A Comprehensive Survey on Orbital Edge Computing: Systems, Applications, and Algorithms

A satellite constellation is a group of artificial satellites working together as a system. Satellites in a constellation have similar parameter such as orbit altitude and orbit inclination. Satellite constellations can provide superior coverage and lowered launch costs. Additionally, they offer unique communication and coverage performance that varies according to the orbital altitude. These distinctions are made clear through Table I. GEO satellite-backed Internet access has been available for decades as they have the advantage of the wide coverage, which can achieve global communication with only three satellites (the small regions around the North and South Poles are normally excluded). However, The high altitude and the requirement to orbit above the equator result in high access latency for consumers, which renders satellite Internet infeasible for many use cases [24, 25].

The development of mobile communication motivates improvement in satellite communication with better performance. LEO satellites attract research attentions due to its better latency and lower launch cost. In 1997, the Iridium constellation was launched to provide voice and message communication service [26]. It consists of 66 active satellites in orbit at a height of approximately 781 kilometres and inclination of $86.4^{\circ}$. Iridium constellation has a remarkable characteristic in which satellites can keep relative stability to each other even in different orbit planes. In providing telephone service, the user has a shorter range to non-GEO satellite network, therefore, less radiated power is required, and the propagation delay is reduced [27]. However, one major drawback of the Iridium constellation is its uneven coverage density, high latitude area can enjoy more dense coverage while it becomes more and more sparse with the latitude lowers. On the other hand, the Walker satellite constellation [28] can realize global uniform coverage, but satellites from different orbit planes cannot maintain a static formation. The Walker constellation is defined by $i:t/p/f$, where $i$ is the inclination, $t$ the total number of satellites and $p$ the number of orbit planes. The phase parameter $f$ is the relative spacing between satellites in adjacent planes. In particular, the Walker constellation is divided into two types: star and delta. The former comprises near-polar orbits at an inclination angle of about 90 degrees. Conversely, the latter includes orbits at an inclination angle generally less than 60 degrees, with the ascending node of the plane distributed over the full range of 360 degrees, and the satellites spaced evenly on the orbit. Emerging massive constellations such as Starlink and Kuiper follow the design of the Walker delta constellation. An example of $75^{\circ}:40/5/1$ Walker $\delta$ constellation is shown as Fig. 3.

Frequency is a crucial factor for satellite communication, microwave (300MHz-300GHz) is suited to satellites because of its wide bandwidth and high frequencies as well as the ability to use small antennas. Table II shows the frequency bands used by satellites. Most communication satellites work in the $C$ frequency band, and with an up link frequency of 5.925 to 6.425GHz and a downlink frequency of 3.7 to 4.2GHz. At present, the $X$, $Ku$ and $Ka$ frequency band has been developed for civil service and broadcasting. However, scarce radio frequency bands cannot keep pace of the rapid increase of satellite applications. Higher frequency bands, such as the $V$ frequency, even up to the optical frequency range (190THz to 560THz), are being explored. For example, Starlink [29] and Micius Satellite [30] has equipped with laser communication payload.

In recent years, thousands of small satellites have been deployed in mega constellations at altitudes below 600 km, i.e., in the LEO. Thus, LEO constellations can also achieve global and continuous coverage thanks to its large number [31, 32]. However, low altitude orbits incur short-time connections between satellites and ground stations (tens of minutes). Inter-satellite communication performs better, especially long-distance, because light propagates faster in a vacuum than in fiber [33]. Inter-satellite links (ISLs) include intra-orbit and inter-orbit links [34]. Intra-orbit links connect satellites in the same orbit plane and form a stable ring network. Satellites in one ring network keep a relatively stationary position with each other. Instead, inter-orbit links connect satellites across orbit planes. In the Walker constellation, inter-orbit links are time-variant. The link between satellites can only keep for a short time.

A satellite network is a dynamic network composed of nodes and links. In this network, nodes are satellites, and links are the communication links between nodes. Due to the time-variant positions of satellite, the network topology is changing continuously. In the Walker constellation, links between satellites can only survive about ten or twenty minutes which depends on the satellite’s orbit. In this subsection, we summarize the algorithm for modeling the topology of a satellite network by refering to [35].

Assuming the links in the network will be broken only if the two satellites are out of sight. The node can be modeled as $S=\{N,P\}$, where $N$ is the identifier of the satellite and $P$ is the position in space. The describe of position could be orbit or coordinates. The model of link $L=\{N_{S1},N_{S2},t_{s},t_{e}\}$ describes the identity of the two satellites $N_{S1},N_{S2}$ and the time $t_{s}$, $t_{e}$ when the link is built and broken. Naturally, the difference $t_{l}=t_{e}-t_{s}$ represents the survival time of the link.

