Air Computing: A Survey on a New Generation Computation Paradigm in 6G Wireless Networks

In the literature, the organization and orchestration of the 3D networking structure is called under different names such as aerial communication, Space-air-ground Integrated Network (SAGIN), and airborne networks [6, 7, 8]. Especially, the aerial term is widely used to point to the utilization of air components in 3D networking. Since air layers and the corresponding air components would be an essential part of the next-generation communication systems rather than an auxiliary, we call this next-generation computation paradigm as air computing.

Even though air computing can be theoretically applicable with current 5G wireless networks, it is not suitable to carry out completely since 5G infrastructure does not offer ultra-low latency (0.1 ms), ultra-high data rate (1 Tbps), and mobility as 1000 km/hr, which are essential for mission critical applications and many future applications [9]. Moreover, seamless coverage, which is crucial for air computing considering the communication between terrestrial systems and air layers, cannot be sufficiently provided by 5G due to its network features [10]. Therefore, 6G communication systems are the genuine candidate for air computing as they ensure those essential features [11]. Moreover, since edge intelligence will be a core part of 6G communication architecture, artificial intelligence (AI) would be used prevalently as the solution for computing processes of applications [12].

As the capabilities of 6G wireless networks allow to satisfy the strict demands of modern applications, a wise deployment of terrestrial servers and air components for the computational needs will change the traditional methods such as edge and cloud computing. Therefore, in this study, we investigate the computing opportunities that would be the result of the intelligent communication between terrestrial servers and air vehicles which we call air computing. These opportunities manifest themselves as an enhancement of QoS and QoE considering both communication requirements and end-user needs [13, 14]. Moreover, as shown in Figure 1, we believe that the coordination between devices and different communication mediums through the air would open new research challenges that will shape the future of the Internet.

We foresee that air computing will be the next generation computation paradigm as a result of the evolution of multi-access edge computing (MEC) [15]. Since the computation paradigm in MEC is typically restricted with the 2D terrestrial networks in which the resources allocated for the application tasks in either an edge server or cloud server, system bottlenecks can limit meeting the diverse requirements of heterogeneous applications. Since air computing is comprised of different air communication technologies, resource allocation alternatives are considerably increased when compared with the 2D settings. Vertical networking structure shown in Figure 1 depicts how different applications would use different resources to ensure their QoS requirements.

In this survey, we focus on the computational requirements of the next generation heterogeneous applications and corresponding solutions formulated as the air computing in which terrestrial servers, LAP, HAP, and LEO layers are coordinated intelligently. To put relevant technologies in perspective, we first investigate the differences between edge computing and air computing in terms of the network architecture, challenges, and use cases. Next, we evaluate the advantages of air computing regarding latency, computation capability, storage, mobility, coverage, and reliability.

Since air computing includes both terrestrial servers and air components, we also examine studies on edge computing, UAVs, and other air components in order to show the benefits of air computing more concretely. Furthermore, we detail the possible scenarios in which air computing would be the only valid alternative.

It is important to note that the technical aspects of the aerial radio access network (ARAN) technologies which can be used in air computing are out of the scope of this survey. We only focus on the computation part of the air computing regarding the aforementioned architectural advantages and possible use cases. Regarding our air computing definition, there are three recent survey papers including [16, 8, 17] that investigated the aerial radio access networks (ARANs). In [16], Cao et al. investigated the mechanisms and protocols for airborne communication networks. They detailed LAP-based communication networks, and HAP-based communication networks regarding their channel models, protocols, and spectrum efficiency. In [8], Baltaci et al. focused on the connectivity requirements and use cases of aerial vehicles considering the challenges of employing wireless communication standards. They introduce the term Future Aerial Communications (FACOM) for aerial connectivity and its use cases. They also examined Radio Frequency (RF) wireless technologies to apply in FACOM. In [17], authors focused on the future network design, system model analysis, and enabling technologies in terms of 6G access infrastructure, transmission propagation, communication latency, and energy consumption. They defined the radio access model as ARAN.

The main contributions of this survey are as follows.

• •

  We introduce air computing which is the next-generation computation paradigm for 6G wireless networks. We define its components including terrestrial, LAP, HAP, and LEO layers and investigate them thoroughly.

• •

  We analyze recent studies that focus on edge computing, UAVs, and other air components in the literature and compare them with air computing in order to show its concrete advantages and possible solutions that cannot be offered by traditional networking paradigms.

• •

  We detail the scenarios for air computing that improve the QoS and QoE for end-users. Especially, we focus on mission-critical applications that may require 1 ms latency which may not be met by traditional computing schemes regarding Metropolitan Area Network (MAN) and Wide Area Network (WAN).

