Fifth Generation (5G) mobile communication technologies are on the way to be adopted all over the world. At the moment, 5G is being deployed in small areas in almost all the continents, with a higher number of available networks in Europe and the USA. In future, 5G is predicted to account for at least 15% of total mobile communications market by 2025 . It is therefore timely to analyze the impact of 5G on key areas of research related to data management and processing, including databases, distributed systems, blockchain and machine learning.
With its increased bandwidth of up to 20 Gbps, low latency of 1 ms, high device density of one million devices per square kilometer, and virtualization technologies , 5G is generating new opportunities in modern computing. New use cases, such as remote healthcare based on virtual and augmented reality (AR), or ultra high-definition (UHD) movie streaming can only be possible in 5G networks . Other applications, such as machine-to-machine (M2M) communication in automotive and smart drones, and high-density Internet of Things (IoT) devices in smart cities can be handled by current technologies, such as 4G, WiFi and Bluetooth, but they can greatly benefit from the improvements of 5G .
In this paper, we provide performance measurements done in a real 5G network showing a maximum download bandwidth of 458 Megabits per second (Mbps) and a minimum round-trip time (RTT) of 10 milliseconds (ms). While these numbers are far from 5G specifications, they represent current 5G networks running in Non-Standalone (NSA) and expose more than better performance, in terms of bandwidth and latency, compared to 4G networks. We plan to use these performance measurements to emulate 5G deployments where data management and processing systems can be evaluated.
Beyond the obvious impact of 5G in well-know areas, we examine the opportunities and challenges in computing areas related to distributed data management and processing. In this paper, we conduct a systematic survey of challenges and opportunities 5G is bringing to key areas in computing. Our focus is on areas related to distributed data management and processing, such as edge computing and IoT (section 3.1), networking (section 3.2), data storage and processing (section 3.3), blockchain (section 3.4), security and privacy (section 3.6
), and artificial intelligence (section 3.5).
5G technology has the potential to bring forth the idea of millions of shared databases which will impact data analytics, federated learning  and security at the edge. Nonetheless, the concept of millions of databases poses challenges in terms of privacy and security.
We highlight security as a major challenge in 5G deployments, due to multiple factors. First, the high density and increased number of IoT devices that can be connected to a 5G network will increase the risk of attacks, such as Distributed Denial-of-Service (DDoS). It is well-known that IoT devices are easier to break and that some of the largest scale attacks were conducted using distributed IoT devices . Second, the full virtualization in 5G networks is posing new challenges in security management. We analyze in this paper what are the risks of network slicing , the key technology in 5G virtualization.
The remainder of this paper is organized as follows. In section 2, we review 5G specifications, evaluate current deployments with performance measurements and examine 5G network simulators. In section 3, we analyze the impact of 5G on major computing areas. In section 4, we discuss how 5G is going to help in the development of new use cases, such as telemedicine, smart cars and smart cities. We conclude the paper in section 5.
2 5G Technologies
In this section, we review the properties of 5G, in comparison with the previous generations of mobile networking technologies and wireless protocols. We provide a summary of current 5G deployments, with performance measurements done with two 5G smartphones.
2.1 An Overview of 5G
5G is the fifth generation of cellular network technologies specified by the 3rd Generation Partnership Project (3GPP). It proceeds 2G, 3G and 4G and their respective associated technologies. In addition to the classical spectrum below 6 GHz used by the majority of wireless communication technologies, 5G will operate in a high-frequency spectrum, from 28 GHz up to 95 GHz [3, 10]. This range is known as the millimeter wave (mmWave) spectrum. Compared to previous generations, 5G is able to use a larger band of frequencies, thus, avoiding congestion. In comparison, 4G operates typically in the range 700-2600 MHz [3, 105]. In addition to higher bandwidths, 5G will comprise densely distributed networks of base stations in small cell infrastructures. This enables processing at the edge, leading to lower latency for cloud-like services. 5G uses the massive multiple-input and multiple-output (MIMO) technology . This technology uses large antenna arrays in both the base station and on the device to create multiple paths for data transmission. With massive MIMO technology, 5G can achieve excellent spectral efficiency  and superior energy efficiency .
Beamforming111The terms beamforming and massive MIMO are sometimes used interchangeably . is a subset of massive MIMO . Beamforming controls the direction of a wave-front by manipulating the magnitude and phase of the signals sent by an individual antenna placed in an array of multiple antennas. In this way, beamforming identifies the most efficient path to deliver the data to a receiver, while reducing the inference for nearby terminals. In addition, 5G uses a full duplex technology which doubles the capacity of wireless links at the physical layer. With full duplex, a device is able to transmit and receive data at the same time, using the same frequency . Based on these new technologies, it is predicted that 5G has the potential to improve services at the edge, support more use cases, accelerate the development of smart cities (section 4), and enhance user experience .
Compared to 3G and 4G, 5G has a lower latency of approximately 1 ms, increased energy efficiency, and a peak throughput of 10-20 Gbps [3, 70]. The increase in bandwidth will not only support better user experience, but also allow for more connected devices, such as drones, vehicles, and AR goggles, among others. While a 4G base station can only support around 100,000 devices, 5G can support up to a million devices per km . A 5G network is designed to be flexible and suited for edge deployment, which further improves the end-to-end latency and overall user experience.
|Bandwidth [Gbps]||10||1||0.4-7||2 Mbps|
|Frequency [GHz]||0.6-6, 28-95||0.7-2.6||5||2.4|
|Connected Devices||1,000,000 / km2||100,000 / km2||200 / gateway||7 / gateway|
|(max) [Mbps]||(RTT) [ms]|
|Chicago, USA||19/5/2019||Verizon||Samsung Galaxy S10 5G||1,385||17|||
|Chicago, USA||30/6/2019||Verizon||Samsung Galaxy S10 5G||1,070||-|||
|New York, USA||30/6/2019||T-Mobile||Samsung Galaxy S10 5G||579||-|||
|New York, USA||1/7/2019||T-Mobile||Samsung Galaxy S10 5G||529||53.5|||
|Bucharest, Romania||19/7/2019||RCS/RDS||Xiaomi Mi Mix3 5G||458/20.6||12||this paper|
5G is designed to be the network for IoT . In current deployments, IoT devices typically connect to a gateway or to the Internet through WiFi, Bluetooth, Long Range (LoRA), Zigbee, among others . These protocols are suitable for short-range communications with low mobility, used in applications such as smart homes and smart offices. However, they may not be suitable for larger deployments, such as smart cities and smart farms, and high-mobility applications such as automotive. Nonetheless, there are some challenges that need to be considered before the successful adoption of 5G. Firstly, is the power usage of a 5G transceiver sufficiently low such that IoT devices with small batteries can incorporate it? Secondly, what is the cost of a 5G subscription per IoT device? Given the large count of IoT devices, a linear pricing scheme is not going to incentivize 5G adoption. To leverage this, hybrid deployment could be used, where multiple IoT devices connect to a 5G gateway using traditional protocols, such as WiFi or LoRA.
In the context of IoT devices and their use cases, we compare the specifications of major wireless communication technologies in Table I. 5G has almost always the best characteristics, except for latency, where 5G and newer generation WiFi are similar. However, median WiFi latency is 0.9 ms and 6.22 ms on an 802.11n router, when 5 GHz and 2.4 GHz frequencies are used for measurements , respectively. The 99th percentile goes up to 7.9 ms and 58.9 ms for the two frequencies , respectively. In practice, current 5G deployments exhibit latencies in the range of 10-50 ms, with jitter than goes up to 125 ms, as shown in the next section.
Similar to Bluetooth and WiFi (i.e. WiFi-Direct), 5G allows devices to communicate with each other directly, with minimal help from the infrastructure . This device-to-device (D2D) communication is a key feature of 5G that has the potential to accelerate the development of edge-centric applications. For example, in automotive applications, vehicles will be able to talk directly to each other, thus, reducing latency and avoiding connection failure to the base stations. Other use cases of D2D 5G communication are federated learning, where end devices could share data among them, and blockchain where devices need to establish peer-to-peer (P2P) connections.
In addition to the improvements at the physical layer, 5G networks are going to be highly virtualized. Among the virtualization technologies used by 5G, we distinguish software-defined networking (SDN), network function virtualization (NFV) and network slicing. SDN is an approach that separates networking data plane (i.e. data forwarding process) from the control plane (i.e. the routing process). This separation leads to easier configuration and management, and higher flexibility and elasticity . Complementary to SDN, NFV  uses commodity hardware systems to run networking services that are traditionally implemented in hardware, such as routers and firewalls. With NFV, network flexibility is greatly improved and the time-to-market is reduced, at the cost of lower efficiency compared to dedicated hardware.
Based on SDN and NFV, 5G networks will employ network slicing to multiplex virtualized end-to-end networks on top of a single physical infrastructure. By separating infrastructure operators from service providers, 5G will better utilize hardware resources while providing a diversified range of services to both businesses and end users . However, all these virtualization technologies pose new challenges in terms of security management, as we shall analyze in this paper.
2.2 Current 5G Deployments
In August 2019, there were some countries and operators offering commercial 5G networks, with limited deployment. According to a map published by Ookla222Ookla developed speedtest.net, a widely-used tool to measure the speed of Internet connections in terms of download and upload bandwidth, and ping latency. , only USA and Uruguay have 5G networks in the Americas. Beside Uruguay, only South Africa and Australia have 5G in the southern hemisphere. In Asia, 5G is available in some countries in the Middle East, such as Saudi Arabia, Qatar, Kuwait and the United Arab Emirates, as well as in South Korea. In Europe, 5G is available in the UK, Spain, Germany, Switzerland, Italy, Romania and Finland.
Table II summarizes existing speed tests on 5G networks around the world. All these tests are conducted using Ookla’s Speedtest Android application. At the bottom of this table, we present our own tests done in Bucharest, Romania where there is a 5G network provided by the local operator RCS/RDS using base stations produced by Ericsson . For these tests, we use a Xiaomi Mi Mix3 5G (Mix3) equipped with a Qualcomm X50 5G modem.
We note that all these measurements include the connection between the base station and the test server. In our case, the number of hops between the device and the test server is in the range 9-13. While this number is relatively high, the backhaul links connecting the base stations to the Internet are of high capacity. These links are usually based on fiber or microwave technology with capacities of up to 20 Gbps . On the other hand, these measurements expose pertinent download and upload throughput as experienced by the end user.
Next, we analyze the results, and present the details in Appendix A.1. With the Mix3 device, we obtain a maximum download bandwidth of 458 Mbps on 5G, compared to a maximum of 86.3 Mbps on 4G. But the upload bandwidth of 5G is similar or even lower compared to 4G. For example, the upload bandwidths corresponding to the same tests as above are 12.9 Mbps and 11.5 Mbps on 5G and 4G, respectively. The latency is measured in terms of round-trip time (RTT), similar to the results reported by the ping Linux tool. RTT is accompanied by the jitter, representing the deviation from the average latency. The RTT is 12 ms with 11 ms jitter on 5G, and 48 ms with 1 ms jitter on 4G. On average, the download bandwidth of 5G across different end servers and measured in different locations is above 300 Mbps, while the upload speed is always less than 30 Mbps.
