RECEIVER ARCHITECTURES AND ALGORITHMS FOR NON-ORTHOGONAL MULTIPLE ACCESS
Talgat Manglayev
A thesis submitted in partial fulfillment of the requirement of Nazarbayev University for the degree of Doctor of Philosophy
May 2020
i
Abstract
Multiple access (MA) schemes in cellular systems aim to provide high throughput to mul- tiple users simultaneously while utilising the network resources efficiently. Traditionally, each user in the network is assigned a fraction of resources (such as slots in time or fre- quency) to operate so that multi-user interference is avoided. These schemes are named as ‘orthogonal multiple access’ (OMA) and are the basis of most cellular standards – from the earliest first generation up to the current fourth-generation systems. Non-orthogonal multiple access (NOMA) on the other hand is a novel method that allows all the users in the network to operate in the entire available spectrum at the same time which enables significant improvement in the system throughput.
While providing increased throughput, NOMA requires high computational power in order to implement sophisticated interference cancellation algorithms at each user termi- nal, as well as power allocation schemes at the base station. As a potential candidate for the fifth-generation networks (5G), NOMA must meet certain requirements, and com- putational efficiency is essential for reduced latency. Recently graphics processing units (GPUs), which were initially intended for outputting images to display, appeared as an alternative to multi-core central processing units (CPUs) for general-purpose computing.
GPUs have thousands of cores with approximately three times less frequency than a CPU core. With their numerous advantages in executing heavy and time-consuming computa- tions in parallel, GPUs have become attractive platforms in a variety of fields.
The overall aim of this research is to significantly increase the scientific understanding and technical knowledge on NOMA. This is achieved by exploring and developing novel methods, models, designs and techniques that will facilitate the implementation of NOMA for future generation networks.
ii First, the achievable data rates for individual users are demonstrated in a success- ful interference cancellation (SIC) based NOMA network. These results were compared against the conventional orthogonal MA schemes with optimum power allocation and varying fairness. In addition, a further investigation was carried out into the deficiency of SIC receivers which can occur when a user in the networks attempts to decode other users’ signal. Presented in the analysis is the findings from the experimental process where the decoding order of a user with a mismatched signal was observed as well as the significant impact on the computation time. The decoding time-difference between cor- rect and mismatched decoding order as a detection method of deficiency or fraudulence in the network is then discussed. Next, a comparison is presented between the compu- tational times of the SIC receiver with another popular interference cancellation scheme named ‘parallel interference cancellation’ (PIC). This was done using different platforms specifically for an uplink NOMA system. The results showed that the computation time of PIC scheme is significantly lower than SIC on the GPU platform even for a very large number of available users in the network. Then, the execution time of NOMA with SIC in the uplink of a cellular network with user clustering was examined. User clustering is a popular method in NOMA networks that eases the sophisticated resource allocation and network management issues. While most works found in the literature review concentrate on the joint optimisation of user grouping and resources, this research project focused on processing the signal detection of each cluster in parallel on the GPU platform at the base station. Following this, parallel interference cancellation (PIC) was implemented and compared with the existing SIC on both CPU and GPU platforms for uplink NOMA- OFDM. Architectures of the receivers were modified to fit into parallel processing. GPU was found applicable to speed up computations in NOMA based next-generation cellu- lar networks outperforming up to 220 times SIC on CPU. Finally, the research presents the power allocation problem from artificial intelligence (AI) perspective and propose a method to predict the power allocation coefficients in a downlink NOMA system. The results of the research show a close-to-optimal sum rate with about 120 times reduced
iii computation time. The achieved results decreases the network latency and assist NOMA to meet 5G requirements.
iv
Contents
1 Introduction 1
1.1 A Brief History of Wireless Cellular Networks. . . 1
1.2 Potential Techniques of Future Radio Access . . . 4
1.3 The Recent Progress of Computation . . . 5
1.4 Aims and Objectives . . . 6
1.5 List of Publications . . . 6
1.6 Thesis Organization . . . 7
2 Background and Preliminaries 10 2.1 Evolution of Mobile Radio Access . . . 10
2.2 Basics Concepts of NOMA . . . 12
2.3 Interference Cancellation . . . 13
2.3.1 Successive Interference Cancellation (SIC) . . . 14
2.3.2 Parallel Interference Cancellation (PIC) . . . 16
2.4 Power Allocation in NOMA . . . 17
2.5 User Clustering in NOMA . . . 18
2.6 Parallel Programming . . . 20
2.7 Artificial Intelligence in Future Networks . . . 24
3 NOMA with SIC and PIC 26 3.1 NOMA Optimum Power Allocation . . . 26
3.1.1 Introduction and Related Works . . . 26
3.1.2 System Model . . . 28
v
3.1.3 Numerical Results and Discussion . . . 32
3.2 Security Threat Detection for SIC . . . 35
3.2.1 Introduction and Related Works . . . 35
3.2.2 System Model . . . 36
3.2.3 Numerical Results and Discussion . . . 37
3.3 NOMA with SIC and PIC . . . 39
3.3.1 Introduction and Related Works . . . 39
3.3.2 System Model . . . 40
3.3.3 Numerical Results and Discussion . . . 41
3.4 Chapter Summary . . . 45
4 GPU Accelerated NOMA 47 4.1 User Clustering in SIC NOMA . . . 47
4.1.1 Introduction and Related Works . . . 47
4.1.2 System Model . . . 48
4.1.3 User Clustering in Uplink NOMA . . . 49
4.1.4 GPU Implementation . . . 51
4.1.5 Numerical Results and Discussion . . . 54
4.2 PIC and SIC for OFDM-NOMA . . . 57
4.2.1 Introduction and Related Works . . . 57
4.2.2 System Model . . . 58
4.2.3 OFDM based SIC and PIC in NOMA . . . 60
4.2.4 GPU Implementation . . . 61
4.2.5 Numerical Results and Discussion . . . 63
4.3 Chapter Summary . . . 67
5 Artificial Intelligence in NOMA 69 5.1 Introduction and Related Works . . . 69
5.2 System Model . . . 71
vi
5.3 Data Preparation . . . 72
5.4 AI Implementation . . . 73
5.5 Computational Complexity Analysis . . . 74
5.5.1 Complexity of Brute-Force Algorithm . . . 75
5.5.2 Complexity of Normal Equation . . . 75
5.5.3 Complexity of Deep Learning Model . . . 75
5.6 Numerical Results and Discussion . . . 76
5.7 Chapter Summary . . . 78
6 Conclusion and Future Research 81 6.1 Conclusion . . . 81
6.2 Future Research . . . 83
6.2.1 Short Term . . . 83
6.2.2 Long Term . . . 83
Bibliography 85
1
Chapter 1
Introduction
This chapter deals with a history of wireless cellular networks (Section 1.1); potential techniques of the future radio access (Section 1.2); history of recent computation (Sec- tion 1.3); novelty, contribution, motivation, aims and objectives of the work (Section 1.4) and finally the thesis organisation (Section 1.6). Specifically, the history section carries background information of the thesis and narrows down to the topic focus. Previously used and potential techniques are given as background information. Similarly, we dis- cussed progress and varieties of computing processors then we describe the main idea of the thesis by presenting novelty, contribution, hypothesis, aim and objectives of the work.