• •

  We investigate the open research problems and challenges for air computing considering the architecture design, request management, utilization of artificial intelligence, a required protocol, energy issues, air regulations, and movement mechanisms. We believe that we point out important spots in the literature so that readers would expand their studies through one of those areas.

The rest of the paper is organized as follows. In Section II, we introduce air computing considering its advantages, and its differences with edge computing. We show the use cases of air computing in different scenarios in Section III. In Section IV, we investigate edge, LAPs, HAPs, and LEOs which are the main components of air computing. We examine corresponding studies and show the reader that unresolved issues in those papers would be solved by air computing. In Section V, we provide the challenges, opportunities, and future research directions. Finally, we conclude our paper in Section VI.

Since several application types such as image rendering, video editing, and simulation require intensive computation, cloud computing is proposed as the possible solution regarding CPU and battery constraints of end-devices [23]. However, with the proliferation of versatile devices and corresponding applications known as the Internet of Things (IoT), cloud computing could not meet the low latency requirements [24]. As a result, several computation architectures including Cloudlet [25], Multi-access Edge Computing (MEC) [21], and Fog Computing [26] emerged considering the replacement of the servers, which are not as powerful as the cloud servers, into the local area network (LAN). In the literature, these architectures are named under the umbrella term edge computing [27] which is also used in this study. The general architecture of edge computing is shown in Figure 2.

The main idea behind the edge computing is that processing the computation-intensive tasks, which cannot be processed in the end-device due to CPU and battery limitation, in the suitable server in the LAN. Moreover, the concept can be enhanced to metropolitan area network (MAN) due to scarce capacity [24]. In that case, the most suitable server in the MAN is selected for the corresponding application. Furthermore, cloud servers can also be used with edge computing for latency insensitive applications in order to serve more users. However, selection of the most suitable edge or cloud server for the long-term optimization would be NP-hard [21]. In general, edge computing is used for many application domains such as agriculture, healthcare, smart home, robotics, data processing, video analytics, and virtual reality [28, 29].

5G systems offers low-latency under 5 ms but it is not sufficient to provide ultra-low latency which is 0.1 ms, high mobility that is 1000 km/hr, and high connection density which is $10^{7}$ devices/sq km [30]. As a result, although edge computing is a crucial networking paradigm in terms of QoS and QoE for diverse application needs, the requirement of recent mission-critical applications that require ultra low-latency may not be met by the existing edge infrastructure that uses 5G wireless networks. Moreover, the capacity of the networks considering 2D terrestrial resources is limited to serve very dense mobile and IoT devices. To solve these issues, air computing provides a third layer which is the air including LAP, HAP, and LEO layer to enhance the 2D computational paradigm into 3D by using the technological advantages of 6G communication systems. As 6G networks provide end-to-end latency under 1 ms, air latency between 10-100 $\mu$s, and reliability of $10^{-9}$, it meets future application requirements as an infrastructure regarding computation paradigms such as air computing [11]. The 3D structure on the other hand is also considered as vertical networking and it provides important solutions that cannot be given by traditional edge computing. Thus, air computing can also be seen as the evolution of edge computing in 6G communication systems.

As shown in Figure 3, one of the most important differences between edge and air computing is the direction of the computational task offloading. In edge computing, the direction is horizontal as computational resources are inside the terrestrial 2D area. Moreover, these directions are always one way that is from the user device to the corresponding server. An offloading of a task or a request comes from the user application to the server and then the server process the task in order to give the suitable service. On the other hand, in air computing, the direction of the computational offloading is both horizontal and vertical. An application would benefit from terrestrial resources and air platforms at the same time. Furthermore, the vertical case of the computational offloading may be in both directions. A typical user in an urban area can use the resources in the air and, conversely, an application in an air vehicle can also benefit from the terrestrial servers. The important differences between edge and air computing are summarized in Table I.

Air computing has many advantages that can be categorized as latency, data rate, computational capability, coverage, and mobility. In this subsection, we explain those advantages in detail.

Natural disasters such as earthquakes, hurricanes, tsunami, and floods wreak havoc on settlements and residential areas. They cause the loss of human lives and the destruction of important resources that people need. Considering communication perspective, there would be two important consequences. First, the communication facilities would be destructed by the natural disaster and as a result people can be deprived of important resources that cause isolation in the disaster site. This is crucial for those people affected by the disaster because they cannot obtain the required aid which must be given to the heavy injury cases and also to humans that expect to be rescued. Without communication resources, the outcome of the disaster would be much worse. Secondly, even in the cases where communication facilities are not seriously damaged, bursty traffic caused by people that want to make an emergency call and to reach their friends, and relatives brings about congestion in the network. Note that this congestion can be also caused by the people outside the disaster site since they also would like to reach the people that are affected by the disaster. Similar to the first case, as a result, people can be isolated in the disaster site so that the outcome of the disaster would be heavier.