While the download speed is significantly higher compared to 4G, the upload speed and latency are surprisingly low, compared to the specifications. This is caused by two main factors. First, the speed test consists of sending requests to servers that are not very close to the base station. Hence, both the latency and bandwidth are influenced by both (i) the 5G wireless link to the base station and (ii) the wired or optic path from the base station to the server. Second, current 5G setups are using the Not-Stand-Alone (NSA) mode . Only the Stand-Alone (SA) 5G mode is supposed to achieve an ultra-low latency of 1 ms . In terms of upload throughput, network operators tend to limit it because typical users are affected more by the download speed.
2.3 5G Simulators
With the limited amount of both 5G deployments and 5G-ready terminals, it is mandatory to explore simulation and emulation solutions for 5G in an effort to develop and analyze applications targeting this new technology. Since our focus is on the impact of 5G on data-driven software platforms, the simulator should be able to reproduce networking behavior at a high level, in terms of throughput, latency, jitter and packet loss.
Many 5G simulators, such as MATLAB 5G Toolbox , NYUSIM , NetTest 5G Network Emulators333http://www.polarisnetworks.net/5g-network-emulators.html, ns-3444https://www.nsnam.org/, , focus on the physical layer (i.e. radio access network (RAN)). Such a detailed simulation at the physical level could offer useful insights to network engineers and mobile operators, but it is time-consuming and resource-intensive.
Instead, simpler solutions could be used to emulate a 5G environment. For example, the tc Linux tool555https://linux.die.net/man/8/tc is able to introduce delays with different patterns to emulate higher latency with custom jitter distributions. Moreover, tc can limit the bandwidth of a given interface. One can use tc in a cluster with Gigabit Ethernet or higher bandwidth links to emulate 5G networking conditions. However, it remains to be investigated how to emulate both D2D and device-to-base station communications on top of an Ethernet network.
In conjunction with network virtualization, which is a key feature of 5G, existing solutions for quick prototyping with SDN can be used, such as Mininet  and its fork, Mininet-WiFi . Some researchers explored the idea of using Mininet as a platform to emulate 4G and 5G on top of wired or wireless networks [89, 81, 87]. However, some of the results reported by these projects are far from both the specifications of 5G and our preliminary measurements. For example, the throughput reported in  is below 100 Mbps. Hence, special care needs to be adopted when using these prototyping platforms to conduct performance measurements for data management and processing frameworks.
3 Areas of Impact
5G and its revolutionary features are going to impact multiple computer science domains and create new use cases, as highlighted in Figure 2. In this section, we present our view on the domains impacted by 5G, such as edge computing, database systems, artificial intelligence, security, among others. In the next section, we present some use cases where 5G and related technologies are going to have a significant impact.
Motivating Use Case. Before diving deeper into each area impacted by 5G, we motivate our analysis by presenting a use case that covers all the areas we mentioned. Representing the growing market of healthcare (section 4.1), this use case assumes that patients (end users) take ownership of their medical data, also called electronic health records (EHR). This is a global trend that tries to put the patient in the center of the healthcare system. For example, Apple allows user to download and keep their medical records on the iPhone .
By storing medical records locally, the user’s smartphone becomes a -database (section 3.3), as shown in Figure 3. In addition to medical records downloaded from clinics and hospitals, this database stores data collected by a variety of IoT devices, such as a smart watch and a mobile electrocardiogram. Within a 5G network, these -databases could be interconnected either through the base station or directly, using the D2D feature of 5G.
These data could be used to train medical deep learning models, such as disease progression models, which are then distributed to the devices to analyze new data and to send alerts to doctors. To make the training more efficient, federated learning (section 3.5) is used to distribute the work among multiple devices with the help of a coordinator. The coordinator could be placed at the edge (section 3.1, in a virtualized micro-datacenter or cloudlet .
Nevertheless, the biggest concern in modern healthcare is the security and privacy of the patients’ data. Recent data breaches [28, 98] motivate the research of new security protocols and techniques. With the adoption of 5G, its virtualization feature could help in isolating the healthcare use case from other verticals. However, virtualization does not always ensure privacy and the security management is problematic in the presence of network slicing (section 3.6). Moreover, the physical infrastructure and SDN configurations need to be verified to ensure a secure environment (section 3.2).
Another method to increase the security and privacy of distributed micro-databases is the use of blockchain technology (section 3.4). For example, some startups, such as MediLOT  and Medicalchain , propose patient-centric healthcare based on blockchain. However, the scalability of blockchain  remains an open problem in the context of 5G networks.
3.1 Edge Computing and Internet of Things
Edge computing is an emerging paradigm which proposes to move cloud services closer to the users and to the devices that produce data, at the edge of the network . With the increasing number of devices connected to the Internet , the pressure on the Internet backbone links is growing. Edge computing alleviates this issue by performing some or all computations closer to the devices that produce the data. Depending on the location of these computations, we distinguish between edge and fog computing. In edge computing, the processing is done on the device or one hop away from the device, for example in a mini-datacenter connected to the 5G base station , as shown in Figure 3. On the other hand, in fog computing, introduced by Cisco in 2015 , the computation could be done anywhere between the edge and the cloud, in switches, routers, base stations or other networking devices.
Edge computing has multiple flavors, among which we distinguish (i) the fog, (ii) Multi-access Edge Computing (MEC), and (iii) cloudlets [91, 97]. The fog is an extension of the edge, where the processing can be done on the way to the cloud, in the backbone’s switches and routers. MEC is a set of standards addressing the diversity of protocols, applications, services and providers of edge computing. A cloudlet  is a small datacenter connected to a networking access point. While typically connected to a base station, one hop away from the devices, a cloudlet can also be placed in the fog, as shown in Figure 3. A cloudlet is using virtualization to provide computing and storage services. With 5G being a heavily-virtualized technology, we are expecting an accelerated deployment of cloudlets.
With the adoption of 5G, which enables higher bandwidth and more connected devices compared to 4G, edge computing becomes a necessity because current cloud interconnections are not able to sustain the traffic. Our measurements on Google Cloud Platform (GCP) and Amazon Elastic Compute Cloud (EC2) show that bandwidths between different cloud datacenters (regions) can hardly hit 100 Mbps, while the majority of our measurements are below 10 Mbps. Only closely-located regions, such as those in Western Europe, exhibit bandwidths of up to 92.8 and 126 Mbps, for GCP and EC2, respectively. These bandwidths are far from being able to sustain the demands of 5G edge devices, where a single device could upload with a throughput of up to 1 Gbps .
The high device densities of 5G are impacting a domain that is in close connection with edge computing, namely, the IoT. On the one hand, low-end IoT devices may not need 5G connections since other technologies, such as WiFi, Bluetooth and Zigbee are sufficient. For example, 5G is not needed for smart door locks, smart refrigerators, smart lights, among others. On the other hand, we identify a series of devices that can benefit from 5G features, as listed below.
Surveillance systems. These systems allow users to remotely scan the area inside or around their homes from the comfort of a smartphone. They can see who is snooping around while overseas, and alert the authorities if needed. For such a service, there is a need for good video quality and high frame rate. 5G would have the bandwidth to deliver such lag-free, high-quality UHD video while allowing flexible reconfiguration which is not possible when using wired connections. On the other hand, edge computing helps in pre-processing the image stream in a cloudlet and send only the alerts to the cloud.
Autonomous cars. Undoubtedly, smart cars need fast response times. The theoretical latency improvement from 50 ms in 4G to 1ms in 5G may be the difference between a crash and a clean pass. Moreover, the D2D communications in 5G can have a positive impact on Vehicle-to-Vehicle (V2V) messaging, further reducing the latency compared to going through a base station. Using D2D, automotive communications can avoid the problem of a single point of failure, represented by a faulty base station. As such, we predict that smart cars will adopt 5G as a reliable means to deliver lots of data as quickly as possible to avert disaster.
Drones. In emergencies and dangerous situations, such as search-and-rescue, fire fighting, surveying, delivery services, having a good bandwidth allows a drone to send high quality sound and video back to its pilot, at the edge, while low latency allows superior control over the drone itself. Similar to autonomous cars, smart drones could benefit from D2D communications, especially in remote areas with no access to a base station. Using D2D, a group of closely-located smart drones can form a swarm to work towards a common objective .
Healthcare devices. With its generous bandwidth and low latency, 5G would improve the monitoring of patients with chronic diseases. Vital statistics can be taken as often as the doctor requires, and alerts can go out the moment the device detects something wrong.
Some of the features of 5G address the challenges faced by the IoT domain. First, the increasing number of IoT devices could be handled by the superior device densities supported by 5G. According to a survey by IoT Analytics Research, there were 7 billion IoT devices in 2018 . This number will be triple by 2025 . This increase in the number of IoT devices raises concerns about security and connectivity, among others. According to an online survey of IoT development conducted by the Eclipse Foundation in 2019 with 1717 participants , the top three concerns are security (38%), connectivity (21%), and data collection and analytics (19%). In the next section, we identify some challenges that edge computing and IoT could face in the era of 5G.
3.1.2 Challenges and Opportunities in Edge Computing
The adoption of edge computing is challenging due to a mix of factors. On the one hand, there is high diversity of edge hardware, communication protocols, service providers and processing frameworks. To overcome this, the European Telecommunications Standards Institute created a special group of interest to propose standards for MEC . 5G could address this issue using virtualization, where hardware functions and software protocols can be virtualized on commodity infrastructure, thus, decreasing the prototyping and deployment time.
Edge services could be driven by the 5G end-to-end network virtualization. For example, network slicing would allow different application planes to run in isolation on the same infrastructure, as shown in Figure 4. With network slicing, cloud-like services at the edge, the broadband connectivity plane and smart city applications could all run in isolation. Nevertheless, the security of such a setup is challenging, as we shall see in the next sections.
Another factor that hinders edge adoption is the high cost of installing, protecting and maintaining edge devices in remote areas . 5G is able to partially address these issues with its high bandwidth and low latency features. The former means that advanced, high-definition security monitoring solutions can be deployed together with the edge hardware. The latter helps in detecting and acting on problems faster, from a centralized command facility.
Remote edge devices may face energy constraints due to the lack of connections to the power grid. Operating on alternative sources of energy, such as solar panels or batteries, imposes constraints on computation and communication. While communication is often more energy-expensive compared to computation, it is a challenge to decide when to process the data at the edge and when to offload it to a cloudlet . With its superior energy efficiency, 5G could help in improving the overall efficiency of edge computing.