Finally, the thesis organisation is given.
1.1 A Brief History of Wireless Cellular Networks
Wireless communications as a branch of information and communication technologies have undergone a rapid development tendency. According to Guturu [1], a new gener- ation of cellular networks traditionally launches at the beginning of each decade since 1980s. First generation (1G) which introduced standard voice calls started in 1982. The voice calls used the frequency division multiple access (FDMA) technique in the physical layer. They were digitally encrypted in the second-generation (2G) a decade later. Digi- tal signals introduced new services such as sending multimedia and short text messages.
Chapter 1. Introduction 2 Moreover, 2G achieved spectral efficiency of available bandwidth. The 2G standards im- plemented the time division multiple access (TDMA) scheme in Europe (as in GSM) and the code division multiple access (CDMA) scheme for North America (as in cdmaOne).
New possibilities such as mobile TV, data services and video calls appeared later in the third generation (3G) networks. The 3G Partnership Project (3GPP) and the International Mobile Telecommunications (IMT2000) launched standardisation towards 3G in 1998.
The latter required 200 kbits/s data rates as standard to be called ‘3G’. The Universal Mobile Telephone System (UMTS) and the CDMA 2000 became the 3G standards. They began in Europe, Japan and China in 2001 and were later launched in North America and Korea in 2002. Further improvements in 3G resulted in the release of 3.5G and 3.75G which made broadband internet available on a massive scale by providing data rates of up to 14.7 Mb/s for smartphones and mobile modems. Fourth generation (4G) started in 2009 with two systems: mobile World Interoperability for Microwave Access (WiMAX) based on IEEE 802.16 standards and long-term evolution (LTE) standards. Mobile services were extended with internet access, internet protocol telephony, high-definition mobile TV, video conferencing and video games in 4G. High data rates attracted a large number of connected devices and the appearance of high bandwidth applications.
At the moment both the industry and academia are engaged in researching the re- quirements and solutions for the next generation (5G). The main technical challenges are:
to achieve a data rate between 1 to 10 Gb/sec, reduce latency to less than 1 millisecond (ms), extend bandwidth for a large number of devices with full availability and cover- age and finally reduce energy consumption. These goals for 5G are based on insights of telecommunication companies such as Nokia, Huawei and Docomo. The enterprises are aiming for massive connection of devices, internal communications, ultra-high-speed and deployment of the efficient spectrum on top of enhancement in all existing network fea- tures. The Internet of Things (IoT) and Machine-to-Machine (M2M) communications are examples of such massive connectivity and offer a variety of network services. For exam- ple, ultra-high-speed throughput will make possible virtual reality, high definition video
Chapter 1. Introduction 3 calls and 4K video broadcasting. Similarly, the deployment of the efficient spectrum will enable the launch of demanding applications and a large number of equipment per user with a variety of complicated scenarios. Such predictions are based on the evolution of previous generations and related technologies e.g. mobile hardware, applications etc.
Wireless communication systems are going to change ordinary daily life in various areas including health, industry and logistics [2–4]. Complex cellular network architec- ture and design with novel technologies cover every layer and promise to reach the set objectives. Some of the on-going and expected advancements for cellular networks are briefly mentioned in this paragraph. One of the simulations and tests of 5G with com- mercial instruments allows recognition of data flow and configuration of the cell with different parameters [5]. Cell design involves advanced technology like centralised radio access network (C-RAN) [6] and multiple-input multiple-output (MIMO) [7]. As well as the graphics processing unit (GPU) cluster is offered [8] to outsource complex gaming computations for mobile devices.
Radio frequency selection is a challenge for industry and scholars whereas its allo- cation remains an issue for the governments. On the one hand frequency bands higher than 29 GHz require deployments of lots of base stations, whereas lower frequency bands may be already reserved for military, rescue or other needs. For example, the expected frequency bands for Europe in the 5G as mentioned in [9] were below 5 GHz. Massive machine communication expected in 5G has another challenge in realisation. Preliminary requirements allow from 10 to 100 times more connected devices and long-life battery for more than 5 years with lower costs and 99% coverage. However, the issue is in the number of machines that are not widespread enough for such needs [10]. The multiple access scheme is established as a distinguishing physical layer feature of any wireless communication network. Orthogonal multiple access techniques have been implemented in wireless cellular networks from the beginning and up to the current fourth genera- tion [11].
Chapter 1. Introduction 4
1.2 Potential Techniques of Future Radio Access
One of the proposed schemes to meet 5G requirements discussed above provides multiple connections either in-power or in-code domain and remains non-orthogonal in time and frequency [12]. Examples of non-orthogonal multiple access (NOMA) schemes in the power domain are based on power level differences of signals of the connected devices.
On the other hand, NOMA in code domain assigns a unique code to the signal which is used to demodulate the message at the receiver side. Since non-orthogonality intro- duces interference among the signals the receiver requires an interference cancellation technique. It must be reliable, practical and most of all, it must meet 5G requirements.
Successive interference cancellation (SIC) and parallel interference cancellation (PIC) techniques are offered with NOMA within the power domain in the literature. The SIC decodes the signals of each user sequentially. It starts by decoding the signal that belongs to the user with the highest power, then subtracts it from the received signal, and then it iterates the same process for the second-highest power signal. In PIC, which is an alternative way, the signals of all users are decoded concurrently [13], [14]. NOMA’s performance promises to meet the 5G requirements which is what made it ‘trendy’ among candidate technologies, however, feasibility challenges are preventing the deployment of the scheme. One of them is controlling the cellular network during the different level of traffic load. Also, the reliability requires correct interference cancellation, for example, in every iteration in SIC receiver. The work in [15] indicated key obstacles of NOMA deployment in 5G. Firstly, the authors argue that the power domain is still on an infancy research level rather than the practical one. Secondly, the signal processing technology of the chip at the receiver side tends to require design improvements as SIC computation is still complex.
Chapter 1. Introduction 5
1.3 The Recent Progress of Computation
Computation cost has been a challenge for both software and hardware engineers. Soft- ware engineers keep reducing the number of instructions and search for creative algo- rithms to reduce the running time of a program, whereas hardware engineers devise so- lutions that save physical resources. Blelloch in [16] describes a parallel programming concept and programming language named “the NESL" which allows parallel operations that reduce the number of operations. Massive pioneer hardware solutions of parallel pro- gramming appeared a decade later than the idea. ®Core™ Duo processors introduced by Intel Inc. in 2006 offered architecture with two cores and shared cache memory. The chal- lenges involved virtualisation, power and temperature control as described in [17]. The latest huge core i9 processors of the 9th generation by Intel Inc. have up to 18 physical cores with 3.0 GHz frequency [18]. Multi-threaded programs are needed in order to use those multiple cores to realise tangible performance benefits. There are various common programming languages such as assembler, C, Java, C# etc., where multi-threading is possible [19]. If the number of initiated threads exceeds the number of cores then threads will not be executed simultaneously.
