθ 2 + COTD rθ 3 TMax TTh f (x)dx. (3.35) Thus, CODrθ+ CHD rθ 2 + COTD rθ 3 = 1. (3.36)
In the right-hand side of (3.34), the ratio of first term, second term, and third term should be the ratio of αθ, βθ, and γθ, which means the ratio of the number of SC mode OD RRHs, HD RRHs, and OTD RRHs. Therefore,
COD = αθ rθ(αθ+ βθ+ γθ), (3.37) CHD = 2βθ rθ(αθ+ βθ+ γθ), (3.38) COTD= 3γθ rθ(αθ+ βθ+ γθ). (3.39)
T
ThT
Max Traffic load x0
Probability Density fRRH,LC(x) fRRH,SC(x) Transform to fRRH,SC(x)Figure 3.4: Example of fRRH,LC(x) and fRRH,SC(x). c 2019 IEEE
Fig. 3.4 graphically demonstrates (3.29) and (3.30). When the range is from 0 to TTh, the transformed function fRRH,LC(x) is equal to f (x). Conversely, because
traffic loads in the range between TThand TMaxare shared by SC mode OD RRHs,
HD RRHs, and OTD RRHs, the width of the range of TTh–TMax becomes narrow
when f (x) is transformed to fRRH,SC(x). In addition, because the traffic load x may be shared by more than one SC mode RRHs, the value of fRRH,SC(x)
Using (3.29)–(3.33), we calculate the expected numbers of active each mode of RRHs to calculate the expected total power consumption of the network. Let
nLC, nODSC, nHDSC and nOTDSC be the expected number of active LC mode RRHs, SC
mode OD RRHs, SC mode HD RRHs, and SC mode OTD RRHs, respectively. They are given as follows:
nLC = NLC TTh 0 fRRH,LC(x)dx, (3.40) nODSC = NSC rθTMax rθTTh fRRHOD ,SC(x)dx, (3.41) nHDSC = NSC rθTMax/2 rθTTh/2 fRRHHD ,SC(x)dx, (3.42) nOTDSC = NSC rθTMax/3 rθTTh/3 fRRHOTD,SC(x)dx. (3.43)
Finally, the expected number of active RRHs is represented by a function
fn(x), and this function does not discriminate between LC mode and SC mode
RRHs or the types of RRHs.
fn(x) = NLCfRRH,LC(x) + NSCfRRH,SC(x). (3.44)
Using (3.44), we derive the expected number of active BBUs after aggregation in the next subsection.
3.3.2
Expected Number of Active BBUs after
Aggrega-tion
BBU aggregation is similar to the bin packing problem, which is a typical NP-hard problem [18], [33]. Therefore, we estimate the expected number of active BBUs after aggregation by assuming that the BBUs can accommodate the traffic load ideally.
3.3.2.1 First Allocation (FA)
In FA, the low and high traffic loads are compacted into one BBU. For example, if the low traffic load is x and the high traffic load is TBBU− x, these two traffic
loads can be compacted into one BBU. If the expected number of active RRHs carrying x and TBBU − x differ, the excess traffic loads will not be allocated in
FA.
To calculate the expected number of active BBUs in FA and the remaining traffic loads, we introduce a horizontal reflection function (fn(x)) of fn(x) over
x = TBBU/2 as:
fn(x) = fn(TBBU− x). (3.45)
Further, the expected number of remaining active RRHs with traffic load x can be determined as follows: g(x) = ⎧ ⎪ ⎨ ⎪ ⎩ fn(x)− fn(x) (fn(x)≥ fn(x)), 0 (fn(x) < fn(x)), (3.46)
where g(x) is the rest of the expected number of active RRHs that have traffic load x. If the remaining traffic load after FA is known, the expected number of
active BBUs in FA (nBBU,FA) can be calculated as: nBBU,FA = 1 TBBU TBBU 0 x fn(x)− g(x)dx, (3.47)
where fn(x)− g(x) is the expected number of active RRHs allocated by the FA; hence, the traffic loads distributed in FA are given by TBBU
0 x
fn(x)− g(x)dx.