3.1.3 Challenges and Opportunities in IoT
One of the biggest challenges IoT development has to encounter in a 5G environment is the new waves of security threats. Since the number of connected IoT devices continue to rise, there is a higher risk that systems will be attacked by malware and ransomware to steal sensitive data or to perform DDoS attacks . This problem is more stringent with the IoT devices being used in automation and security systems at home or in vehicles. These systems may be compromised, leading to more serious threats, such as home intrusion or remote vehicle hijack. A piece of common advice to ensure the security of IoT devices is to keep their firmware and security patch up-to-date to avoid any vulnerabilities exploitable by attackers. Also, users need to change their default account and password periodically on the IoT devices to prevent unauthorized access by brute-force attacks. Finally, data transmission and communication between devices or from the devices to the 5G network need to be encrypted to prevent any leak of confidential data.
The second challenge that needs to be considered is how to guarantee data privacy when IoT devices have access to private data such as surveillance videos, daily habits, health data. Although users can review the coarse-grain access controls of these devices to sensitive information , they generally could not supervise how this data is collected and utilized. For example, if users let IoT applications access their data, they will not know when the devices send the data to developers, advertisers, or any other third parties. To ensure the privacy of data, multiple solutions need to be considered. For example, it is important to set dedicated policies, regulations, rules, or laws to ensure that IoT service providers and developers take necessary actions to protect sensitive data of users . Besides, strong security solutions also need to be set up in IoT devices, to prevent any breach or exploitation of sensitive data.
The third challenge is that most IoT devices might not be tested sufficiently during their production and also might not receive enough firmware updates after deployment. This might be due to the fierce competition in IoT industry where manufacturers often focus on quickly producing and selling devices and do not pay enough attention on ensuring security issues. For example, many manufacturers only offer firmware updates on new devices: they stop the update of old-generation devices when they start working on the new generation. This bad practice might leave IoT devices vulnerable to potential attacks due to the outdated firmware. To overcome this problem, manufacturers are encouraged to test their products properly and update firmware regularly. Furthermore, they should use safe programming languages and automatic program testing and verification techniques to avoid any potential bugs during the product development.
3.2 Network Testing and Verification
Ensuring modern networks, including 5G, operate properly as designed is crucially important to telcos, banks, content providers and other businesses. The failure of these networks might lead to severe consequences. For example, according to a report of IHS Inc. in 2016 , network failure caused the loss of billions of US dollars annually in North American businesses. The failure of a network can occur statically due to its misconfiguration or dynamically at runtime . The misconfiguration errors are often introduced by human mistakes, and they can lead to problems such as unreachable servers, or security holes. On the other hand, runtime errors are due to failures in network links and hardware, or bugs in network software.
Huge efforts have been spent by researchers to develop testing and verification techniques to find network errors . However, this task is known to be very complex since modern networks include a large number (thousands) of not only servers, routers, and end-user devices, but also many middleboxes such as firewalls, optimizers, transcoders, proxies, load balancers, and intrusion detection systems . Furthermore, the software controlling these devices is very complicated: it contains millions of lines of code and runs in a highly distributed environment. For 5G, this testing and verification task will be more challenging, due to the network’s growth in complexity and flexibility. In particular, 5G networks can support a massive number (up to millions) of connected devices. It is also equipped with a novel network slicing feature which allows a slice (a virtual network) to be dynamically created, used, and deleted. In the following, we will discuss more about such difficulties and challenges.
3.2.2 Network Testing
Up to date, testing is the main method that has been used to discover errors in modern networks . Network engineers often find bugs by using a wide range of tools, from the rudimentary ping and traceroute, to advanced tools like nmap, tcpdump, netcat, acunetix, ip scanner. They mostly conduct ad-hoc validations of existing networks (3G, 4G, or enterprise networks) via active monitoring to detect potential problems. For example, a network often needs to be validated after a configuration change, such as when new remote sites are installed, routing policies are changed, or firewall rules are updated.
However, the existing network testing solutions might need to evolve to cope well with the scale of 5G. Unlike 3G and 4G networks, which are limited only to the telecom industry, a 5G network will comprise millions of connected devices from many industry verticals  grouped in network slices, as shown in Figure 4. Therefore, various ad-hoc cases of the network need to be considered for validation. Also, monitoring the network will be more challenging since network slices can be flexibly created, used, and deleted, based on user requirements . Hence, it is difficult for network engineers to manually design testing strategies that thoroughly cover all behaviors of the network.
In order to make 5G network testing more effective and efficient, we should automate this task as much as possible. Firstly, we need to design tools that can automatically analyze the network configuration and generate tests to cover all ad hoc cases. Secondly, these test cases need to be run automatically and periodically on candidate networks to discover any possible errors. In fact, although automatic testing is new in the context of network testing, this idea has been well-studied by the software engineering community. Recently, several efforts have been made to automate network testing. For example, Zeng et al.  have proposed a technique to automatically generate test packages for testing the forwarding behavior of simple networks. However, the size and dynamicity of 5G will be much complicated than existing networks. Hence, there will be many challenges yet opportunities to develop automatic testing tools for 5G networks.
3.2.3 Network Verification
Although testing is commonly used to detect network errors, this method has two limitations. Firstly, it is often used to check the behaviors of a production network but not to examine the network’s configuration. Secondly, this approach cannot guarantee that a network is implemented correctly according to its design, since it is impossible to generate test cases that cover all the possible behaviors of the network. In reality, it is often desired that a network’s configuration can be examined before being deployed to prevent any possible errors in the future.
Inspired by recent advances in software verification, researchers have proposed to treat networks like programs so that they can apply similar techniques to formally verify the forwarding behavior of networks [102, 75]. In essence, a network consists of two planes, the data plane and the control plane . The data plane decides how a network packet is handled locally by a router: when the packet arrives at one of the router’s input links, it will determine which output link to forward it to. The control plane determines how a packet is routed among routers along a path from the source host to the destination host. In traditional networks, both these two planes are implemented in routers. However, in modern networks, the control plane can be implemented as a separate service in centralized servers.
Recent works have focused on verifying simple networks, which are configured by static and immutable forwarding rules, e.g., , or small-scale networks with limited number (hundreds or thousands) of devices [8, 21, 40]. In reality, modern networks often contain various middleboxes, whose states are mutable and can be updated in response to received packets. Hence, the behaviors of routers and middleboxes in these networks are affected by not only their configurations but also by the incoming packets. Further, 5G networks will be even more complex since they allow a massive number (millions) of connected devices. These two challenges, namely, complexity and mutable states, will be the key factors that need to be considered when verifying the forwarding behavior of 5G networks.
As previously mentioned, 5G supports network slicing, where a network slice is a software-based, logical network that can span across multiple layers of the network and could be deployed across multiple operators. Furthermore, the isolation of slices can be flexibly configured at different levels to satisfy the customers’ needs . For example, some users may not mind sharing network resource with others, but would require isolation for the computing resources. Therefore, it will be challenging to formally verify if the isolation property of network slices in a deployed network satisfies its design.
3.3 Data Management and Processing
5G could be the P2P network layer in a system comprising millions of inter-connected micro databases. With a projected density of one million devices per km , 5G is able to connect a few million devices in a smart city, where each device stores a (-)database. This either gives rise to (i) a network of inter-connected -databases or (ii) millions of independent -databases owned by individuals.
The realization of the first scenario in a datacenter equipped with traditional server systems is problematic due to networking and power constraints. Firstly, the connectivity in a datacenter’s cluster is done through switches or routers that become wither bottlenecks or sources of network failure. Secondly, a typical server uses more than 50 W of power, while often reaching 100 W . With one million servers, the power requirements of such a datacenter reach 100 MW, more than the fastest supercomputer in Top500 .
On the other hand, low-power systems based on ARM CPU, such as smartphones and IoT devices, typically use less than 10 W while active [59, 57]. Besides, 5G is predicted to be more energy efficient , hence, will further reduce the power usage of the node. Previous research projects connecting low-power nodes in distributed data management and processing systems [6, 2, 59, 56] show that these devices can significantly reduce the energy usage, while trading-off performance in terms of response time and throughput. Indeed, a distributed network of one million low-power 5G devices has a high possibility to exist in the near future.
The second scenario is that of a P2P database where each individual user owns its data and decides how to share it. In addition, the user may chose to share resources, such as storage space or computing units. In this scenario of -databases, sharing requires a fine-grain access control mechanism to ensure security and privacy, especially in the context of strict data protection and privacy rules, such as European Union’s General Data Protection Regulation (GDPR) .
3.3.2 Challenges and Opportunities
With the ownership of the data being passed back to the users and with the implementation of strict data protection frameworks, such as GDPR, big data analytics has to be efficiently supported. First, it is challenging to perform efficient batch or stream processing in the presence of a fine-grain access control mechanism. For example, users may choose to share only part of the data which could affect the final results of the analytics task. Second, the highly dispersed and volatile nature of distributed nature of P2P -databases make fault-tolerance and task scheduling stringent issues in big data processing frameworks. It is well-known that strugglers affect the performance of data analytics [111, 108].
When large amounts of fragmented data are stored across a large number of devices, data have to be processed locally and/or transferred to the cloud for large scale analytics. For local data management and resource sharing over devices, efficient and light data management is required, possibly with some form of distributed shared memory .
The D2D and density of connections in 5G will present a great opportunity for human-in-loop data processing. A complex job may be partitioned into computer-oriented tasks and human-oriented tasks , which is in line with the exploitation of AI for tasks that machines can do best. Decomposition and classification of tasks need to designed for specific application domains, since domain knowledge and availability of experts are key to high accuracy. Fast response from human is needed to improve the overall data quality and decision-making. However, all these must be examined to ensure that humans do not introduce noise into the system and cause further irregularity.
In the last decade, we have witnessed the rapid proliferation of blockchain platforms, both in public, permissionless networks and private, permissioned setups . From the performance point of view, blockchains are known to exhibit low transaction processing throughput, high latency and significant energy usage [20, 55]. In the context of 5G adoption, it is useful to investigate how blockchain systems are going to be impacted.
From the applications point of view, blockchains could help in providing trusted services at the edge while connecting multiple mutually untrusted entities. For example, mobile number portability (MNP) is an application where different telecommunication companies that do not trust each other need to work together to offer this service to their clients . Using MNP, a client can keep her mobile number while switching the telecommunication provider. For this application, the blockchain could keep a unified database with keeping track of mobile numbers, client ids and telecommunication providers.
With the adoption of 5G, the number of devices that can potentially connect to a blockchain will increase significantly; thus, traditional blockchains are expected to exhibit even lower performance. To improve the scalability of blockchain, researchers have looked into reorganizing the structure of the network. There are two key approaches to do this re-organization, namely, (i) sharding where the network is split into smaller partitions [15, 61] and (ii) hierarchical where there is a main (root) network and many secondary networks [80, 79]. These approaches become more relevant in the era of 5G, edge computing and network virtualization.
Both sharding and hierarchical networks could improve the performance of blockchains. Shards or secondary blockchains running at the edge, in close proximity to 5G base stations, are supposed to run much faster compared to global networks. For example, Hyperledger Fabric 0.6 with PBFT exhibits up to 5x higher throughput in a local cluster with Gigabit Ethernet networking compared to Google Cloud Platform distributed across 8 regions .