Considering both cases, the essential requirement for providing the communication and computation in a disaster site is to enhance the capacity dynamically. This can be carried out by using air components of air computing including UAVs, HAP vehicles, and LEOs. Note that the utilization of those components depends on the disaster type. For example, if the disaster is a hurricane, UAVs and HAP vehicles cannot be used during the disaster. Similarly, if the disaster is an earthquake or a flood, using UAVs would be more effective considering the latency which is a crucial metric for these scenarios.

Even though networking paradigms propose several solutions to mission-critical applications that require ultra-low latency, they may not always prove successful because of the underlying technology and incapability of terrestrial resources. Thanks to 6G wireless networking capabilities and air computing, the requirements of mission critical applications for a desired QoE can be met. In this section, we list some of those applications to show their benefits.

The communication infrastructure in terrestrial areas is built considering fixed resources in which once the facility is deployed, it can be changed with difficulty and significant additional cost. Considering capital expenditures (CAPEX) and operating expenses (OPEX) of companies, investing on fixed resources was plausible over years. For example, if there are limited resources such as base stations for users that increase in years, improving the capacity which means adding new base stations could be the solution for this problem. However, especially after the proliferation of the versatile application types in smartphones, the fixed infrastructure may not be sufficient for the user needs that change dynamically.

One of the most important examples of this situation is the sport and concert activities where thousands of people rally and use their applications. Statically built infrastructures expose a significant amount of traffic and computational offloading which may be an order of magnitude higher than the expected requests. Even though network slicing, network function virtualization (NFV), software-defined networks (SDN), and edge computing are used to provide a solution, the issue is still open. On the other hand, by using air components and directing them to the corresponding places where activities occur based on the current network requirements, we can increase the capacity of the network dynamically as shown in Figure 8. Hence, even though the network resources cannot meet the number of application requests, new resources and routes can be created through air computing. As a result, QoS and QoE would be enhanced.

Since people that perform outdoor activities including sailing, kayaking, climbing, and trekking may not access corresponding resources due to lack of communication infrastructure, they can face isolation in the natural environments. This situation would be crucial considering several issues. In the case of an injury or a health problem such as a heart attack, emergency services must be accessed immediately in order to obtain the aid. Regarding the current networking conditions, the communication from secluded areas including sea, mountain, and forest cannot be provided properly. Thus, even though there would not be such a problem, people carrying out the activities do not feel safe about those possible issues. Apart from the health, daily routines such as social media, communication via chat, and telephone call also cannot be carried out in those conditions. Hence, even though those activities provide important relief for people, their life quality may have deteriorated.

Through air computing, the issues for outdoor activities can be solved as shown in Figure 9. By deploying UAVs, HAP vehicles, and LEOs in suitable places, people in secluded areas can reach important contents and communication infrastructure. We assume that HAP and LEOs would be more useful for such activities as UAVs may face battery reload problems which is similar to the situation in rural areas.

Even though vertical networking adds another dimension to the current network infrastructure in air computing, the essential element in an urban area would be edge computing as it is deployed widely. As a result, determining a profitable edge strategy is still important [38]. In this section, we investigate the most recent edge computing studies considering resource allocation, task offloading, edge caching, and energy efficiency issues. Moreover, we give an evaluation of selected edge studies in Table II and analyze them considering the benefits of the air computing.

Since providing Line of Sight (LoS) links has many advantages in terms of connectivity, service provision, and latency, UAVs have been deployed in many areas especially remote locations. However, this deployment also brings its own issues including mobility management, UAV networking management, and flight formation [67, 68]. To this end, we categorize and evaluate these issues under trajectory planning, task offloading, placement of UAVs, and energy consumption.

Considering the limited battery energy and cell-based coverage of UAVs, HAP and LEO layers provide important advantages regarding long-distance communication, energy consumption, and management opportunities. Moreover, their performance would not be affected by the weather conditions due to their high altitudes as 10 - 30km for HAP components, and 160 - 200km for LEO satellites.