The processing speed of a graphics processing units (GPU) core is usually approxi- mately two to three times less than the frequency of a CPU core. For example, one of the latest models of NVIDIA GPU RTX 2080 has a 1515 MHz clock speed. In contrast, the number of cores in GPU is counted in hundreds, so the same RTX 2080 has 2944 CUDA cores against 18 cores in the latest produced CPU as mentioned above. Hardware char- acteristics of GPU allow general-purpose computing and high-performance computing (HPC) using the advantage of numerous multiple cores upon parallel algorithms.
Indeed, digital signal processors (DSP) are naturally suitable devices for encoding and decoding. Moreover, they are flexible and suit to a number of operating systems. The processors may be built-in to both the trendy IoT devices and to the BS. Hence, they fit both downlink and uplink channels. However, the classic DSPs are obsolete and are not able compete with modern GPUs with a major parameter like GFLOPS [20]. The DSP
Chapter 1. Introduction 6 enthusiasts lost the computation competition to GPU because of the ability to run tasks in parallel of the latter one [21].
In [22] first HPC cluster with 30 GPUs for scientific computation was assembled and demonstrated. Some of the most popular platforms which enable general-purpose computing and HPC are OpenCL from Intel®, CUDA from NVIDIA® and MATLAB® parallel computing toolbox from Mathworks. This leads to a speedup of heavy and time- consuming research simulations and computations in different areas. Different ways of accelerating NOMA related techniques with parallel programming using multiple CPU and GPU threads are described in this thesis.
1.4 Aims and Objectives
This thesis aims to increase the scientific understanding and the technical knowledge on non-orthogonal multiple access (NOMA) systems by exploring and developing novel methods, models, designs and techniques that will facilitate the implementation of future radio access. Towards this aim the following objectives are pursued:
1. To develop a novel interference cancellation mechanism for improved capacity.
2. To involve GPU for multi-user decoding in NOMA networks.
3. To investigate the feasibility of parallel programming approaches for multi-user decoding.
1.5 List of Publications
Journal Publications
[1] T. Manglayev, R. C. Kizilirmak, N.A.W.A. Hamid “GPU Accelerated PIC and SIC for OFDM-NOMA"Electronics, vol. 8, no. 3, pp. Feb. 2019.
Chapter 1. Introduction 7 [2] T. Manglayev, R. C. Kizilirmak, Y.H. Kho, N.A.W.A. Hamid “GPU Accelerated Suc- cessive Interference Cancellation for NOMA Uplink with User Clustering",Wireless Per- sonal Communications, vol. 103 no. 3, pp. 2391-2400, Dec. 2018.
Conference Publications
[1] T. Manglayev, R.C. Kizilirmak, Y.H. Kho “Comparison Of Parallel And Successive Interference Cancellation For Non-Orthogonal Multiple Access"Comparison Of Parallel And Successive Interference Cancellation For Non-Orthogonal Multiple Access, Astana, 2018 pp. 74-77.
[2] T. Manglayev, R.C. Kizilirmak, Y.H. Kho, N. Bazhayev, I. Lebedev “NOMA with imperfect SIC implementation", IEEE EUROCON 2017 -17th International Conference on Smart Technologies, 2017, Ohrid, pp. 22-25.
[3] T. Manglayev, R.C. Kizilirmak, Y.H. Kho “Optimum power allocation for non-orthogonal multiple access (NOMA)", 2016 IEEE 10th International Conference on Application of Information and Communication Technologies (AICT), Baku, 2016, pp. 1-4.
1.6 Thesis Organization
There are six chapters in the thesis. The first chapteris the introduction to the thesis. It discusses the historical milestones that led to the current stage of wireless communications and discusses a forward view of future radio access. The chapter also states the aim and objectives of the research and describes the contributions of the work.
Thenext chapterpresents preliminary and fundamental work which is relevant to the research. Firstly, works related to multiple access schemes are reviewed followed by a re- view of NOMA. After discussing the basics of NOMA in both uplink and downlink chan- nels, the two most popular interference cancellation techniques were introduced, namely
Chapter 1. Introduction 8 successive interference cancellation (SIC) and parallel interference cancellation (PIC).
Next, power allocation and its vital role in power domain NOMA are discussed followed by user grouping in the cellular systems. Then, we switch our focus to parallel program- ming concept. Examples of advanced engineering solutions are offered where parallel programming is applied to wireless communications. Finally, the recent role of artificial intelligence (AI) in the upcoming mobile systems are presented with a particular focus on 5G.
Chapter three has three subsections. The first topic is optimum power allocation for NOMA networks with SIC receiver. Power allocation coefficients are obtained to reach the maximum sum rate for the users with a predefined fairness constraint. The sum- rate results of NOMA and OMA are compared and the superiority of NOMA is demon- strated. The second subsection discusses the potential threat in the receivers and discusses a method that identifies it. Decoding time for the users in a network is proportional to their distance to the base station. The possible attack can be detected by tracking the decoding times of the users in the network. The third part studies NOMA with SIC and PIC re- ceivers in the uplink channel. These two receivers are the most common in the literature and their decoding times will be measured and compared using different platforms: MAT- LAB using ordinary CPU; Java programming language that runs multithreaded CPU and CUDA platform with GPU.
Chapter four proposes solutions for enhancing the computation time of the NOMA receiver for reduced latency. The first part considers user clustering for NOMA with SIC in the uplink channel. The second part focuses on OFDM-NOMA with SIC and PIC receivers. Both parts present the receivers implemented on a GPU device as an alternative to CPU and present the results for comparison.
Chapter five demonstrates machine learning (ML) and deep learning (DL) from the AI family to assist the power allocation mechanism in NOMA scheme. Numerical results evidence that AI-enabled power allocation gives very close to optimum power allocation in terms of sum capacity and the execution times of ML and DL is much faster than
Chapter 1. Introduction 9 exhaustive search.
Chapter sixconcludes the thesis and discusses potential future works that may further develop the idea. The long-term and short-term perspectives are presented separately.
FIGURE1.1: Info graphics of the main parts of the thesis.
10
Chapter 2
Background and Preliminaries
This chapter deals with the research and findings relating to multiple access schemes in general (Section 2.1), NOMA in power domain (Section 2.2), interference cancellation in NOMA (Section 2.3), NOMA user clustering (Section 2.4) and parallel programming (Section 2.5). Specifically, multiple access schemes in earlier cellular networks and vari- ants of NOMA which have the potential to be implemented into 5G are reviewed. The SIC and PIC decoding schemes at the NOMA receiver are also discussed. Following this, literature about user-clustering in NOMA that is grouping the users in a cell for improved performance is then reviewed. Finally, the focus is switched to the computation part of the topic and a review of the literature is offered to discuss the parallel programming on multi-threaded CPU and GPU wireless cellular networks.