3.5 Artificial Intelligence and Federated Learning
With higher connection density and bandwidth, billions more devices are expected to be connected to the 5G network, including mobile phones, tablets, wearables, automobiles and drones. The increased number of interconnected devices (e.g., mobile phones, wearable devices, drones, smart homes) and the accompanying sensors would generate a tremendous amount of data on a daily basis. At the same time, there is a surging demand for personalized services on the devices to enhance user experience. For example, companies may want to provide real-time personalized recommendations to users. The unprecedented amount of data residing in the edge devices is the key to build personalized machine/deep learning models for better user experience. This trend following 5G presents new opportunities as well as challenges for machine learning (ML) and deep learning (DL).
Over the past few years, various hardware accelerators like APU (AI processing unit), NPU (neural processing unit) and VPU (vision processing unit), have been added to mobile chip platforms, including Qualcomm, HiSilicon, MediaTek and Samsung chipsets, to support fast inference of ML and DL models in the edge devices. Corresponding software libraries are also developed, e.g., SNPE SDK, Huawei HiAI SDK, NeuroPilot SDK, Android NNAPI and tensorflow-lite. Currently, the training is typically done on the cloud. The model is then converted into a certain format to be deployed in edge devices. Although 5G enables fast data transfer between mobile devices and base stations, the edge-cloud links have limited capacity in terms of the large scale of data from the connected 5G devices. Consequently, some training tasks need to be shifted from the cloud to the edges to save the communication cost from data transferring. At the same time, training in edge devices resolves the data privacy issue as the data is not shared to the cloud.
Edge devices are expected to handle the training process in some specific scenarios. With datasets getting larger at the edge, the training has to be conducted locally or in the fog, rather than on the cloud, since edge-cloud links typically have limited capacity. In this case, each device or sensor is no longer merely a data carrier; it will also handle processing and data requests from servers or other devices. With faster data flow between mobile devices and 5G base stations, carriers can directly transfer the required data to the client via the edge, fog or D2D, without involving cloud servers. Even ML/DL models can be transferred directly between devices without server.
Data from 5G devices has some special properties that are quite different from the assumptions on training data for traditionally centralized training. Therefore, training in edge devices requires substantial adaptations of model design, training and deployment algorithms. The main properties of data sets for learning in the 5G era can be summarized as:
Highly-distributed. The data is distributed among end devices instead of being collected in a centralized datacenter, where the number of end devices would easily surpass the number of training samples per client.
Unstructured. The majority of samples in a local data set are expected to be in unstructured and diversified formats since raw data are collected from various applications or sensors.
The local data set is gathered from a particular client, and thus great variance is expected among different local data sets. Hence, the local data set is not independent, identically distributed (IID) sampled from the population distribution.
Unbalanced. The amount of training samples varies in different clients and this unbalance is also specific to local data sets due to the bias of the sensors and user preferences.
These data properties pose challenges to the traditional ML/DL training that requires centralized structured training data . First, processing the raw data with these properties for various learning tasks requires substantial efforts on data pre-processing and cleaning. Second, providing real-time personalized services while reducing the server-side burden is highly demanded. Third, the large number of client-side data islands and increased communication efficiency of 5G connections will undoubtedly require substantial adaptation of model design, training and deployment. In response to these challenges, the research was focused on techniques for edge device architecture engineering [36, 62, 88], model compression [18, 33, 34], neural architecture search [116, 83, 51].
3.5.1 AI Model Training and Deployment
New opportunities for a wide range of research directions in AI are arising with the adoption of 5G. In terms of model design and deployment, more client-side models are anticipated. This is possible because (i) edge storage and compute resources are more powerful with various system-on-chip (SoC) technologies and (ii) there is a data-privacy practice to keep personal data private. Further, due to its inherent capability of adaptive modeling and long-term planning, reinforcement learning presents potential in building interactive and personalized models, such as interactive recommendation systems[114, 112, 113].
How to construct the correct machine learning model on edge devices remains a challenging problem. Since the computational capability of edge devices are mostly limited by battery and storage space, four key factors should be taken into consideration before the deployment of an ML/DL model: the power consumption, the storage space occupation, the service latency, and the model performance. In real-world applications, those end-user perceptible requirements/constraints should also be considered in the network structure engineering and hyper-parameter configurations during the training procedure.
Existing techniques such as model compression and quantization may potentially benefit the training procedure on edge devices. For instance, model compression techniques reduce the model complexity in various ways with controlled accuracy degradation. The quantization techniques not only lower the power consumption and storage space cost but also take advantage of the emerging edge-based hardware such as neural processing units that support for integer operations, which is usually much faster than half-precision floats.
Further, the edge devices for diverse IoT applications in the 5G era vary a lot in hardware capacity, and thus device-dependent constraints such as energy, power or latency. The difference requires automation in the model building workflow to satisfy the model deployment in diversified environments. Automated machine learning and Neural Architecture Search (NAS) could provide technical clues to tackle the challenge. For example, the meta-information of datasets, local resource profiles, and service-time constraints can be gathered to model the automation procedure and recommend candidate model configurations.
3.5.2 Federated Learning
Increasingly, mobile devices become the primary computing power for many people. The various embedded sensors and also popular applications collect an unprecedented amount of data on a daily basis. 5G technology will undoubtedly accelerate this trend. Privacy issue should be taken into consideration more seriously when applying a large amount of personal data to ML/DL. Further, directives like GDPR pushes for strict personal data processing, which suggests individuals or organizations to handle data in an appropriate manner.
To preserve data privacy, Google proposes an alternative training technique Federated Learning (FL)  that keeps the personal data residing in the mobile devices, and learns a shared model by aggregating locally-trained updates, as illustrated in Figure 6. In federated learning, data are only accessible to the data owner and training processes run on mobile devices. The server can only receive intermediate results such as models or model updates from clients. Consequently, it helps preserve privacy and reduce the communication costs of dataset transfer.
Federated learning is a feasible and ideal solution for the data privacy concern in the 5G era. The 5G network and federated learning will further complement each other. In recent years, most researches on federated learning focus on communication efficiency and privacy-preserving. With the high-bandwidth and low-latency property of the 5G network, 5G powered technologies will largely improve the communication efficiency of federated learning and make up for the communication overhead caused by privacy-preserving protocols. In addition, the more stable connections brought by 5G can mitigate the dropout issues of clients during the federated learning training. Therefore, federated learning provides a privacy-preserving solution for learning in the 5G era, while 5G makes federated learning more practical and robustness. Further, vertical federated learning  is proposed to tackle the data islands problem. Succinctly, vertical federated learning is a collaborative privacy-preserving learning approach to handle scenarios that multiple parties separately hold datasets of different attributes. For example, when a bank and an e-commercial company decide to collaboratively train a model without disclosing sensitive data, the datasets involved are divided into multiple data islands and vertical federated learning can come to rescue. In addition, smart cities are applications of great potential in the 5G era, and different parties or platforms in a smart city will definitely benefit from resolving these data islands issues. With the 5G technology and vertical federated learning, the world with Internet of Everything is to be anticipated.
3.6 Security and Privacy
Current practices in enterprise security rely on collecting and analyzing data both at endpoints and within the network to detect and isolate attacks. 5G brings more endpoints and vastly faster networks. More endpoints mean larger attack surface, raising the probability of the network being attacked to near certainty. Faster networks impact data collection, as it becomes unfeasible to store, and later analyze, highly granular data over long periods of time. Therefore, 5G demands a fundamentally new security analytics platform. We note that existing solutions, for example Splunk666https://www.splunk.com and LogRythm777https://logrhythm.com, are inadequate for 5G scale, since they stitch together general-purposed data analytics platforms. The desired solution should not have been designed to target general data management workloads, but specifically optimized for 5G workloads.
One distinguishing feature of 5G is network slicing, which enables different applications with different requirements to share the same network. A generalization of virtualization, network slicing works across all layer of the application stack. The radio network layer is multiplexed through spectrum sharing. Networking layer are multiplexed at the telco providers via SDN and NFV. Cloud resources, especially ones near the edge, are multiplexed via virtual machines. The fact that one 5G network slice comprises multiple virtualized resources managed by multiple providers makes it difficult to ensure isolation. In particular, it may be possible to achieve virtual machine isolation with the secure design of hardware vitalization, but does this still hold when, for example, slices at NFV layers are compromised. To ensure slice isolation, it seems necessary for the layers to coordinate, i.e., to agree on a cross-layer protocol. Figure 7 illustrates an example where a slice consists of several sub-slices at different systems. At each system the slice needs to be isolated.
5G is a key enabler for machine-to-machine communication. Applications based on device locations, for instance, may see new devices moving in and out of range at high velocity. This type of ad-hoc communication with high churn rate poses a new challenge for device authentication. In particular, devices must establish identities of each other before communicating, for example, by knowing the mapping of devices to their public keys. The scale of 5G requires an in identity system that supports a large number of users and avoids a single point of trust. Existing public key infrastructures (PKIs) are too heavy weight because they are designed for enterprise identities. Large-scale consumer systems such as those used for end-to-end encryption, for examples iMessage and WhatsApp, meet the performance and scalability requirement, but they still rely on a centralized party. To decentralize the existing identity system, we envision a blockchain based solution which maintains a highly available and tamper-evident ledger storing identity information. However, existing blockchains are severely limited in their throughput and latency. Therefore, novel blockchain systems are needed to meet the performance requirement of future 5G applications.
Apart from the security aspects, 5G also presents new challenges and opportunities in terms of privacy, as the improved bandwidth and reduced latency of 5G open up the possibility of transforming mobile devices into private databases that could be queried in real-time. For example, consider an online shopping service that provides recommendations to users based on their shopping histories. With current technologies, performing such recommendations requires the service provider to store users’ shopping histories at the server side, which has implications for privacy. In contrast, with the help of 5G, we may keep each user’s shopping history in her local device, and let the service provider join hands with the users to perform recommendations in a privacy preserving manner, e.g., by offloading to the users the part of the recommendation task that requires accesses to private data. Such a computation paradigm, however, poses a number of challenges from the technical perspective. First, how should we manage each user’s private data on her local device, so that different service providers could access data through a unified and efficient interface? Second, how could we enable users to make educated decisions regarding which service provider should be allowed to access what data item? In addition, given that each user may have a considerable amount of heterogeneous private data stored on her local device, how could we alleviate users’ overhead in setting up access controls for a sizable number of service providers? Third, when a service provider and a user jointly compute the result of a certain task, the service provider may infer sensitive information from the computation result, even if he does not have direct access to a user’s private data. For example, based on the result of recommendation computed from a user’s shopping history, the service provider may infer partial information about the items that the user purchased in the past. How should we prevent such inference attacks without degrading the accuracy of the result jointly computed by the user and the service provider? Addressing these issues could lead to the development of new techniques that advance the state of the art in privacy preserving data analysis.
We conclude this section by highlighting the key challenges that 5G is introducing in areas related to data management and processing, as depicted in Figure 8.