The main goal of the deployment of the HAP components is to provide connectivity such as Internet access, and to perform as a controller node for the UAVs [18, 96]. However, in certain circumstances, such as a congestion in the lower layers, they can be used as an edge computing server or a relay node. Since their coverage is on a regional scale, UAVs that can be considered as dynamic cells can reach the corresponding resources in a particular area via HAP components. Even though this is one of the reasons that the deployment of HAP components is generally in urban and suburban areas, maintenance issues based on energy consumption also cause an important restriction for their deployment in rural areas.

On the other hand, the essential use case of the LEO satellites is to meet the coverage problems of rural areas where infrastructure for the communication is insufficient. They can be deployed for months thanks to their efficient energy consumption, however they are not recoverable after their deployment. Moreover, they cannot be used for low latency applications due to their high propagation delay. Therefore, they can be used as a complementary resource for urban and suburban areas in order to get access to cloud computing solutions in WAN. Besides, exploiting services in continental or beyond regional distances would be more efficient using LEO satellites since terrestrial nodes may cause high latency [97]. Figure 10 depicts reaching regional and inter-regional resources using HAP and LEO layers.

It is important to note that when the LEO layer is exploited, SAGIN is the term that is generally used by studies to indicate the corresponding communication system [7]. In [98], Tang et al. developed an efficient offloading mechanism for SAGIN. They benefited from the communication between the LEO satellite network, LAP vehicles, and terrestrial resources. To minimize the total delay and to handle the high mobility of nodes, they proposed a deep reinforcement learning traffic offloading approach. Chen et al. [99] benefit from satellite constellations to apply Federated Learning considering communication overhead and privacy issues. They use four different modes including remote cloud learning, onboard satellite learning, federated learning with data sharing, and federated learning with no data sharing to evaluate the performance of their system. On the other hand, Guo et al. focus on service coordination between different air layers in SAGIN [6]. Therefore they separate the requirements of the environment using three service coordination scenarios: (1) fine-grained, (2) medium-grained, and (3) coarse-grained. In the fine-grained scenario, the network in the air is used as complementary regarding the need for the terrestrial network and ubiquitous coverage. Considering the delay-sensitive applications, coordination of the data processing and data communication services is provided by using medium-grained coordination. Finally, mobility and the corresponding service migration are ensured using coarse-grained service coordination.

In [100], Zhou et al. investigate dynamic scheduling problem in task offloading considering SAGIN environment. They deploy UAVs as flying gateways in order to perform the offloading decision. Considering the dynamic environment, they formulate the corresponding problem as an MDP and then apply linear programming. Similarly, Cheng et al. focus on computational offloading in SAGIN considering energy and computation constraints [101]. In their design, UAVs provide edge computing while satellites ensure access to cloud computing. To learn the optimal policy in a dynamic environment regarding large action space and mobility, they use an actor-critic DRL algorithm. In [102], Zhang et al. analyze the architecture and corresponding application scenarios of satellite MEC. They propose network function virtualization and cooperative task offloading methods in order to integrate computing resources and improve the efficiency regarding delay and energy consumption.

Resource allocation and controller placement problems are also essential for LEO studies. In [103], Chen et al. examine the dynamic assignment and placement of controllers in SDN-based LEO satellite networks. They consider two challenges including highly dynamic topology, and randomly fluctuating load. To this end, they take propagation and queueing delays into account and then formulate the adaptive controller placement and assignment problem considering management costs. On the other hand, Zhang et al. investigate resource allocation in non-orthogonal multiple access (NOMA) terrestrial-satellite networks in which terrestrial and satellite components use the same spectrum for the communication [104]. Since the original optimization problem is non-convex, they divided the original problem into three subproblems including user association, bandwidth assignment, and power allocation.

Since each layer is crucial for the performance of an air computing environment, we also summarize their essential features in Table III based on the critical issues. Thus, which layer should be used for particular requirements can be more distinct.

Since air computing has a 3D structure with four major layers including terrestrial, LAP, HAP, and LEO, the networking architecture in terms of offloading mechanism and routing is crucial [106]. One of the main concerns in networking is how the requests would be handled considering the mesh connected different nodes in the 3D structure. To investigate this, we evaluate three candidate design approaches including hierarchical, free, and hybrid designs. Moreover, we also make a comparison between a distributed approach and orchestration in air computing regarding task offloading.

In an air computing environment, handling user and IoT requests are crucial to meet the required QoS. Moreover, since there are different layers in the air, the request management would be more complex than other networking paradigms such as edge computing. Even though architecture design is essential for request management, deciding where to offload and when to offload is crucial for the performance. To this end, the benefits of a distributed system and an orchestrator-based system must be investigated thoroughly.