2.1 Evolution of Mobile Radio Access
A wireless cellular network is a one that has distributed base stations over a region each serving several mobile users connected to them. The communication between a mobile terminal and a base station is two-way, i.e. there is a dedicated uplink channel for mobile terminals to transmit as well as a downlink channel for base stations to transmit their data to the mobile users. In both uplink and downlink, the common channel is shared by multiple users. Multiple access schemes address the way the overall resources of the common channels are shared among the users within the network [23]. Chapter I
Chapter 2. Background and Preliminaries 11 noted the cellular network’s standards that have been implemented so far. The multiple- access techniques that have been used in those standards enabled the first voice call to be launched in 1981 which was based on FDMA. Nordic Mobile Telephone (NMT) system by Ericsson AB achieved a combination of several cells and released the world’s first cellular network. The system remained analogue due to radio transmission technology even though the network had digital switching technology [24]
The cellular technology then experienced rapid developments and many digital tech- nologies were implemented. In particular, the multiple access techniques over the years experienced the most dramatic changes. For example, in 2G, TDMA and FDMA were used in GSM. In the third generation cellular networks such as W-CDMA and UMTS, direct sequence CDMA (DS-CDMA) were used and the receiver employed the Rake receiver to counter multipath fading. Orthogonal multiple access (OMA) based on or- thogonal frequency division multiple access (OFDMA) or its single carrier counterpart SC-FDMA was adopted in the 3.9G and 4G networks such as LTE [25] and LTE Ad- vanced [26] [27]. These approaches rely on the idea that the common channel is parti- tioned and shared among the users either in time, frequency and code domain orthogo- nally, i.e., without overlapping each other hence the name ‘orthogonal multiple access’
(OMA). These technologies were successful to meet the demand for mobile services at the time.
In order to maintain the sustainability of wireless cellular networks during the next decade, novel solutions which will face the challenges are required [28]. Considering the 1000-fold leap in the size of mobile traffic, another advancements such as network capac- ity, quality of service and better user experience are needed, The cost-effective network capacity improvements are reached by smart design of radio access technology. Non- orthogonal multiple access (NOMA) scheme is proposed to accomplish those goals.
NOMA implements a novel approach for multiplexing users. Unlike in the previ- ous generations of cellular networks, NOMA multiplexes users in power domain and let them operate in the same frequency and time. In NOMA, users are de-multiplexed at the
Chapter 2. Background and Preliminaries 12 receiver side using advanced interference cancellation techniques such as a successive in- terference canceller (SIC) [29], [30]. From the perspective of information theory, NOMA with ideal SIC is an optimal multiple access technique in terms of achievable user data rates both in uplink [31] and downlink channels [32].
2.2 Basics Concepts of NOMA
This section introduces NOMA in the power domain and discusses it as a candidate of 5G. Interest around NOMA trend can be recognised in academic journals and conference proceedings approximately from 2012. The study continues both in academia and indus- try. As highlighted earlier in this paper, multiple access schemes share resources so that users in the cell simultaneously and actively connect to the base station. Massive usage of mobile devices and the popularity of the internet lead to a shortage of these natural resources such as frequency and time. NOMA in the power domain with an accurate in- terference cancellation at the receiver is proposed as a solution for both downlink (Fig.
2.1) and uplink channels (Fig. 2.2). From an information-theoretic perspective, NOMA with SIC is an optimal multiple access scheme from the view point of the achievable mul- tiuser capacity region, in the downlink [32] [33] [34] and in the uplink [31]. In studies related to practical cellular networks [13], [30], again NOMA is more spectrally efficient when compared to orthogonal multiple-access techniques.
The benefits of NOMA does come with its challenges. Firstly, there is a need of an acute and reliable interference cancellation at the receiver. The lack of such will lead to error propagation and further deteriorates the performance of the NOMA receiver [35].
Secondly, since the user signals are distinguished in the power domain, the power allo- cation among the users is essential for improved capacity. Even with perfect interference cancellation, if the power is not allocated among the users properly, the expected perfor- mance outcome will not be reached. The efforts put on both the interference cancellation and power allocation are expected to pay back, however, these are usually sophisticated
Chapter 2. Background and Preliminaries 13
BS
UE1
UE2
power user1
f
user2
interference
cancellation UE1 signal
interference
cancellation UE2 signal
FIGURE2.1: NOMA in power domain with one BS and two UEs in down- link channel.
and computationally heavy algorithms. Thirdly and often the challenge most overlooked, NOMA algorithms should run fast enough to meet the ultra-latency requirement of 5G networks.
2.3 Interference Cancellation
In downlink NOMA (see Fig. 2.1), the signals of each user are scaled and added on each other at the BS and then transmitted. The same superimposed signal is received by each UE in the cell. In uplink NOMA (see Fig. 2.2), on the other hand, each UE transmits their signal at the same time and at the same frequency. The signals are scaled in proportion to
BS
UE1
UE2
power user1
f
user2 interference cancellation
UE1 signal UE2 signal
FIGURE2.2: NOMA in power domain with one BS and two UEs in uplink channel.
Chapter 2. Background and Preliminaries 14 their distances to the BS and added on each other when they reach the BS. In both uplink and downlink, to extract the desired signal from the superimposed signal, interference cancellation is applied. In this Section, the two most commonly used interference can- cellation schemes in NOMA are introduced, successive interference cancellation (SIC) and parallel interference cancellation (PIC). Both the scholarly and industry literature on these techniques are discussed.
2.3.1 Successive Interference Cancellation (SIC)
Fig. 2.3 illustrates the working principle of a SIC receiver. The receiver decodes the received signal and the first decoded signal belongs to the one with strongest power. The remaining signals are seen as interference to this signal. The decoded data is then used to regenerate the signal as it was travelled through the channel. This can be done by modulating a carrier wave, multiplying with estimated channel coefficient and adjusting its phase. The regenerated signal is then subtracted from the received signal and the processes are iterated until the desired signal is obtained. The BS needs to iterate the process as many times as the number of users to decode each of them, whereas, UEs will iterate until they find their signal.
decoding UE1 data
with residual interferene received
signal
decoding UE2 data
with residual interferene
decoding UEK data
regenerate UE1 signal
regenerate UE2 signal
FIGURE 2.3: NOMA with SIC receivers forK UEs with channel gains:
UE1 > UE2 > .. > UEK.
Chapter 2. Background and Preliminaries 15 First versions of SIC were mentioned as early as 1994 [36] where SIC was discussed as a candidate scheme for direct sequence CDMA (DS-CDMA) networks with different and equal power distribution. The work implies that the more SIC iterations occur, the better BER (bit error rate) performance is achieved. Nevertheless, there was a tremendous decoding complexity due to the number of iterations. The work demonstrates this by analysing BER for 60 UEs without implementing complexity reduction techniques, such as user pairing, and finally compares the results of SIC with the conventional interference cancellation techniques.