Security and Privacy. Some of the key features of 5G may have a significant impact on security and privacy. First, the support for massive number of connections increases vulnerability area and provides an ideal setup for large DDoS attacks. Second, network slicing and end-to-end virtualization is challenging in terms of security management in the presence of multiple service providers. Third, D2D communication introduces security and privacy challenges in an era when people are more concern about their private data and when strict data protection frameworks, such as GDPR, are enforced. In this context, there is a need for security standards and ways to ensure consensus among entities participating in a 5G network.
Network Infrastructure. With its impressive bandwidth and high device densities, 5G allows more data to be downloaded or uploaded from and to the cloud. But this will exert a high pressure on both the (i) backhaul links from the base stations to rest of network operator’s infrastructure and (ii) backbone of the Internet, including inter-cloud connections. Our networking performance measurements among different cloud regions show that current connections are not ready for the speeds of 5G. For example, inter-region connections can hardly reach 100 Mbps in throughput and 10 ms in delay, in the best case, while 5G specifications require 10 Gpbs and 1 ms throughput and latency, respectively (section 2.1). We assert that there is a need for both (i) better backbone connectivity and (ii) smart edge-fog-cloud data offloading strategy to cope with the demands in services, data movement and data processing.
Service and Business Models. With the explosion of IoT and mobile devices connected to the Internet through 5G, there is a need for new service delivery and business models. 5G is considered the ideal network for connecting IoT devices, but creating a separate subscription for each device may be inconvenient for the end-user. On the other hand, the high mobility specific to mobile devices introduces new challenges in service delivery and accountancy, in the context of virtualization and edge computing.
4 Use Cases and Challenges
While previous sections analyzed the impact of 5G on relevant computer science domains, this section presents 5G use cases with a focus on analyzing challenges and identifying research opportunities.
The healthcare industry is rapidly expanding, mainly due to the advancements in machine learning which are applied to the medical domain 
. In a recent study, Deloitte estimates the healthcare market to grow to 10 trillion US dollars by 2022. With the adoption of 5G, new smart healthcare use cases are taking shape, such as telemedicine, telesurgery, and smart medical devices.
5G will be the foundation of telemedicine in countries where wired infrastructure is not well developed. 5G mobile services will enable a more effective delivery of remote diagnosis and support for paramedics. This allows for a new and seamless way of delivering cost-effective and direct-to-consumer healthcare as it is no longer limited to traditional face-to-face consultations in healthcare facilities. In order to have connected care and telemedicine, 5G is needed to guarantee low latency and high quality video streaming.
Telesurgery can also benefit from the low latency and high bandwidth of 5G. Telesurgery allows surgeons to execute a real-time surgery, even when they are not physically in the same location, using a remote control to carry out the surgery. Although 4G is sufficient for real-time video transmission under ideal conditions, its relatively high latency renders it unusable for telesurgery. It remains to be studied if 5G, with its improved latency and increased bandwidth, is able to meet the requirements of telesurgery.
One of the main reasons patients with chronic diseases visit the hospital is the lack of medical equipment at home to measure and monitor vital body signs. 5G will alleviate the burden of hospital checks by transferring this functionality to the community (e.g to local clinics and homes). Devices that are community-deployable should be equipped with vital signs sensing, biomarker sensing, video analytics, a chatbot and an AI-enabled intervention mechanism (e.g. a model that can predict disease progression [115, 47]). All these features are more feasible in the 5G era.
Massive Internet of Medical Things (IoMT) market is predicted to grow from 8 to 33 million shipments in the period 2016-2021 . IoMT are clinical wearables consisting of low-power medical monitoring devices that allow for tracking a patient’s status. Such an integrative device receives information from various sensors and sends pre-processed data to healthcare providers who may adjust medication doses or change the behavioral therapy.
We assert that the security and privacy challenges in the era of 5G pertain to the field of healthcare and IoMT, as well. With a series of recent security breaches in medical data management systems [28, 98], security and privacy are one of the biggest concerns in the digitalization of healthcare. Moreover, strict personal data processing directives, such as GDPR , require special attention. It remains to be studied if 5G virtualization could address these concerns.
4.2 Smart City
The key features of 5G, such as high speed, massive connections, and virtualization, will enable the development of smart cities. A smart city is a sustainable city that utilizes smart solutions to improve the infrastructure and provide better services to the community . Among smart city solutions, we mention correlated traffic systems, public safety, security and surveillance. A key objective of a smart city implementation is to provide cohesion among the variety of deployed systems.
Below, we enumerate a few smart city applications that may be enhanced by 5G technologies. Firstly, smart homes can be implemented, where many devices must be inter-connected and where fast Internet access may be needed, especially for security monitoring and alerts. Secondly, smart education could be enabled by stable connectivity and high bandwidth. Students will be able to access a massive number of online courses, and even participate remotely in real-time classes. Thirdly, smart safety and surveillance could be enabled by reliable connections and the integration of real-time video observation in various locations. This allows cities to have real-time emergency response and surveillance of traffic conditions, accidents, banks and ATMs, stores, roads, among others. Lastly, smart power, which uses the smart grid technology  consisting of smart meters, sensors, and data management systems. A smart power solution reduces the energy usage and fuel consumption, while identifying power outages in real-time.
Currently, smart cities are not efficiently implemented due to the lack of powerful connectivity . Low-latency, stable connectivity is required anywhere and anytime within a smart city. For example, a smart city needs ultra-reliable communications, with a reliability of 99.9999 % or above , and must be able to support an immense amount of IoT devices. 5G suits the requirements of smart city connectivity, with its low latency of 1 ms, and high device density of up to one million devices per square kilometer.
Another challenge in smart cities is ensuring the energy efficiency of monitoring solutions . This is challenging in the context of maximizing the life of battery-operated sensors, and requires smart deployment of devices, as well as algorithms to compute an optimal communication-to-computation ratio per device. Nonetheless, the energy efficiency of 5G terminals could improve the battery life of remote monitoring devices.
While the benefits of 5G in smart cities are obvious, there are some challenges that need to be addressed. First, with the inter-connection of vital city infrastructure there is a high security risk in case attackers manage to capture critical nodes. Network slicing is a partial solution to this, where different smart city applications are isolated. However, we discussed in subsection 3.6 that 5G virtualization presents some security risks that need to be addressed. Second, the high volume of data from surveillance and monitoring systems will exert high pressure on the networking infrastructure connecting 5G base station with central facilities. A solution to this is the use of edge and fog computing where partial processing with discarding of fruitless data can be done closer to the data source.
The automotive industry will be significantly impacted by 5G, as it opens up the potential for vehicles to be connected to roadside infrastructure, pedestrians, and other vehicles. Currently, autonomous vehicles are not fully supported by the IT infrastructure due to the lack of mobile antennas and sensors, which does not allow for efficient and stable communication .
4G technology is unable to reach the handling, processing and analyzing standards needed by autonomous vehicles [13, 54, 39]. In order for autonomous cars, also knows as smart cars or self-driving cars, to be well implemented, the time to transmit and process sensor data needs to match at least the speed of human reflexes .
Existing 4G infrastructure, including the mobile antennas on buildings, is not sufficient for autonomous cars . There is a need for significant amounts of antennas located few hundred meters apart to enable stable car-to-car communications . 5G, with its D2D technology, could help in alleviating this issue. In addition, D2D helps with sensor fusion such that cars can have a better view of the traffic and road condition beyond their line of sight.
Wireless communication enables vehicles to share road information or traffic conditions with other vehicles and roadside infrastructure. For example, Cellular Vehicle-to-Everything (C-V2X) protocol uses multiple communication methods, such as Vehicle-to-Vehicle (V2V), Vehicle-to-Network (V2N), Vehicle-to-Infrastructure (V2I), and more . 5G’s low latency will allow for V2V and vehicle platooning, where vehicles directly communicate with each other to share warnings and real-time road conditions. The predicted reliability of 5G at 99.9999% will allow for V2N to run smoothly as it can share real-time traffic information with the wireless network infrastructure. V2I enables communication with roadside infrastructure elements, such as road signs, intersections and pedestrian crossings. These smart vehicle technologies can anticipate potential risks or help in planning an optimal route given real-time traffic conditions. Moreover, these technologies are predicted to improve safety and reduce deaths, since 90% of fatal car accidents are due to human errors .
While 5G is seen as the natural choice for wireless communications in autonomous cars, there are some challenges that need to be address. Firstly, critical decisions must be taken by the autonomous car based on its own processing, such that the reaction time is kept below 2 ms . Even if 5G has a theoretical latency of 1 ms, this is the best case latency to the base station. If multiple hops are needed to get the required data, the latency may increase above 2 ms. For example, our measurements on a real 5G network show an RTT of 10 ms to a sever that is few hops away. Secondly, there is the challenge of trust and authenticity in the messages received by a vehicle from other entities. In the context of security issues in 5G environments, discussed in subsection 3.6, there is an imperative need to evaluate their impact on critical systems, such as autonomous vehicles.
4.4 Smart Drones
The flexibility in the deployment of unmanned aerial vehicle (UAV), also known as drones, has enabled a series of use cases such as the spread of Internet in remote areas, public safety communications, disaster recovery, flood area detection, and special deliveries. The use of multiple drones (drone swarm) allows for the spread of Internet to areas that lack reliable connectivity. In this use case, multiple drones fly autonomously in close proximity to build a wireless network with no gaps in signal distribution to the ground .
The deployment of UAV base stations (e.g. drone base stations) , can be accelerated by 5G, especially with the usage of mmWave and massive number of connections. Currently, the limited radio frequency spectrum below 5 GHz is not capable of supporting smart drones and UAVs. With the use of a larger spectrum, between 28 and 95 GHz (section 2.1), 5G enables effective communication between drones and ground users. More specifically, 5G will enable enhanced mobile broadband (eMBB), ultra-reliable low-latency communication (URLLC), and massive machine-type communication (mMTC) .
Another feature of 5G, namely energy efficiency, could have an impact on UAV base stations. These UAVs are battery-operated and need to exhibit satisfactory operating time to enable reliable connectivity . While alternative sources of energy, such as solar panels, can be used to enhance battery life, the energy efficiency of 5G is a complementary feature that can extend operating time.
5G-connected drones can aid in emergency situations. Drones can communicate and share real-time information with those on the ground. This will increase the effectiveness of search and rescue missions, and will allow relief teams to dispatch rescue teams. For example, this use case can quickly estimate debris level and distribute resources efficiently.
The adoption of smart drones and 5G will expose challenges related to security, privacy and public safety . In addition to DDoS cyber-attacks, as those launched using IoT devices (section 3.1), drones could be used to conduct physical attacks on the population of a smart city. In the context of UAV base stations, there needs to be a clear separation between the Internet providing service and the UAV control plane. Network slicing and virtualization could help in addressing some of the security challenges of smart drones.