A distributed system in air computing can be described such that the device which offloads the task takes the decision of where to offload. For example, users can select one of the edge servers for offloading, or an edge server relays the offloaded task to one of the UAVs without consulting any intermediate device. One of the most important advantages of a distributed system is its scalability and its low latency. Since there is no intermediate element considering the offloading, the transmission of the request would be faster.

On the other hand, it has a significant drawback considering the current state of the network. An offloading device in the environment cannot be aware of the current condition of the corresponding servers and network elements in terms of the load of the servers, and the number of requests that may affect the communication links. Therefore, if there is no additional communication regarding this information, the decision taken by the offloading device would cause a failed task offloading. Moreover, note that if an additional communication mechanism is deployed in the environment, it must be well-optimized so that the communication links cannot be affected by the transmission regarding congestion.

In contrast to distributed systems, an orchestrator to which the tasks are first sent by the offloading devices can be used in different parts of an air computing environment. In such a system, a user can directly offload its task to the orchestrator, and the decision of where to offload can be taken by the orchestrator itself which is aware of the current condition in the network regarding communication links and corresponding servers.

Even though it has advantages, several points must be investigated and optimized in an orchestrator-based system. First, the deployment of the orchestrator is crucial for the performance of the system since many entities in the network may send their corresponding requests. Therefore, it should be reachable in a suitable time so that it can relay those requests without violating the maximum delay requirement based on application type. Second, the centralized deployment of an orchestrator results in a single point of failure such that the corresponding portion of the network would be heavily affected as users and IoT devices cannot offload their tasks. To this end, a fault management system should be considered with the deployment of an orchestrator. Finally, third, the cost of an orchestrator in terms of new communication links, delay, and congestion must be minimized for the system performance.

Considering the diverse requirements of various applications and IoT devices in an air computing environment, traditional optimization-based solutions and heuristics would provide limited solutions. Therefore, the application of AI-based solutions will be inevitable. Especially, regarding edge and UAV-based systems, there are already many studies which benefit from AI-based solutions including ML, DL, and DRL [107, 108, 109]. As a result, along with the requirements of the air computing paradigm, novel AI solutions should be applied to meet new challenges.

Since there are many resources that produce an enormous amount of data in an air computing environment, processing them and then learning meaningful patterns considering the performance of the system can be feasible using ML techniques. However, in recent years, DL solutions have been preferred rather than traditional ML algorithms since DL is more successful in terms of training, non-linear transformation, efficiency, and required computing power [110].

Even though DL solutions are preferred in recent studies, the data collection phase in an air computing environment would cause serious degradation of system performance. First, the huge amount of data can bring about congestion problems on communication links. Second, for such a heterogeneous environment, providing the privacy of the data would be difficult. Therefore, applying Federated Learning (FL) solutions would be more suitable with air computing [111, 112].

On the other hand, since many decisions must be taken based on dynamic events in the air computing environment, and they may not be labeled due to the nature of the problem, supervised ML and DL based solutions would be inadequate. To this end, recent studies benefit from DRL in which the agent can learn directly from the environment without needing human interaction [113]. The agent in an air computing environment can be the orchestrator, UAV, or edge server since they take actions based on the current state of the system. Thus, we believe that future studies can consider the deployment of DRL solutions along with FL.

As there are many different entities in the air computing environment, the communication between them should be carried out based on predefined rules, which are defined by a protocol. An air computing protocol should be reliable, secure, and fast so that the entities carry out communication easily [114]. Moreover, it should facilitate the management of the network as it must provide data integrity.

The vehicles in air layers use either batteries or fuel. Therefore, their deployment, trajectories, and computing capacities should be well-optimized to ensure energy efficiency. As mentioned in Section IV, energy-related issues are currently evaluated regarding edge and UAV studies. However, considering all layers of air computing, a collaboration between different air vehicles can reduce energy consumption further.

Since each country has different regulation policies considering flying air vehicles, the entities in an air computing environment should comply with those corresponding rules. Moreover, the air computing protocol should also be in compliance with regulations as reliable and fast communication is vital for flying entities in the air. On the other hand, considering the existing cellular network infrastructure, the standardization between existing resources and newly deployed air computing devices must be investigated for 6G and beyond [115, 116].

Even though deployment is an important research issue for 2D terrestrial networks, this problem is more difficult to manage as air vehicles move. Moreover, since their coverage, transmission quality, and power consumption are heavily affected by their vertical movement, optimization of their altitude and trajectory is crucial [117, 118]. Therefore, request management can be handled along with this optimization in order to provide efficient performance.