Even though there are exceptions as in [23], where the iteration order starts from the weak signals, most of the recent works mentioned below employed decreasing order. For example, the strongest signal is decoded and cancelled first. The first decoded signal has the most amount of interference, while the last signal only decodes its signal and in case of perfect interference, the cancellation will experience only a noisy channel [37]. The SIC receiver has also some drawbacks that can be listed as
1. The scheme is dependent on the decoding order and substantial power difference is required.
2. The decoding time is proportional to the number of UEs in the network which is still expected to be compensated by the computation power of new mobile CPUs.
3. The channel information is needed at the receiver to prevent residual interference that leads to capacity erosion.
4. Interference that occurs from the neighbouring cell.
5. The errors made in decoding propagates to other iterations.
In the context of 5G, NOMA with SIC was first mentioned in a technical document by NTT Docomo [38]. In their works, engineers envisage the first weakness in the list above and kept their discussion with only two users in a cell. Soon after NTT Docomo’s initiation, their works were extended to applications with multiple-input multiple-output
Chapter 2. Background and Preliminaries 16
decoding received
signal
decoding
decoding
regenerate UE1 signal with residual interference
regenerate UE2 signal with residual interference
regenerate UEK signal with residual interference
Add all except the desired signal
decoding data of the desired UE
FIGURE2.4: PIC receiver model forKUEs.
(MIMO). In [39] the data throughput of ‘NOMA with SIC maintained by advanced power control’ was compared with that of OMA and demonstrated significant performance im- provement with NOMA. BER results for three UEs were more deteriorated compared to user pairing.
2.3.2 Parallel Interference Cancellation (PIC)
In this part, works related to PIC receivers with NOMA and CDMA schemes will be summarised. In the literature, there are fewer works for PIC compared to SIC, even though the scheme has considerable advantages over SIC. Furthermore, the performance of PIC may be enhanced via implementing it on recently available hardware upgrades such as parallel programming. PIC model for K UEs is illustrated in Fig. 2.4. In PIC receiver, all the signals except the desired one are demodulated in parallel, regenerated and then summed. The sum is then subtracted from the received signal along with the assumption that the difference is the desired signal. The final step is decoding the desired signal.
In comparison with SIC, in the literature, there are contradicting research examples that are biased to one of the schemes. For instance, BER results in [40] with up to 16 UEs showed a worse PIC performance than SIC where the comparison was made in both uplink and downlink channels. The study added OFDM computations to CDMA scheme and
Chapter 2. Background and Preliminaries 17 proposed the receiver for multi-carrier CDMA. Results compared minimum mean square error (MMSE) per carrier, maximum ratio and equal gain combining receiver methods for single-user detection with hard and soft-input decoding methods.
Pioneers of PIC for 5G, Anwar et al. [41], revealed the drawbacks of SIC in downlink and proposed PIC as a more feasible alternative. The works discussed in [39] and [29]
criticise SIC for its requirement for a substantial difference in the power level of each received signal, whereas in practice, this may not be maintained easily. Moreover, most works with NOMA with SIC note its dependency on the accurate decoding at the SIC iterations, otherwise it leads to error propagation [42].
The proposed PIC scheme solves the aforementioned problems [41]. The study demon- strated the probability of bit error with SNR and with the number of UEs, then compu- tational complexity is presented. The study analyses the performance and admits the necessity to practical implementation of channel estimation. In their comparison of com- putational complexities for different number of UEs they find linear dependency for PIC scheme againstO(logn)for SIC. Authors expect powerful smartphones in the next decade to cope with the tasks in parallel.
In this thesis, we evaluate the computation performance of both SIC and PIC receivers on different computing platforms inSection 3.3and for multicarrier transmission NOMA- OFDM in Section 4.2. For SIC, the impact of decoding mismatch on the computation performance in terms of execution time is also discussed inSection 3.2.
2.4 Power Allocation in NOMA
In power domain NOMA, since the users are distinguished by their power levels, proper power allocation among the user signals is needed. As mentioned earlier inSection 2.3.1, for the SIC receiver to properly cancel the interference, the power levels of each con- tributing signal should be well differentiated. In downlink, the BS allocates its available power among the UE’s modulated signals before adding them up. It allocates more power
Chapter 2. Background and Preliminaries 18 the UEs that are located far from itself and less power to the ones that are closer to the BS [13,29,43]. This enables the SIC receivers to accurately cancel the interference. For instance, the farthest UE has the largest component in the received signal (the other sig- nals seem as interference) and the in the first decoding iteration, the UE obtains its signal.
Whereas in the uplink, although the adjusting transmission powers of the geographi- cally distributed UEs is possible with additional signalling, usually the difference in the received power levels of the UE signals is achieved naturally by their distances to the BS [44]. This time, assuming that all the UEs have the equal transmit power, the closest user would contribute more to the received signal and is decoded first at the SIC of the BS.
Both in downlink and uplink, optimum power allocation remains as challenging problem due to its computational complexity and user mobility while satisfying many constraints such as latency and fairness in the network [37,45–49].
In PIC, on the other hand, the receiver performs better than SIC when the received power level of each signal is equal [50]. This can be achieved by a power control mech- anism as in CDMA networks [51] [50]. For PIC to work properly for a large number of UEs, a signature code is required to distinguish the user signals and the receiver becomes subject to code domain NOMA [52]. In the uplink, power control is implemented to or- chestrate the UEs transmit powers so that the closer UE will have less transmit power and the further UE will have higher transmit power. In the downlink, the BS allocates more power to the signal of UE that is the furthest [53]. This power control strategy is routine work and had already been implemented in cdma One and 3G networks.
2.5 User Clustering in NOMA
Andrews and Meng had optimistic views [54] on the future radio networks of that time.
However, their assumptions like“only one decoder for all the users”or“Further as data- rate demands increase in the future, the number of users per cell will remain constant or
Chapter 2. Background and Preliminaries 19 possibly decrease.” do not match to realities that multiple decoders may now serve the same cell. Moreover, the number of UEs per cell was much larger than they expected.
User clustering was a novel NOMA technique that deals with a large number of UEs by grouping them to diminish the complexity or to increase the scope of the network (see Fig. 2.5). User clustering was first mentioned for efficient optimal power allocation in [55] and several works followed on from the idea. In the related literature, user clustering is investigated for both uplink and downlink channels for several decoding schemes and power allocation algorithms.
BS
UE1
UE2
UEK
UE1 UE2
UEK
Cluster 1
Cluster 2
FIGURE 2.5: Cluster based NOMA system
Cluster formation, however, is a challenging task since different grouping strategies may give very different network performance. Most works in the literature consider a basic cluster with two users which is more like pairing rather than grouping the users.