Then, the expected number of active BBUs in FA (i.e. (3.47)) is determined as the traffic load allocated by FA divided by TBBU.
3.3.2.2 Second Allocation (SA)
In SA, we consider two types of traffic load: those that exceed TBBU/2 and
those below TBBU/2 . This is done because two RRHs transporting traffic loads
above TBBU/2 cannot be allocated to the same BBU. However, traffic loads below TBBU/2 can be allocated to the same BBU. Here, we assume that the traffic loads
below TBBU/2 are ideally allocated to the BBUs. For example, let Tex be the
sum of the traffic loads below TBBU/2. Then, the number of required BBUs for
accommodating these loads is Tex/TBBU.
SA proceeds according to the following steps. First, the expected number of active BBUs for allocating the RRHs exceeding TBBU/2, denoted by nBBU,over, is
calculated as follows:
nBBU,over =
TBBU
TBBU/2
g(x)dx. (3.48)
Thereafter, the remaining capacity of the BBUs (CBBU,left) that have already
accommodated RRHs in (3.48) is derived as:
CBBU,left= nBBU,overTBBU−
TBBU
TBBU/2
xg(x)dx. (3.49)
capacity of the BBUs. Referring to the remaining capacity of the BBUs, we determine whether additional BBUs are required. If the total traffic load of RRHs not exceeding TBBU/2 is at least equal to the remaining capacity of the
BBUs, i.e. if TBBU/2
0 xg(x)dx≥ CBBU,left, then the additional number of active
BBUs nBBU,additional is given by
nBBU,additional = 1 TBBU TBBU/2 0 xg(x)dx− CBBU,left . (3.50) However, ifTBBU/2
0 xg(x)dx < CBBU,left, then nBBU,additional is given as follows:
nBBU,additional = 0. (3.51)
Using (3.49), we can rewrite (3.50) as follows:
nBBU,additional = 1 TBBU TBBU/2 0 xg(x)dx− CBBU,left = 1 TBBU TBBU/2 0 xg(x)dx− CBBU,left TBBU = 1 TBBU TBBU/2 0 xg(x)dx− nBBU,over+ 1 TBBU TBBU TBBU/2 xg(x)dx = 1 TBBU TBBU 0 xg(x)dx− nBBU,over. (3.52) The conditions also can be revised in terms of (3.49). Let the condition function be Ψ(TTh), which is given by:
Ψ(TTh) = 1
TBBU
TBBU
0
Therefore, nBBU,additional can be revised as follows: nBBU,additional = ⎧ ⎪ ⎪ ⎨ ⎪ ⎪ ⎩ Ψ(TTh) Ψ(TTh)≥ 0 , 0 Ψ(TTh) < 0 . (3.54)
3.3.2.3 Total Expected Number of Active BBUs
The total expected number of active BBUs (nBBU,active) is the sum of the expected
numbers of active BBUs in FA and SA, i.e. the sum of (3.47), (3.48) and (3.54):
nBBU,active= nBBU,FA+ nBBU,over+ nBBU,additional. (3.55)
Note that nBBU,additional is conditionally changed; in turn, nBBU,active is also
simi-larly changed. If Ψ(TTh)≥ 0, then: nBBU,active
= nBBU,FA+ nBBU,over+ nBBU,additional
= nBBU,FA+ nBBU,over+ Ψ(TTh)
= 1 TBBU TBBU 0 x fn(x)− g(x)dx + nBBU,over+ 1 TBBU TBBU 0 xg(x)dx− nBBU,over = 1 TBBU TBBU 0 xfn(x)dx. (3.56)
On the other hand, if Ψ(TTh) < 0, then:
nBBU,active = nBBU,FA+ nBBU,over + nBBU,additional
= nBBU,FA+ nBBU,over + 0
= 1 TBBU TBBU 0 x fn(x)− g(x)dx + nBBU,over = 1 TBBU TBBU 0 xfn(x)dx− 1 TBBU TBBU 0 xg(x)dx + nBBU,over = 1 TBBU TBBU 0 xfn(x)dx− 1 TBBU TBBU 0 xg(x)dx− nBBU,over = 1 TBBU TBBU 0 xfn(x)dx− Ψ(TTh). (3.57) From (3.56) and (3.57),TBBU
0 xfn(x)dx is the total traffic load in the network
service area; this means that if TTh is involved in the condition Ψ(TTh)≥ 0,
then nBBU,active will not be changed by TTh. Otherwise, if TTh is involved in the
condition Ψ(TTh) < 0, (3.57) is greater than (3.56) which means an increase in nBBU,active because of Ψ(TTh) < 0.