In fact, the topology change of the satellite communication network occurs at some specific moment. Therefore, the time of the simulation model could be discrete. In [35], the topology is fixed during a period which is called timeslice $T_{i}=[T_{si},T_{ei}](i=1,2,3\ldots)$, where $T_{i}$ is the $i$-th timeslice in the simulation time and $T_{si}$ is the time at which it starts and $T_{ei}$ is the time it ends. Naturally, $T_{s(i+1)}=T_{ei}$. Based on this, the time-variant topology can be modeled like $G_{d}=\{T,\{S_{1},S_{2},\ldots,S_{M}\},\{T_{1},T_{2},T_{3},\ldots\},\{G_{s1},G_{s2},G_{s3},\ldots\}\}$, where $T$ is duration of simulation time. $\{S_{1},S_{2},\ldots,S_{M}\}$ is the set of satellite nodes and $M$ is the total number of them. In the $i$-th timeslice, the model $G_{si}(i=1,2,3\ldots)$ describes the satellites(nodes) and links as $G_{si}=\{T_{i},\{S_{i1},S_{i2},\ldots,S_{im}\},\{L_{i1},L_{i2},\ldots,L_{in}\}\}$. It indicates that in the $i$-th timeslice, there are $m$ satellites bulit $n$ links. The nodes are $\{S_{i1},S_{i2},\ldots,S_{im}\}$ and links are $\{L_{i1},L_{i2},\ldots,L_{in}\}$. Based on this model, the network topology is a dynamic graph, and some algorithms of OEC, such as computation offloading, mobility management, and resource allocation, can be simulated. Furthermore, the weight of links in this model can be defined according to particular needs.

The advancements in free space optical (FSO) communication have paved the way for a next-generation technology known as laser communication, which promises to improve communication rates and ranging accuracy in satellite network. Laser ISLs offer significant advantages over traditional microwave communication, including large bandwidth, high data rate, ease and speed of deployability, and low power. The first laser ISLs were demonstrated in 2008 between Terra SAR-X and NFire, two LEO satellites located 5500 km apart and moving at a speed of 25,000 km/h, achieving a capacity of 5.5 Gbps [57]. Studies have shown that laser communication technologies are superior to radio frequency (RF) technologies in terms of data rate, secure communication, and weight reduction [58, 59, 60, 61]. Table IV provides a summary of the comparative features between laser and RF communication. One of the most promising applications of laser communication is SpaceX’s Starlink project, which plans to use inter-satellite laser communication technology to provide internet services up to 1 Gbps with the deployment of 12,000 satellites [29].

A satellite is typically developed for a single or a small number of specific missions, which determine its capabilities, functions, and operations. The integration of hardware and software on a satellite is tightly coupled, making the whole system inflexible and non-universal. However, certain architectural components, such as communication systems and antennas, can be shared between different missions. The software-defined satellite (SDS) has been proposed as a potential solution to the aforementioned challenges, improving satellite network efficiency by using software-defined radio as its universal hardware platform and defining its functions via software [62]. China’s Chinese Academy of Sciences (CAS) has developed and launched the experimental Tianzhi series of software-defined satellites. The Tianzhi No. 2 D-star is capable of supporting super-resolution image reconstruction on the satellite due to its 40 TOPS computational power and only 19 kg weight. The successful application of SDS has facilitated the development of smart satellites. However, conventional task scheduling requires ground stations to pre-arrange instructions, which are uploaded to the satellite when it comes into range. As satellite tasks become more complex, this method is inefficient. In the future, autonomous task scheduling could alleviate these limitations, enabling satellites to make decisions online with minimal manual intervention. Advances in smart satellite research, such as the creation of satellite computing network [63] and space-air-ground integrated network, will stimulate numerous OEC applications.

In recent years, researchers have been drawn to SAGIN, a network architecture that interconnects space, air, and ground segments to take advantage of their respective strengths, such as the space network’s extensive coverage and the air network’s flexibility [8]. As illustrated in Fig. 4, SAGIN consists of three primary segments: space, air, and ground, along with the addition of marine communication network in the SAGSIN [64]. Smart satellites, particularly LEO satellites, play a crucial role in SAGSINs by offering irreplaceable services for the IoT [65]. Therefore, researchers are concerned with OEC in SAGIN. In the architecture proposed by [19], satellites can be divided into edge and cloud nodes and sliced through software defined Dnetwork or network functions virtualization (SDN)/(NFV) technology.