One major limitation of large cluster size is that it results in increased complexity and delay in SIC decoder. In [56], the two objective of cluster formation is listed as follows:
1) finding the optimum cluster size and 2) matching the users to clusters to maximise the sum capacity. The cluster formation is coupled with optimum power allocation and proportional fairness subjects as discussed earlier inSection 2.4.
Chapter 2. Background and Preliminaries 20 In NOMA related literature, user clustering found its place in earlier works. For two UEs in downlink NOMA system, in [57] power allocation is studied using cognitive radio perspectives. [58] derives a suboptimal solution for max-min resource allocation and [59]
studies the optimal resource allocation for maximising the sum capacity. Joint optimiza- tion of beamforming parameters and power allocation is studied in [60]. For the uplink in [61], user paring is studied and resource allocation is performed among these pairs.
The aid of user clustering in full-duplex communication is another challenge for NOMA systems which is also investigated in [62]. Joint optimization of beamforming and MIMO parameters while detecting the users to be paired is proposed in [63]. User clustering is still an ongoing research field. The problems of optimal resource allocation and finding the best group are still be investigated. In this thesis, we approach the clustering problem from a computing perspective and propose parallel processing to challenge the computa- tional difficulties in NOMA receivers inSection 4.1.
2.6 Parallel Programming
This section switches from wireless communications to the computer science field. It dis- cusses articles parallel programming, multi-threading on CPU and then multi-threading on GPU. Furthermore, works which offer acceleration of wireless cellular networks with parallel programming concept are presented.
Future radio access is aiming for the simultaneous connection of multiple devices in one domain. The same concept is applied in parallel processing, where CPUs are multi- plied physically and their tasks are assigned on multiple threads. Engineers keep trying to enable running more tasks concurrently and in parallel as possible. Physical resources such as hardware, special instructions and software are required to use the privileges of multi-threading. The hardware for running instructions is CPU which allows being in- structed on a rich variety of programming languages. For example, java programming language allows direct instructions upon initiated virtual threads. Each thread executes a
Chapter 2. Background and Preliminaries 21
FIGURE 2.6: Sequential and parallel programming for summing numbers from 1 to 16.
particular assigned task and there is a possibility to initiatennumber of threads. Multiple physical cores of each CPU allows it to initiate and stop a thread at a time.
The efficiency of parallel algorithms with work and depth terms are described in [16].
The term ‘step’ substitutes the word ‘depth’ in the latest studies which define the longest sequential chain of computations. The term ‘work’ stands for a number of all operations.
Fig. 2.6illustrates sequential and parallel algorithms for summing numbers from 1 to 16.
In sequential summing, there are 15 works which run after another, whereas in parallel programming the works have another structure. Each depth has half as many works as the previous level until there was only one summing work left. The idea is that those summing operations are independent and maybe run on different processors, cores or threads simultaneously.
There are early works that discussed computation requirements for hardware and also considered CPUs in the BS [64], [65]. The study revealed the scarce of hardware possi- bilities for updated algorithms of those times. The authors compared CPUs of the BSs, which were used for GSM e.g. in the 1990s. Intel® Pentium 166, SUN Ultra 170 and
Chapter 2. Background and Preliminaries 22 DEC 3000/80 processing units were compared. Authors analysed each phase of the re- ceiver such as channel coding, interleaving, modulation and detecting errors separately.
The simulations were done with C++ programming language which was modern at the time. Both works expect CPU accelerations for modules. Interestingly, computations in [64] had foreseen clustered workstations or multiprocessor server appearances and be- came widely available in 4-5 years. Those workstations had to run computations similar to theirs. The works also introduced a method to benchmark algorithms and were done on MATLAB software.
About a decade later a study proposed multi-core digital signal processor architecture for cellular networks [66]. A discussion included debugging of parallel codes and Field Programmable Gate Arrays (FPGAs) and even obsolete Application Specific Integrated Circuits (ASICs). The work only proposes conceptual architecture without a practical discussion. Then multi-core physical layer for the super base station was proposed for time-domain long term evolution (TD-LTE) in [67]. Proposed digital signal processor for data control, uplink and downlink data processing were located on multiple cores.
Another work proposed speed enhancement via parallel computations in multiple ac- cess MIMO channels [68]. CPU cores or threads were not involved in the offered al- ternative optimization algorithm. Results presented the average worst-case MSE of the proposed transceiver design. Design of the proposed robust transceiver outperformed non-robust transceiver one. There are several works which modify algorithms and im- plement parallelisation in information theory [69–72]. These show the benefit of parallel concept became recognisable in different layers of wireless communications and availed in the design/implementation of both hardware and software.
The number of cores in conventional CPUs nowadays reaches up to 18, whereas the graphics processing unit has built-in thousands of cores. Those GPU cores are also avail- able for general-purpose programming along with graphics visualisations. Thus, GPU solves problems with a large amount of data or many similar small-sized tasks in a rela- tively short time. Open CL by AMD and CUDA by NVIDIA® are software platforms for
Chapter 2. Background and Preliminaries 23 GPU acceleration. CUDA is being widely used in machine learning, computational fluids dynamics, industries which use weather/climate data and generally applied in research of different directions [73]. Telecommunications are also becoming popular and LTE BS is based on GPU as proposed in [74]. After DSP and FPGA based BSs, the new compu- tation hardware is offered and the results are compared with the benchmark. The study involved two GPUs and concluded that in a real-time scenario more GPUs are required.
These are the benefits of GPUs highlighted from the articles:
1. GPUs have thousands of threads and offer general-purpose programming 2. Data-level parallelism.
3. Threads parallelism.
4. Consume less power and therefore cost-efficient.
Concluding the benefits listed above, the article finds GPU as the best hardware al- ternative for an LTE BS. Authors suggest allocating different threads per connected UE.
Antennas, symbols and subcarriers may also be served by GPU threads in parallel. CPU is needed to operate GPUs. The researchers used two GPUs and reached the data rate of 75 Mbits/sec in a setup of the article. Another recent study applied the CUDA platform to accelerate complex computations in massive MIMO systems [75]. The GPUs are not widely implemented in accelerating problems of wireless communication. Neither the al- gorithms of computation are discussed. NVIDIA® is researching 5G and has a special di- rection of telecommunications in famous annual GPU technology conference [76], [77].It may be concluded that there is a credible premise for GPUs in 5G which comes naturally rather than a crammed mixture of technologies. This work aims to show some practical unbiased appliances of parallel programming on GPU for NOMA. Indeed, NOMA with related procedures is expected to be applied in wireless communications cellular networks of the future generation.
Chapter 2. Background and Preliminaries 24
2.7 Artificial Intelligence in Future Networks
The aim of AI assistance in wireless communications is that upcoming mobile networks can learn several system parameters from users’ behaviour to the changes in physical layer characteristics to autonomously optimise itself for improved performance. This smart ap- proach of self-configuration requires the implementation of sophisticated AI tools. How- ever, ML and more recently proposed DL tools have already found their places in the context of AI-assisted wireless communication, though it is more conceptual than prac- tical today. In Fig. 2.7, the illustration shows a smart base station that learns through observations and then executes actions based on its evaluations, i.e. optimisation of se- lected system parameters.