3.3.3
Expected Total Power Consumption
The expected total power consumption of the network (Ptotal,expected) is decided by
the expected number of active/sleep RRHs and BBUs, and the active/sleep power consumption of the RRHs and BBUs. Therefore, by (3.19)–(3.21), (3.40)–(3.43), (3.56), and (3.57), Ptotal,expected is calculated as follows:
Ptotal,expected = nLCPRRH,LC+ NSC− nODSC − n HD SC − n OTD SC − nLC PRRH,sleep + nODSCPRRHOD ,SC+ nHDSCPRRHHD ,SC+ nOTDSC PRRHOTD,SC
+ NBBUPBBU,sleep + nBBU,active(Pstatic− PBBU,sleep)
+ Pload TMax 0 xf (x)dx = Na αθPRRHOD ,SC+ βθPRRHHD ,SC+ γθPRRHOTD,SC + nLC PRRH,LC+ (αθ+ βθ+ γθ− 1)PRRH,sleep − αθPRRHOD ,SC− βθP HD RRH,SC− γθP OTD RRH,SC + NBBUPBBU,sleep + nBBU,active(Pstatic− PBBU,sleep)
+ Pload
TMax
0
xf (x)dx,
(3.58) From (3.58), we note that nLC and nBBU,active are affected by TTh. When TTh
increases, nLC evidently increases, and nBBU,active may also increase because of
the increased traffic loads transported by each RRH are less flexibly packed into the BBUs. As assumed in (3.12) and (3.14), we expect a tradeoff between the power consumption of the RRHs and BBUs, i.e., the total power consumption of the network should be minimized at some optimal traffic threshold (Topt).
3.4
Proposed Optimal Traffic Threshold
Although the power consumption relationship between the RRHs and BBUs de-pends on their particular parameters, a BBU typically consumes more power than a small-cell RRH, which has small transmission power [21], [35]. Therefore, we consider that the total power consumption is more affected by the number of active BBUs than by the number of active RRHs.
As mentioned from (3.56) and (3.57), TTh does not alter the expected number
of active BBUs when Ψ(TTh)≥ 0. Conversely, TTh increases the expected number
of active BBUs when Ψ(TTh) < 0. Thus, the starting point for increasing the
expected number of active BBUs is Ψ(TTh) = 0, where TTh = Topt minimizes the
total power consumption.
Our procedure for finding Topt is based on the bisection method. The
pseu-docode is given in the Procedure 1. First, if Ψ(TTh)≥ 0 is satisfied by setting
Procedure 1 Finding the Optimal Traffic Threshold. c 2019 IEEE Input: f (x), Na, θ, TBBU, Output: Topt 1: TTh ← TBBU 2: if Ψ(TTh)≥ 0 is satisfied by TTh then 3: return TTh 4: else 5: a← 0, b ← TBBU 6: while |Ψ(TTh)| ≥ do 7: TTh ← (a + b)/2 8: if Ψ(TTh) > 0 is satisfied by TTh then 9: a ← TTh
10: else if Ψ(TTh) < 0 is satisfied by TTh then
11: b ← TTh
12: end if
13: end while
14: return TTh
TTh = TBBU, then Topt is TBBU because the expected number of active BBUs is
not increased by accommodating LC mode RRHs. Next, if Ψ(TTh) < 0 is satisfied
by setting TTh = TBBU, i.e., the expected number of active BBU is increased by
accommodating LC mode RRHs, we find the TTh satisfying|Ψ(TTh)| < through
the bisection method, where denotes the precision. The smaller the value, the more precise Topt will be. If the convergence is satisfactory, the value of TTh is
returned as Topt.