4.5 Remote Desktop Computing and Virtual Reality
In contrast to the trend of decentralization, as promoted by edge computing and the IoT, we envision that 5G will enable better experiences in working remotely, from devices connected to a centralized facility. The high bandwidth of 5G allows UHD video streaming, and the low latency of around 1 ms allows playing video games where the 3D engine runs on remote servers, for example on cloudlets. The 1 ms latency is much lower compared to the refresh rate of ordinary displays, which is 60 Hz. In such cases, the terminals connected to 5G do not need to be fat clients running an entire operating system: they can be low-power devices equipped with a browser.
Huawei has started its cloud desktop service for both enterprise clients  and individual clients . In a 5G environment, this cloud desktop service will support image quality of up to 4K. In this way, a mobile device connected to a 5G network can serve as a portable workstation or as a mobile game console. Nvidia introduced RTX server for cloud-based GPU computing . At the moment, these services do not deliver excellent user experience, especially for mobile users, because of high network latency. However, we expect an improvement with the adoption of 5G.
Cloud-based Virtual Reality and Augmented Reality (VR/AR) services can be classified as another kind of remote desktop technologies. The potential use cases of these interactive and immersive experiences are wide and varied, but the platform behind these revolutionary technologies is the same: a combination of cloud and 5G connectivity. Currently, the challenges faced by VR, AR and mixed reality are mainly related to the lacking of mobility and bad user experience in terms of lag and low video quality. Under the umbrella of 5G, the distributed cloud will be the main technology to tackle those issues. With the high bandwidth and low latency of 5G, cloud and edge VR/AR can deliver the high resolution content to the VR glasses. 5G will enable computation offloading from the VR glasses to the edge cloudlets or directly to the cloud. In this way, content delivery will be faster, enabling smooth VR/AR experiences that will be more engaged in terms of interactivity and immersiveness.
Beyond entertainment, 5G will connect the front-end and back-end workers in big organizations. Front-end workers are always the first to interact with a potential or existing customer, or make a product demo for the company. It is often critical for big organizations to connect the customer, the front-end worker, and the leader, across geographic boundaries. With 5G, communication tools will support real-time feedback which will allow distributed workers to overcome the communication delay and respond to customer needs timely. This is very useful, especially in a fast-paced environment.
In this section, we have presented key use cases that are going to be enabled by 5G, as depicted in Figure 9. We summarize by highlighting the trends and challenges we foresee in the event of 5G adoption.
Efficient Healthcare. Healthcare sector has a huge market size which is going to increase with population’s aging. 5G, together with Machine Learning and IoMT, is going to enable more efficient and affordable healthcare, even in under-developed countries.
Smart City. Smart cities, including related domains such as smart cars, smart drones and smart grid, are going to benefit from 5G, as it reduces latency, enables massive IoT, and offers highly-reliable connectivity.
Virtual Reality. Virtual and augmented reality is a sector with huge potential that spans both entertainment and work-related activities. With its increased bandwidth and low latency, 5G could create an immersive experience for VR/AR users, as well an increase the productivity of businesses that use remote desktop computing.
With 5G on the verge of being adopted as the next mobile network, it is necessary to analyze its impact on the landscape of computing and data management. In this paper, we have analyzed the broad impact of 5G on both traditional and emerging technologies and shared our view on future research challenges and opportunities. We hope the review serves as a basis for further study and development of relevant technologies. 5G will make the world even more densely and closely connected, and will present us the vast amounts of possibilities and opportunities to contribute to the challenges ahead of us.
Acknowledgement: We thank Dan Banica for helping us with 5G measurements in Bucharest. This research is supported by Singapore Ministry of Education Academic Research Fund Tier 3 under MOE’s official grant number MOE2017-T3-1-007.
-  5G Automotive Association, The Case for Cellular V2X for Safety and Cooperative Driving, http://archive.fo/cFOfW, 2016.
-  F. Ahmad, S. T. Chakradhar, A. Raghunathan, T. N. Vijaykumar, Tarazu: Optimizing MapReduce on Heterogeneous Clusters, Proceedings of 17th International Conference on Architectural Support for Programming Languages and Operating Systems, pages 61–74, 2012.
-  N. Al-Falahy, O. Y. K. Alani, Technologies for 5G Networks: Challenges and Opportunities, IT Professional, 19(1):12–20, 2017.
-  V. Albino, U. Berardi, R. M. Dangelico, Smart Cities: Definitions, Dimensions, Performance, and Initiatives, Journal of Urban Technology, 22(1):3–21, 2015.
-  M. Alzenad, A. El-Keyi, F. Lagum, H. Yanikomeroglu, 3-D Placement of an Unmanned Aerial Vehicle Base Station (UAV-BS) for Energy-Efficient Maximal Coverage, IEEE Wireless Communications Letters, 6(4):434–437, 2017.
-  D. G. Andersen, J. Franklin, M. Kaminsky, A. Phanishayee, L. Tan, V. Vasudevan, FAWN: A Fast Array of Wimpy Nodes, Proceedings of 22nd ACM SIGOPS Symposium on Operating Systems Principles, pages 1–14, 2009.
-  Apple, Empower Your Patients with Health Records on iPhone, http://archive.fo/ukyrm, 2019.
-  R. Beckett, A. Gupta, R. Mahajan, D. Walker, A General Approach to Network Configuration Verification, Proceedings of Conference of the ACM Special Interest Group on Data Communication, pages 155–168, 2017.
-  S. K. Biswash, A. Ziviani, R. Jain, J. Lin, J. J. P. C. Rodrigues, Editorial: Device-to-Device Communication in 5G Networks, Mobile Networks and Applications, 22(6):995–997, 2017.
-  F. Boccardi, R. W. H. Jr., A. E. Lozano, T. L. Marzetta, P. Popovski, Five Disruptive Technology Directions for 5G, IEEE Communications Magazine, 52(2):74–80, 2014.
-  Q. Cai, W. Guo, H. Zhang, D. Agrawal, G. Chen, B. C. Ooi, K. Tan, Y. M. Teo, S. Wang, Efficient Distributed Memory Management with RDMA and Caching, The Proceedings of the VLDB Endowment, 11(11):1604–1617, 2018.
-  C. Chiasserini, A. Magnan, 5G for the Automotive Domain, https://bit.ly/2KNv0Ss, 2018.
-  J. Choi, V. Va, N. González Prelcic, R. C. Daniels, C. R. Bhat, R. W. H. Jr., Millimeter-wave Vehicular Communication to Support Massive Automotive Sensing, IEEE Communications Magazine, 54(12):160–167, 2016.
-  Cisco, Fog Computing and the Internet of Things: Extend the Cloud to Where the Things Are, http://archive.fo/joKFb, 2015.
-  H. Dang, T. T. A. Dinh, D. Loghin, E. Chang, Q. Lin, B. C. Ooi, Towards Scaling Blockchain Systems via Sharding, Proceedings of International Conference on Management of Data, pages 123–140, 2019.
-  J. Dean, S. Ghemawat, MapReduce: Simplified Data Processing on Large Clusters, Proceedings of 6th Symposium on Operating Systems Design and Implementation, pages 10–10, 2004.
-  Deloitt, 2019 US and Global Health Care Industry Outlook, http://archive.fo/qiR69, 2019.
-  E. L. Denton, W. Zaremba, J. Bruna, Y. LeCun, R. Fergus, Exploiting Linear Structure Within Convolutional Networks for Efficient Evaluation, Advances in Neural Information Processing Systems, pages 1269–1277, 2014.
-  T. T. A. Dinh, R. Liu, M. Zhang, G. Chen, B. C. Ooi, J. Wang, Untangling Blockchain: A Data Processing View of Blockchain Systems, IEEE Transactions on Knowledge and Data Engineering, 30(7):1366–1385, 2018.
-  T. T. A. Dinh, J. Wang, G. Chen, R. Liu, B. C. Ooi, K. Tan, BLOCKBENCH: A Framework for Analyzing Private Blockchains, Proceedings of the 2017 ACM International Conference on Management of Data, pages 1085–1100, 2017.
-  M. Dobrescu, K. J. Argyraki, Software Dataplane Verification, Proceedings of the 11th USENIX Symposium on Networked Systems Design and Implementation, pages 101–114, 2014.
-  R. dos Reis Fontes, S. Afzal, S. H. B. Brito, M. A. S. Santos, C. E. Rothenberg, Mininet-WiFi: Emulating Software-defined Wireless Networks, Proceedings of 11th International Conference on Network and Service Management (CNSM), pages 384–389, 2015.
-  R. Du, P. Santi, M. Xiao, A. V. Vasilakos, C. Fischione, The Sensable City: A Survey on the Deployment and Management for Smart City Monitoring, IEEE Communications Surveys Tutorials, 21(2):1533–1560, 2019.
-  J. B. Ekanayake, N. Jenkins, K. Liyanage, J. Wu, A. Yokoyama, Smart Grid: Technology and Applications, John Wiley & Sons, 2012.
-  Ericsson, DIGI — RCS/RDS and Ericsson Go Live with 5G Mobile Service in Romania, http://archive.fo/siuxf, 2019.
-  Ericsson, Non-standalone and Standalone: Two Standards-based Paths to 5G, http://archive.fo/ppIGd, 2019.
-  European Union, Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the Protection of Natural Persons with Regard to the Processing of Personal Data and on the Free Movement of Such Data, and Repealing Directive 95/46, Official Journal of the European Union (OJ) 59, 1-88 (2016), 294, 2016.
-  L. Evenstad, NHS Data Breach Caused Details of 150,000 Patients to be Shared, http://archive.fo/X5cRR, 2018.
-  X. Foukas, G. Patounas, A. Elmokashfi, M. K. Marina, Network Slicing in 5G: Survey and Challenges, IEEE Communications Magazine, 55(5):94–100, 2017.
-  S. Gridelli, Why Network Testing is Important, http://archive.fo/alS0w, 2016, Accessed: 2019-06-27.
-  I. Grigorik, High Performance Browser Networking, Shroff Publishers & Distr, 2014.
-  B. Han, V. Gopalakrishnan, L. Ji, S. Lee, Network Function Virtualization: Challenges and Opportunities for Innovations, IEEE Communications Magazine, 53(2):90–97, 2015.
-  S. Han, H. Mao, W. J. Dally, Deep Compression: Compressing Deep Neural Network with Pruning, Trained Quantization and Huffman Coding, CoRR, abs/1510.00149, 2015.
S. Han, J. Pool, J. Tran, W. J. Dally,
Learning Both Weights and Connections for Efficient Neural Network,Advances in Neural Information Processing Systems, pages 1135–1143, 2015.
-  Huawei, Desktop Cloud - Cloud-based Workspace for Enterprises, Ultimate Experience of Agility and Efficiency, http://archive.fo/EnDFz, 2019.
-  F. N. Iandola, M. W. Moskewicz, K. Ashraf, S. Han, W. J. Dally, K. Keutzer, SqueezeNet: AlexNet-level Accuracy with 50x Fewer Parameters and ¡ 0.5 MB Model Size, CoRR, abs/1602.07360, 2016.