ML techniques have been considered for wireless systems to estimate massive MIMO channels, user location learning, spectrum sensing, device-to-device user clustering, re- source allocation, spectrum sharing, energy harvesting and HetNet selection (see [78] and the references therein). DL was also incorporated into wireless communications in the context of channel coding for MIMO in [79–81]. In [82], DL was used to develop a code- book adaptively to increase the error performance of the SCMA systems. The authors also demonstrated the superiority of DL-aided methods in terms of computational time.
AI has also been applied to multicarrier systems for channel estimation and signal detec- tion in [83] and its ability to detect signals directly as opposed to conventional receivers which estimate channel first and then detect the signal. Moreover, DL aided data traffic
BS
observations Learning algorithm
Evaluations Action
selection
Intelligent Radio
FIGURE2.7: Illustration of learning platform of AI assisted radio.
Chapter 2. Background and Preliminaries 25 management was considered in [84–87].
The recent popularity of AI and useful assistance from ML and DL results in a plethora of works. It can be seen that advancements in algorithms barely challenge with the non- linearity rules and policy obtained with the AI. InChapter 5, the power allocation problem in the downlink NOMA predicts the power allocation coefficients to maximise the sum- rate by applying AI algorithms is explored. Although many other recent papers have also attempted sum rate and reliability optimisation, these methods require high compu- tational complexity due to the nonlinear optimisation. This paper further compares the computation times of AI-aided techniques with optimal numerical search methods and demonstrate superiority.
26
Chapter 3
NOMA with SIC and PIC
This chapter deals with optimum power allocation mechanism for NOMA networks (Sec- tion 3.1), improved security thread detection for NOMA with SIC (Section 3.2) and a method which decreases the running time of PIC and SIC receivers during decoding the signal (Section 3.3). These solutions are meant to fill the gaps identified earlier in the pre- vious chapters. Specifically, the literature referring to power allocation in NOMA mostly consider applications with user pairing. We first demonstrate optimal power allocation with exhaustive search algorithm that is meant to determine the feasibility of the idea and to build a guideline for our research. Moreover, the algorithm keeps a balanced through- put ratio among UEs via predefined fairness index. Since there is scarce attention paid to the NOMA security in the literature, a thread detection that takes into account the ex- ecution time of the SIC receivers is proposed. Finally, we present our numerical results related to the execution time for SIC and PIC on multi-threaded central and graphical processors.
3.1 NOMA Optimum Power Allocation
3.1.1 Introduction and Related Works
Power allocation is an open research field of significant interest for NOMA with SIC. In the literature, there are many power allocation techniques to increase the overall spectral efficiency. Ideal power allocation is exceptionally challenging since it also determines the
Chapter 3. NOMA with SIC and PIC 27 interference level in the network. The problem gets more complicated for multicarrier transmission. Earlier attempts on power allocation for multi-carrier NOMA systems, for example [29], [88], [30], are far from being optimal. The works in [49], [89] consider only two UEs and reduce the number of subcarriers. Then they can propose optimal resource allocation for a hybrid orthogonal and non-orthogonal access scheme. Other works [90]
incorporated subcarrier allocation and obtain close to optimum performance with convex programming but again for two users only.
In optimal power allocation, the objective is usually to maximise the sum rate of the network. The optimisation becomes a non-convex problem due to the interference term in the objective function. This research [91] proposes an analytical suboptimal solution for power allocation that maximises the sum rate by employing a water-filling algorithm.
In [92] authors defined a closed-form suboptimal solution for joint subcarrier and power allocation for multicarrier NOMA system combining Lagrangian duality and dynamic programming. Most of the works in the literature consider maximising the sum through- put, however, fairness is another important objective to be considered in resource alloca- tion. There is a trade-off between the fairness and the sum data rate in cellular networks.
Optimal resource allocation solutions incorporated different fairness measures to their problem definitions in early works such as [44,45]. In [93], authors derive a closed-form optimal solution for multicarrier NOMA with fairness constraint for two users available in the network. The recent study in [93] proposes proportional fairness performance with and without user pairing scenarios for power allocation in NOMA uplink. The work in- volves various decoding orders for both cases. However, the algorithm complexity of both scenarios isO(N2), which is time-consuming comparing to works previously described.
In [94], power allocation is proposed for two and three multiplexed UEs with Karush- Kuhn-Tucker conditions and the performance is compared with exhaustive search. The results are obtained considering the 3GPP scenario. On the one hand, the algorithm has O(N)complexity and aims to maximise sum capacity rather than balancing the ratio or max-min rate of each UE.
Chapter 3. NOMA with SIC and PIC 28 In addition to the fairness index, max-min methods are also employed to demonstrate fair scheduling. For example, [95] employs power allocation for NOMA by maximis- ing the minimum data throughout in the network using channel state information. Their problem is non-convex, however, the proposed solution is low-complexity and provides a close-to-optimal resolution. Research that considered max-min fairness in [96] performed power allocation for outage balancing problem for downlink NOMA that maximises the minimum outage probability.
Due to the computational complexity and challenges in fairness, recent works con- centrated more on user pairing within the cell. The articles [97–100] discuss the NOMA scheme with SIC receiver for user pairing. The works include interference rejection com- bining algorithm with MIMO system; optimum against random user pairing scenarios;
different number of antennas at the BS and the UEs side without inter-cell interference as well as compare user pairing with a tree search based transmission algorithm.
3.1.2 System Model
In our system model, a downlink transmission was considered in the wireless cellular network with NOMA scheme in the power domain. The network has a BS and K UEs each having SIC receivers. The distance of each UE to the BS is different. UEs are ordered starting from the UE1 as the closest one to the BS and ending with UEK, which is the furthest from the BS (Fig. 3.1). Since the NOMA scheme is in power domain, the signal for UE1 is assigned the least amount of power and thus has the weakest channel conditions. Whereas the channel conditions of the farthest UEK are compensated with the largest amount of power. Therefore, UEKhas the strongest signal. As for the receiver side, all UEs receive the same transmitted signal with messages for all UEs and then each UE runs SIC for decoding. Received signals are decoded starting from the strongest one, then the decoded message is subtracted from the received signal. These decoding procedures and subtractions are iterated until the signal which needs to be decoded reaches its order at the SIC receiver. The nearest UE cancels out the decoded signals of all further UEs and
Chapter 3. NOMA with SIC and PIC 29
BS
UE1
UE2
power
f
UE1 UE2
SIC(UEK, UEK-1,...UE2) DECODE(UE1)
UEK
UEK SIC(UE
K, UE
K-1,...UE
3) DECODE(UE2)
DECODE(UEK)
FIGURE3.1: Downlink NOMA with SIC forKUEs
the farthest just decodes its signal due to the signal strength. The remaining part of the signal is considered as interference.