3.5
Summary
In this chapter, we presented our considered network model and assumptions. We described the power consumption models of RRHs and BBUs, and the total power consumption of the network. Based on the system model and assumptions, we formulated the expected active number of RRHs and BBUs through a stochastic method. Thereafter, we calculated the expected total power consumption of the network. We found that there is a starting point of the increase in the expected number of active BBUs, which is the optimal point that minimizes the total power consumption of the network. Finally, we proposed a procedure to find the optimal point to reduce the total power consumption of the network.
Performance Evaluation
4.1
Introduction
In this chapter, we evaluate our proposed scheme through numerical and simula-tion analyses. First, we evaluate the expected active number of RRHs and BBUs with our formulated equations and then calculate the expected total power con-sumption. Second, the proposed procedures for finding the optimal traffic thresh-old and total power consumption of the network are evaluated in a simulation analysis. Finally, we apply the proposed scheme to time-varying traffics and evaluate the average power consumption of the network in a simulation analysis. Some parts of the content in this chapter are presented in the following paper, which was written by the author of this thesis.
• Yunseong Lee, Keisuke Miyanabe, Hiroki Nishiyama, Nei Kato, and Takashi
Yamada, “Threshold-Based RRH Switching Scheme Considering Baseband Unit Aggregation for Power Saving in a Cloud Radio Access Network,”
4.2
Numerical Analysis
In the numerical analysis, we evaluated the expected number of active LC mode RRHs, SC mode RRHs, and BBUs. Then, the expected total power consumption of the network was calculated through our formulated equations. The evaluated network service area comprised 10×10 subareas, and the configuration of subareas
θ was 2, i.e., Na = 100, αθ = 1, βθ = 6, γθ = 0; hence, NLC = 100 and NSC = 700.
NBBU was set to 700. The traffic loads were expressed as percentages of BBU
capacity; therefore, TBBU was set to 100 %. For example, consider two subareas, a1 and a2, with traffic loads Ta1 = 60 % and Ta2 = 40 %, respectively. Both
loads can be allocated to one BBU. However, if the traffic load in subarea a3 is
Ta3 = 200 %, the subarea must operate in SC mode, and at least two BBUs are
required to accommodate Ta3.
The power consumption parameters of RRH and BBU are listed in Table 4.1. Active RRH power consumption is calculated through the equations of the RRH
power consumption model as mentioned in Section 3.2.2. We set dLC to 360 m,
dSC to 180 m, PRX to −100 dBm, PRF to 5.4 W, and η to 0.285. P L(d) is set to
36.7 log10(d) + 22.7 + 26 log10(fc) dB (d in m and fc in GHz),
and fc (the center frequency) to 2.5 GHz. Most of these parameters of RRH and BBU are referenced from [36]–[38], considering the urban micro non line-of-sight scenario and hexagonal cell layout in [36].