-  D. Jiang, B. C. Ooi, L. Shi, S. Wu, The Performance of MapReduce: An In-depth Study, The Proceedings of the VLDB Endowment, 3(1-2):472–483, 2010.
-  P. Kazemian, G. Varghese, N. McKeown, Header Space Analysis: Static Checking for Networks, Proceedings of 9th USENIX Symposium on Networked Systems Design and Implementation, pages 113–126, 2012.
-  KeySightTechnologies, How 5G Will Influence Autonomous Driving System, http://archive.fo/6hmn9, 2018.
-  A. Khurshid, X. Zou, W. Zhou, M. Caesar, P. B. Godfrey, VeriFlow: Verifying Network-Wide Invariants in Real Time, Proceedings of 10th USENIX Symposium on Networked Systems Design and Implementation, pages 15–27, 2013.
-  C. Kolias, G. Kambourakis, A. Stavrou, J. M. Voas, DDoS in the IoT: Mirai and Other Botnets, Computer, 50(7):80–84, 2017.
-  D. Kreutz, F. M. V. Ramos, P. J. E. Veríssimo, C. E. Rothenberg, S. Azodolmolky, S. Uhlig, Software-Defined Networking: A Comprehensive Survey, Proceedings of the IEEE, 103(1):14–76, 2015.
-  D. Krishnaswamy, K. Chauhan, A. Bhatnagar, S. Jha, S. Srivastava, D. Bhamrah, M. Prasad, The Design of a Mobile Number Portability System on a Permissioned Private Blockchain Platform, Proceedings of IEEE International Conference on Blockchain and Cryptocurrency (ICBC), pages 90–94, 2019.
-  J. F. Kurose, K. W. Ross, Computer Networking: A Top-Down Approach (6th Edition), Pearson, 6th edition, 2017.
-  B. Lantz, B. Heller, N. McKeown, A Network in a Laptop: Rapid Prototyping for Software-defined Networks, Proceedings of 9th ACM SIGCOMM Workshop on Hot Topics in Networks, pages 19:1–19:6, 2010.
-  E. G. Larsson, L. V. der Perre, Massive MIMO for 5G, IEEE 5G Tech Focus, 1(1):1–4, 2017.
-  C. Lee, Z. Luo, K. Y. Ngiam, M. Zhang, K. Zheng, G. Chen, B. C. Ooi, J. W. L. Yip, Big Healthcare Data Analytics: Challenges and Applications,, Handbook of Large-Scale Distributed Computing in Smart Healthcare, pages 11–41. Springer International Publishing, 2017.
-  R. Leventhal, Report: Patient Monitoring Wearables Market to Grow 35 Percent by 2021, https://bit.ly/2ZH4QFl, 2017.
-  F. Li, B. C. Ooi, M. T. Özsu, S. Wu, Distributed Data Management Using MapReduce, ACM Computing Surveys, 46(3):31:1–31:42, 2014.
-  Y. Li, X. Yin, Z. Wang, J. Yao, X. Shi, J. Wu, H. Zhang, Q. Wang, A Survey on Network Verification and Testing With Formal Methods: Approaches and Challenges, IEEE Communications Surveys and Tutorials, 21(1):940–969, 2019.
-  H. Liu, K. Simonyan, Y. Yang, Darts: Differentiable Architecture Search, CoRR, abs/1806.09055, 2018.
-  X. Liu, M. Lu, B. C. Ooi, Y. Shen, S. Wu, M. Zhang, CDAS: a Crowdsourcing Data Analytics System, The Proceedings of the VLDB Endowment, 5(10):1040–1051, 2012.
-  M. W. Live, Cloud and 5G Bring the Reality to AR and VR, https://bit.ly/2ZvCTAd, 2019.
-  R. Llanasas, 5G’s Important Role in Autonomous Car Technology, http://archive.fo/an5CN, 2019.
-  D. Loghin, G. Chen, T. T. A. Dinh, B. C. Ooi, Y. M. Teo, Blockchain Goes Green? An Analysis of Blockchain on Low-Power Nodes, CoRR, abs/1905.06520, 2019.
-  D. Loghin, L. Ramapantulu, O. Barbu, Y. M. Teo, A Time-energy Performance Analysis of MapReduce on Heterogeneous Systems with GPUs, Perform. Eval., 91(C):255–269, 2015.
-  D. Loghin, L. Ramapantulu, Y. M. Teo, On Understanding Time, Energy and Cost Performance of Wimpy Heterogeneous Systems for Edge Computing, Proceedings of IEEE International Conference on Edge Computing (EDGE), pages 1–8, 2017.
-  D. Loghin, L. Ramapantulu, Y. M. Teo, Towards Analyzing the Performance of Hybrid Edge-Cloud Processing, Proceedings of IEEE International Conference on Edge Computing (EDGE), pages 87–94, 2019.
-  D. Loghin, B. M. Tudor, H. Zhang, B. C. Ooi, Y. M. Teo, A Performance Study of Big Data on Small Nodes, The Proceedings of the VLDB Endowment, 8(7):762–773, 2015.
-  K. L. Lueth, State of the IoT 2018: Number of IoT Devices Now at 7B, http://archive.fo/fFAcD, 2019, Accessed: 2019-06-28.
-  L. Luu, V. Narayanan, C. Zheng, K. Baweja, S. Gilbert, P. Saxena, A Secure Sharding Protocol For Open Blockchains, Proceedings of ACM SIGSAC Conference on Computer and Communications Security, pages 17–30, 2016.
N. Ma, X. Zhang, H. Zheng, J. Sun,
Shufflenet v2: Practical Guidelines for Efficient CNN Architecture
Proceedings of European Conference on Computer Vision (ECCV), pages 116–131, 2018.
-  M. Machowinski, Businesses Losing $700 Billion a Year to IT Downtime, Says IHS, http://archive.fo/F4wdH, 2016.
-  MATLAB, 5G Development with MATLAB, http://archive.fo/99aPV, 2019.
-  H. B. McMahan, E. Moore, D. Ramage, B. A. y Arcas, Federated Learning of Deep Networks using Model Averaging, CoRR, abs/1602.05629, 2016.
-  MediaTek, 5G: What are eMBB, URLLC and mMTC?, http://archive.fo/c7Jnc, 2018.
-  Medicalchain, Medicalchain Whitepaper 2.1, http://archive.fo/ndl1J, 2018.
-  MediLOT, MediLOT White Paper, http://archive.fo/AWcrt, 2019.
-  K. Mekki, E. Bajic, F. Chaxel, F. Meyer, A Comparative Study of LPWAN Technologies for Large-scale IoT Deployment, ICT Express, 5(1):1–7, 2019.
-  E. Mohyeldin, Minimum Technical Performance Requirements for IMT-2020 Radio Interface(s), http://archive.fo/WbPkW, 2016.
-  Nvidia, NVIDIA RTX Server, http://archive.fo/Uq0eR, 2019.
-  B. C. Ooi, K. Tan, S. Wang, W. Wang, Q. Cai, G. Chen, J. Gao, Z. Luo, A. K. H. Tung, Y. Wang, Z. Xie, M. Zhang, K. Zheng, SINGA: A Distributed Deep Learning Platform, Proceedings of the 23rd ACM international conference on Multimedia, pages 685–688, 2015.
-  Ookla, Ookla 5G Map, https://bit.ly/2W2Ipwu, 2019.
-  C. Page, Huawei Cloud PC for Android Brings Windows 10 to the Huawei P20, http://archive.fo/xNa3X, 2018.
-  A. Panda, K. J. Argyraki, M. Sagiv, M. Schapira, S. Shenker, New Directions for Network Verification, Proceedings of 1st Summit on Advances in Programming Languages, pages 209–220, 2015.
-  S. Parkvall, E. Dahlman, A. Furuskar, M. Frenne, NR: The New 5G Radio Access Technology, IEEE Communications Standards Magazine, 1(4):24–30, 2017.
-  M. Passoja, 5G NR: Massive MIMO and Beamforming – What Does It Mean and How Can I Measure It in the Field?, http://archive.fo/OPlvR, 2018.
-  N. Patriciello, S. Lagen, B. Bojovic, L. Giupponi, An E2E Simulator for 5G NR Networks, Simulation Modelling Practice and Theory, 96(101933):1–19, 2019.
-  J. Poon, V. Buterin, Plasma: Scalable Autonomous Smart Contracts, https://plasma.io/plasma.pdf, 2017.
-  J. Poon, T. Dryja, The Bitcoin Lightning Network:Scalable Off-Chain Instant Payments, https://lightning.network/lightning-network-paper.pdf, 2016.
-  K. Ramantas, E. Kartsakli, M. Irazabal, A. Antonopoulos, C. V. Verikoukis, Implementation of an SDN-Enabled 5G Experimental Platform for Core and Radio Access Network Support, Interactive Mobile Communication Technologies and Learning, pages 791–796. Springer International Publishing, 2018.
-  S. K. Rao, R. Prasad, Impact of 5G Technologies on Smart City Implementation, Wireless Personal Communications, 100(1):161–176, 2018.
-  E. Real, A. Aggarwal, Y. Huang, Q. V. Le, Regularized Evolution for Image Classifier Architecture Search, CoRR, abs/1802.01548, 2018.
-  Riccardo Bassoli and Fabrizio Granelli, Rapid Deployment of 5G Services Using Drones and other Manned and Unmanned Aerial Vehicles, http://archive.fo/Z9arE, 2010.
-  R. Roman, J. López, M. Mambo, Mobile Edge Computing, Fog et al.: A Survey and Analysis of Security Threats and Challenges, Future Generation Computer Systems, 78:680–698, 2018.
-  M.-A. Russon, Will 5G Be Necessary for Self-driving Cars?, http://archive.fo/nlXsD, 2018.
-  S. B. H. Said, B. Cousin, S. Lahoud, Software Defined Networking (SDN) for Reliable User Connectivity in 5G Networks, Proceedings of IEEE Conference on Network Softwarization (NetSoft), pages 1–5, 2017.
M. Sandler, A. G. Howard, M. Zhu, A. Zhmoginov, L. Chen,
Mobilenetv2: Inverted Residuals and Linear Bottlenecks,
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4510–4520, 2018.
-  I. Santos, P. Vieira, R. Borralho, M. P. Queluz, A. Rodrigues, Emulating a Software Defined LTE Radio Access Network Towards 5G, Proceedings of International Conference on Communications (COMM), pages 371–376, 2018.
-  M. Satyanarayanan, P. Bahl, R. Cáceres, N. Davies, The Case for VM-Based Cloudlets in Mobile Computing, IEEE Pervasive Computing, 8(4):14–23, 2009.
-  W. Shi, J. Cao, Q. Zhang, Y. Li, L. Xu, Edge Computing: Vision and Challenges, IEEE Internet of Things Journal, 3(5):637–646, 2016.
-  O. Shrit, S. Martin, K. A. Agha, G. Pujolle, A New Approach to Realize Drone Swarm Using Ad-hoc Network, Proceedings of the IEEE 16th Annual Mediterranean Ad Hoc Networking Workshop (Med-Hoc-Net), pages 1–5, 2017.