The BS modulates the message signals of each UE with a single carrier modulation scheme such as quadrature phase-shift keying (QPSK), quadrature amplitude modulation (QAM) etc. Then superimposed signalx(t) (Eq. 3.1) is obtained by summing up of all the modulated individual waveforms. Finally, the BS transmits the signal (see Fig. 3.2 (a)), which can be written as
x(t) =
K k=1
∑
p
αkPTxk(t). (3.1)
As can be seen in (3.1), it is the sum of xk(t)messages of every UE scaled with the power allocation coefficientα for UEkand the total powerPT is the available power at the BS. It turns out that the power of UEksignal isPk=αkPT with coefficient values assigned according to the distance as described earlier.
The signal received by each UE is written as
yk(t) =x(t)gk+wk(t) (3.2) gkis the channel attenuation between the BS and each particular UEk,wk(t)is the additive white Gaussian noise at the UEk with mean zero and power densityN0(W/Hz) andx(t)
Chapter 3. NOMA with SIC and PIC 30
FIGURE3.2: Block diagrams of downlink NOMA with SIC (a) transmitter side (b) receiver side.
is the superimposed signal.
The data throughput for NOMA can be represented as follows [23]:
Rk=Wlog2 1+ Pkg2k N+∑k−1i=1Pig2k
!
(3.3) whereW is the transmission bandwidth andNis the total noise powerN=N0W. For OMA the total bandwidth of the channel and power of the transmission are shared equally for all UEs, giving
Rk=Wklog2
1+Pkg2k Nk
(3.4) whereWk=W/KandNk =N0Wk. The sum capacity is an aggregated result of all UEs is expressed as
RT =
K k=1
∑
Rk. (3.5)
Chapter 3. NOMA with SIC and PIC 31 The fairness index of our system model is defined by [101] as
F= (∑Rk)2
K∑R2k (3.6)
which measures the fairness of capacity of each UEs. The value of F varies between 0 and 1. As higher the index as more adjacent becomes capacity values of neighbouring UEs.
Algorithm 1:Optimum power allocation (OPA) initialization
initialize powerMatrix include all possible PAs set fairnessConstraint toF0
foriin powerMatrix do calculate capacity calculate fairnessIndex
if fairnessIndex≤fairnessConstraint then set capacity(i) to zero
end if end for
set maximumCapacity to zero foriin capacity do
calculate capacity(i)
if capacity(i)≥maximumCapacity then set maximumCapacity to capacity end if
end for
The optimum power allocation aims at maximising the total capacity of connected UEs within the fairness index constraint.
maximize
αk
K
∑
k=1
Wlog2 1+ Pkg2k N+∑k−1i=1Pig2k
!
subject to:
K
∑
k=1
Pk≤PT
Pk≥0, ∀k F=F0
where F0 is the targeted fairness index. Capacity is measured along with the fairness index. Each UEkis assigned a power allocation coefficientαk, which is obtained from the
Chapter 3. NOMA with SIC and PIC 32 exhaustive search described in Algorithm 1.
3.1.3 Numerical Results and Discussion
This section presents the numerical results which compare achieved data rates from the OMA and NOMA with different power allocation coefficients. The results were obtained with the following parameters (Table3.1).
TABLE3.1: Simulation Parameters Parameter Name Parameter Value Number of UEs (K) 5
Bandwidth (W) 50 MHz
Total Power (PT) 1 Watt Distance between UEs 50 Meters Carrier Frequency 1 GHz Noise Density 10−17 W/Hz Propagation Model Okumura-Hata
Channel Gains for 5 UEs [-33.21 -36.23 -37.99 -39.24 -40.20]
Fairness Index 0.9
Data throughput for OMA with equally divided power allocation and bandwidth ob- tained by (3.6) with a fairness index equal to 0.9 is given in Fig. 3.3. The UE1 achieved the highest data rate of 9x107bps, the second one has close to 7x107bps, the data rate val- ues decrease until it reached slightly more than 4x107bps for the UE5. The sum capacity for all UEs in the OMA scheme equalled 3.05x108bps.
Compared to the results obtained with OMA, the NOMA evaluation of the data through- put with optimally allocated power using (3.3) was considered. Numerical results prove that NOMA outperformed OMA in all the selected three fairness constraints that were 0.5, 0.7 and 0.9. Data throughput results of NOMA with the lowest fairness index equal to 0.5 are imbalanced (see Fig.3.4). The overall sum capacity roughly reached 4.37x108 bps which was substantially higher than that of OMA. The first UE achieved 2.5x108 bps and the second UE barely reached 1x108 bps. Obtained optimum power allocation coefficients for all UEs become[α1. . .α5]= [0.07 0.2 0.23 0.24 0.26].
Chapter 3. NOMA with SIC and PIC 33
1 2 3 4 5
User equipment 0
1 2 3 4 5 6 7 8 9
Throughput (bps)
×107
FIGURE 3.3: OMA maximum capacity data throughput rate performance for five UEs with 0.9 fairness index.
Fig. 3.5shows the capacity for NOMA with a fairness index of 0.7. Farthest two UEs obtained less data throughput comparing to data throughput of OMA. However, the first three UE achieved more capacity. The highest data throughput 16 x 107 bps belonged to the first UE, and 12 x 107 bps had a UE 50 meters farther. 5 x 107, 3.8 x 107 and 3 x 107bps were achieved by the last three UEs. Optimum power allocation coefficients for this capacity were[α1. . .α5]= [0.02 0.14 0.23 0.30 0.31]. The sum capacity became 4.17 x 108 bps, which was still higher than that of the OMA and slightly less than the sum capacity of NOMA with a fairness index equal to 0.5.
1 2 3 4 5
User equipment 0
0.5 1 1.5 2 2.5 3
Throughput (bps)
×108
FIGURE 3.4: NOMA data throughput for 5 UEs with 0.5 fairness index.
Chapter 3. NOMA with SIC and PIC 34
1 2 3 4 5
User equipment 0
2 4 6 8 10 12 14 16 18
Throughput (bps)
×107
FIGURE 3.5: NOMA data throughput for 5 UEs with 0.7 fairness index.
1 2 3 4 5
User equipment 0
2 4 6 8 10 12 14
Throughput (bps)
×107
FIGURE 3.6: NOMA data throughput for 5 UEs with 0.9 fairness index.
The toughest fairness constraint of 0.9 significantly decreased data throughput of NOMA for 5 UEs (see Fig. 3.6). The sum capacity for all UEs became 3.95 x 108 bps, comparing to 3.05 x 108bps for OMA. The furthest UE reached 5 x 107, whereas in the OMA worst results are 4.4 x 107 bps. The highest data throughput was achieved by the UEs closest to the BS in both schemes. In NOMA the first UE had only 0.01 allocated power coefficient. Data rates were 12.1 x 107bps in NOMA and 4.4 x 107bps in OMA for the first UE. In NOMA with OPA coefficients, all UEs became as [α1. . .α5] = [0.01