Table 4.1: Power Consumption Parameters for Performance Evaluation c 2019 IEEE
Power consumption of LC mode RRH in active state 7.10 W
Power consumption of SC mode OD RRH in active state 5.53 W
Power consumption of SC mode HD RRH in active state 5.47 W
Power consumption of SC mode OTD RRH in active state 5.44 W
Power consumption of all types of RRHs in sleep state 0 W
Static power consumption of BBU 120 W
Coefficient of dynamic power consumption of BBU 60 W
Power consumption of BBU in sleep state 3 W
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0 20 40 60 80 100 Probability Density Traffic Load [%] E(X) = 60 [%], σ = 0.1 E(X) = 60 [%], σ = 0.2
Figure 4.1: Log-normal distribution with identical parameter E(X) but differing parameters σ.
imated by a log-normal distribution [39]–[41]. Thus, the following log-normal distribution for the PDF of the traffic load is considered in our present network:
f (x) = √ 1 2πσxexp −(lnx− μ)2 2σ2 , (4.1)
where μ and σ are distribution parameters. The arithmetic mean E(X) of the traffic loads following f (x) is given as follows:
E(X) = exp μ + σ 2 2 , (4.2)
which itself is a function of μ and σ. Because the total traffic load NaE(X)
should be constant for any value of σ, we set the distribution by fixing E(X) and σ, then calculate μ to generate f (x). Fig. 4.1 shows an example of log-normal distribution. Note that as σ decreases, the traffic load in a given subarea will closely approximate E(X). The traffic load distribution parameters in the numerical analysis were set to E(X) = 60 % and σ = 0.1.
0 100 200 300 400 500 600 0 20 40 60 80 100
T
Th= 61.85%
Expected Active Number
Traffic Threshold [%]
LC mode RRHs SC mode OD RRHs SC mode HD RRHs BBUs
(a) Expected number of active LC mode RRHs, SC mode RRHs, and BBUs as functions of TTh
14 15 16 17 18 0 20 40 60 80 100
T
Th= 61.85%,
P
total,expected= 14.56kW
Expected Power Consumption [kW]
Traffic Threshold [%]
(b) Expected total power consumption of the network versus TTh
Figure 4.2: Numerical analysis results for E(X) = 60 % and σ = 0.1. 2019c
plots the expected number of active LC mode RRHs, SC mode RRHs, and BBUs, calculated by (3.40)–(3.43), (3.56), and (3.57), as functions of traffic threshold. We did not plot the expected number of active SC mode OTD RRHs because
there was no OTD RRHs in the evaluated network, i.e. γθ = 0. Increasing the
traffic threshold increases the number of active LC mode RRHs and decreases the number of active SC mode RRHs. The number of active BBUs is unchanged for
TTh up to 61.85 % but increases from this value onward. When TTh = 61.84 %,
Ψ(TTh)≈ 0.0494, which satisfies Ψ(TTh) > 0. However, when TTh = 61.85 %,
Ψ(TTh)≈ −0.0112, which satisfies Ψ(TTh) < 0. From these results, we can
con-clude that the point at which TTh satisfies Ψ(TTh) = 0 changes the value of the
above formula from positive to negative, which is the starting point in the
in-crement of the number of active BBUs. When TTh exceeds 61.85 %, the number
of active BBUs approaches the number of active LC mode RRHs because
al-most all of the LC mode RRHs transport traffic loads are above TBBU/2 = 50 %.
Consequently, only one LC mode RRH can be allocated to one BBU.
Fig. 4.2 (b) shows the expected total power consumption of the network as a function of traffic threshold. The expected total power consumption (Ptotal,expected)
is minimized when TTh is 61.85 %, which is the same as the start point of
in-creasing the number of active BBUs. Therefore, our formulated equations can calculate the optimal traffic threshold that minimizes the power consumption of the network.
4.3
Simulation Analysis
In the simulation analysis, we generated a number of random samples following a log-normal distribution and evaluated how our proposed optimal traffic threshold reduces the power consumption of the network. The parameters related to power consumption were unchanged from the numerical analysis. The precision in Procedure 1 was set to 0.1. The simulations were programmed in Python 3.6 and the BBUs were aggregated using the general bin packing algorithm in [42]. We randomly generated 100 samples for each distribution by varying E(X) as 60 %, 65 %, 70 %, 75 %, and 80 %, and σ as 0.05, 0.1, 0.15, and 0.2. Note that we consider time-homogeneous traffic loads, which means that the spatial traffic distribution is not time-varying.