-  I. Spectrum, 5G Bytes: Full Duplex Explained, https://spectrum.ieee.org/video/telecom/wireless/5g-bytes-full-duplex-explained, 2017.
-  M. Spoonauer, First T-Mobile 5G Speed Tests: The Results Are In, http://archive.fo/IA3eH, 2019.
-  S. Sun, G. R. MacCartney, T. S. Rappaport, A Novel Millimeter-wave Channel Simulator and Applications for 5G Wireless Communications, Proceedings of IEEE International Conference on Communications (ICC), pages 1–7, 2017.
-  M. Swider, 5g Speed Test: 1.4 Gbps in Chicago, but Only If You Do the ’5G Shuffle’, http://archive.fo/jwk7z, 2019.
-  T. Taleb, K. Samdanis, B. Mada, H. Flinck, S. Dutta, D. Sabella, On Multi-Access Edge Computing: A Survey of the Emerging 5G Network Edge Cloud Architecture and Orchestration, IEEE Communications Surveys Tutorials, 19(3):1657–1681, 2017.
-  I. Tham, Personal Info of 1.5m SingHealth Patients, Including PM Lee, Stolen in Singapore’s Worst Cyber Attack, http://archive.fo/8F4fw, 2018.
-  The Eclipse Foundation, IoT Developer Survey 2019 Results, http://archive.fo/teblE, 2018.
-  A. Tobin, 5G Will Account For 15% Of Global Mobile Market By 2025, http://archive.fo/vSfix, 2019.
-  TOP500, Top 500 of the Most Powerful Commercially Available Computer Systems in June 2019, http://archive.fo/uplr7, 2019.
-  G. Varghese, Technical Perspective: Treating Networks Like Programs, Commun. ACM, 58(11):112–112, 2015.
-  E. Vattapparamban, I. Guvenc, A. I. Yurekli, K. Akkaya, S. Uluagac, Drones for Smart Cities: Issues in Cybersecurity, Privacy, and Public Safety, Proceedings of International Wireless Communications and Mobile Computing Conference (IWCMC), pages 216–221, 2016.
-  C. Welch, T-Mobile’s 5G Network Can’t yet Top Verizon’s Speed, but It Has Promising Coverage, http://archive.fo/pZJQN, 2019.
-  Wikipedia, LTE Frequency Bands, http://archive.fo/vUwmx, 2019.
-  Q. Yang, Y. Liu, T. Chen, Y. Tong, Federated Machine Learning: Concept and Applications, ACM Transactions on Intelligent Systems and Technology, 10(2):12:1–12:19, 2019.
-  J. Yu, 5G Mobile Backhaul Isn’t All Fiber, http://archive.fo/1J80n, 2018.
-  M. Zaharia, A. Konwinski, A. D. Joseph, R. H. Katz, I. Stoica, Improving MapReduce Performance in Heterogeneous Environments, Proceedings of the 8th USENIX Conference on Operating Systems Design and Implementation, pages 29–42, 2008.
-  H. Zeng, P. Kazemian, G. Varghese, N. McKeown, Automatic Test Packet Generation, Proceedings of Conference on Emerging Networking Experiments and Technologies (CoNEXT), pages 241–252, 2012.
-  H. Zeng, P. Kazemian, G. Varghese, N. McKeown, Formal Network Testing, http://archive.fo/0Sdvn, 2019, Accessed: 2019-08-21.
-  H. Zhang, G. Chen, B. C. Ooi, K. Tan, M. Zhang, In-Memory Big Data Management and Processing: A Survey, IEEE Transations on Knowledge and Data Engineering, 27(7):1920–1948, 2015.
-  X. Zhao, L. Xia, L. Zhang, Z. Ding, D. Yin, J. Tang, Deep Reinforcement Learning for Page-wise Recommendations, Proceedings of 12th ACM Conference on Recommender Systems, pages 95–103. ACM, 2018.
-  X. Zhao, L. Zhang, Z. Ding, L. Xia, J. Tang, D. Yin, Recommendations with Negative Feedback via Pairwise Deep Reinforcement Learning, Proceedings of 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1040–1048, 2018.
-  G. Zheng, F. Zhang, Z. Zheng, Y. Xiang, N. J. Yuan, X. Xie, Z. Li, DRN: A Deep Reinforcement Learning Framework for News Recommendation, Proceedings of Conference on World Wide Web, pages 167–176. International World Wide Web Conferences Steering Committee, 2018.
-  K. Zheng, W. Wang, J. Gao, K. Y. Ngiam, B. C. Ooi, J. W. L. Yip, Capturing Feature-Level Irregularity in Disease Progression Modeling, Proceedings of ACM Conference on Information and Knowledge Management, pages 1579–1588, 2017.
-  B. Zoph, V. Vasudevan, J. Shlens, Q. V. Le, Learning Transferable Architectures for Scalable Image Recognition, Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 8697–8710, 2018.
a.1 5G Measurements
In this section, we extend the measurements in subsection 2.2. These measurements were run in an existing 5G deployment provided by RCS/RDS in Bucharest, Romania. The base station used in this setup is produced by Ericsson , while the smartphone is a Xiaomi Mi Mix3 5G (Mix3) with a Qualcomm X50 5G modem. We used Ookla’s Speedtest Android app to run our measurements, as shown in Figure 10. We selected three test servers hosted by major telecom companies in Romania. The results, summarized in Table III, compare download and upload throughput, latency and jitter of 5G and 4G.
First, the results show a significant difference between the download throughput of 5G and 4G. The difference in peak download throughput is almost 5, with 458 Mbps for 5G and 86.5 Mbps for 4G. At the other extreme, the difference between the lowest download throughput is 35, with 244 Mbps for 5G and 6.9 Mbps for 4G. We observe a high variability in 4G download throughput, where the highest and lowest values are 86.5 Mbps and 6.9 Mbps, respectively. We attribute this to different network conditions during the measurements, such as connected base station, location, congestion, among others. While we tried to use the same location for all tests, we were not always connected to the same base station, based on the gateway’s IP displayed on our phone.
The second observation is that the upload throughput is similar on 5G and 4G, with values in the range 8.5-23.9 Mbps. 5G upload throughput is well below the 10 Gbps peak provided by the specifications . The third observation is that 5G latency is relatively high, being in the range of 10-41 ms in our measurements. While the measured value is the RTT, it is still far form the 1+1 ms RTT based on the specifications . The measured 5G latency is comparable to the measured 4G latency which is in the range of 13-48 ms.
We attribute these results to at least two factors, as discussed in subsection 2.2. First, current 5G deployments run in the Not-Stand-Alone (NSA) mode , together with 4G networks and, thus, exhibit lower performance compared to the specifications of Stand-Alone 5G. Second, the test servers are more than one hop away from the phone. Two of the servers are at 9 hops and the other at 13 hops away from the phone. Usually, the backhaul links are based on fiber or microwave  and have high capacity, but other network parameters and conditions (e.g. routing, congestion) may affect the performance. The type and configurations of backhaul links could explain the low upload throughput of 5G, because operators tend to prioritize the download throughput which is more perceptible to the end-user.
a.2 Big Data in a Hybrid 5G Edge-Fog-Cloud Setup
In an analysis that combines edge computing and data processing systems, we analyze the effect of 5G in a hybrid edge-fog-cloud environment. For this analysis, we use our existing analytic model for hybrid edge-cloud processing . While the existing model considers only edge and cloud, we introduce the fog by assuming that base stations and attached cloudlets are part of the fog. Only the end devices represent the edge in this analysis.
5G has a direct impact on the bandwidth and latency of the wireless link that connects the end devices and the base stations. We neglect the effect of communication between the base station and the cloudlet, by considering a strong networking connection between these two systems (e.g., it can be a 10 Gbps link with 0.1 ms latency). We employ the well-known MapReduce [16, 37, 49] framework which can easily be generalized to modern workloads, such as sensor analytics in smart cities or federated learning. The Map phase is embarrassingly parallel, since all the processing can be done at the same time on all the nodes involved in the computation. In contrast, the Reduce phase requires intermediate data from the Map phase and usually is done on fewer nodes.
The performance model in  requires three parameters to compute the speedup of edge-only and hybrid processing compared to cloud-only processing. These parameters are (i) cloud hardware speedup over the edge for a given MapReduce application (), (ii) MapReduce selectivity which represents the ratio of input and output size per task or per application (), and (iii) the computation-to-communication (C2C) ratio for a given application, reported to cloud computation time (). 5G only affects this last parameter, where higher bandwidth decreases the communication time.
In our comparison with the fog, we assume that fog cloudlets have the same type of hardware as cloud computing, being based on x86/64 architecture. Hence, we use the same speedups as in 
where mobile devices at the edge are based on ARM CPUs, while the cloudlets and the cloud have x86/64 architecture. Next, we consider three applications suitable for hybrid processing which expose different selectivity and communication-to-computation ratios. K-means (KM) classified a set of multi-dimensional points intogroups, or clusters, and it is representative for Machine Learning applications. KM has high end-to-end selectivity when is relatively small, because it outputs only centroids. However, the intermediate selectivity is low (), because each point is sent to the Reduce phase in order to compute the new centroids. Word Count (WC) represents the classic programming example for MapReduce, being a data-intensive application. WC has low end-to-end and intermediate selectivity. Finally, Grep (GR) allows the search of a regular expression in a huge log, being representative for security analysis, among others. In general, GR has high selectivity both end-to-end and after Map phase, when the regular expression is found in a small number of records.
In a hybrid setup, we distinguish four scenarios:
edge-only processing where both Map and Reduce tasks run at the edge, and the output of the edge processing is uploaded to the fog for storage or visualization
fog-only processing where both Map and Reduce tasks are processed in the fog, and all the input data is transferred to the fog
cloud-only processing where both Map and Reduce phases run on the cloud, and all the input data is uploaded to the cloud to enable the processing
hybrid edge-fog processing where Map tasks run at the edge and the Reduce phase is done in the fog. In this case, all the intermediate data need to be transferred to the fog.
We first analyze the effect of 5G in an edge-fog cluster, where we compare 5G with 4G using their maximum theoretical bandwidths of 10 Gbps and 1 Gbps, respectively. (a) shows the speedup of fog-only versus edge-only and hybrid edge-fog, as defined earlier. We observe that 5G increases the speedup of fog-only because data is transferred faster to the cloudlet where the processing power of the x86/64 architecture is superior to the low-power ARM CPU on the edge devices.
(b) shows the speedup of cloud-only processing compared to hybrid edge-fog. First, we show the processing speedup of the cloud compared to the edge with no data transfer. This is an estimation of the difference in processing power between the x86/64 cloud and the low-power ARM at the edge. Second, we show that 5G induces a negligible improvement in speedup because the main bottleneck is the processing time, rather than communication time. On the other hand, in term of communication, the bottleneck is represented by the fog-cloud and inter-cloud links, as we mentioned before.