Augmented reality (AR) and virtual reality (VR) display exhibit potential to trigger attractive applications, including but not limited to health care, education, engineering design, manufacturing, retail, and entertainment [76], and urtra-high definition (UHD) video can provide better visual differentiation and better depth perception for users [77]. However, they face many challenges on complex computation operations, including encoding/decoding and rendering, thus introducing a longer processing time [78]. Edge computing provides opportunities for these applications by transfering part of computing pressure to edge device. Howerver, terrestrial networks may fail to provide ubiquitous coverage to surban and rutal areas, and the latency increase obviously with the transmission distance. As an effective solution, OEC provides ubiquitous access and ultra-long distance with low-latency transmission, opening up further possibilities for these applications [79].

Celestial [18] evaluated the performance of WebRTC video meetings using OEC based on the Starlink network. Three users sent a high-definition video stream with 2.6 Mb/s while receiving video from the others. The experiment showed that the end-to-end latency of both the satellite servers and cloud servers for at least 80% of the video conference duration was below a maximum round-trip time (RTT) of 16 ms and 46 ms, respectively. This confirms that satellite servers have the potential to significantly improve the quality-of-service (QoS) for clients in latency-sensitive applications.

The Internet of Vehicles (IoV) comprises a combination of vehicle auxiliary devices and roadside infrastructure that aims to enhance the safety, security, and efficiency of roadway transportation. It also offers different services to various types of vehicles, including private cars, taxis, and buses. Satellites have become a vital element in providing positioning and navigation services for vehicles, and they are an essential component of IoV [80]. Yu et al. [81] suggest addressing the challenge of providing vehicular networking services in remote or disaster-stricken areas using edge computing and satellite network. Their proposed fine-grained joint offloading and caching scheme leverages the orbit-ground collaboration to provide communication services for vehicles in rural areas, isolated islands, and remote places where terrestrial infrastructure is not available. The scheme aims to provide real-time Enhanced Cellular Satellite Assisted Global Navigation System (EC-SAGINs) services for terrestrial vehicles.

The integration of multiple information, including satellite navigation, road conditions, and traffic flow, among others, enable the implementation of autonomous driving technology, with artificial intelligence (AI) serving a notable role in decision-making processes [79]. However, AI-based decision-making requires significant computing resources. Edge computing has emerged as a potential solution, providing lower-latency and greater computing resources for autonomous driving. Satellite edge computing could serve as an auxiliary solution to the inadequate and uneven distribution of ground computing resources. This auxiliary process can provide a more stable and reliable AI decision-making scheme for autonomous driving.

Satellites are essential for capturing earth data in various fields, such as weather, vegetation and urban development. The data generated is growing rapidly in terms of size and variety, which creates immense pressure on space-ground communication [82]. However, OEC can analyze this data, provide near real-time analysis, and significantly enhance decision-making processes. One specific application of this technology is the real-time monitoring of natural disasters. By quickly detecting and transmitting critical information to ground stations, response times can be greatly improved [40]. Additionally, satellite edge computing can identify invalid images caused by cloud cover in orbit, which reduces the cost of data transmission to the ground [15]. Environmental monitoring is another potential application of satellite edge computing. Tracking changes in land use, forest cover, and ocean temperature provide an opportunity to quickly identify changes and send alerts to stakeholders for timely action. Moreover, multiple agile satellites can utilize satellite edge computing to solve the scheduling problem, enabling on-orbit decision-making and intelligent mission planning [83]. These applications demonstrate the potential of satellite edge computing to enhance the efficiency and effectiveness of earth observation missions.

Scientific Experiments satellites such as DArk Matter Particle Explorer (DAMPE), Quantum Experiments at Space Scale (QUESS), etc. are used to gather data for scientific research. Scientists obtain a large amount of experimental data from these satellites for space exploration and other missions, which often require transmission via satellite-ground links for subsequent analysis on the ground, a process that is inefficient. Edge computing on satellites, however, seeks to minimize data transmission and even enable in-orbit data analysis and processing.

Several scientific experiment satellites, including the Dark Matter Particle Explorer (DAMPE) and Quantum Experiments at Space Scale (QUESS), are instrumental in gathering data for space exploration and other scientific research missions. Although a considerable amount of experimental data is obtained from these satellites, transmitting the data via satellite-ground links for subsequent analysis on the ground is often an inefficient process. Therefore, OEC can assist scientific satellites in completing on-orbit data processing, reducing unnecessary data transmission.

With the increasing demand for data-driven solutions in various fields, such as agriculture, healthcare, and environmental monitoring, federated learning has emerged as an attractive technique that addresses the challenges associated with centralized data processing and privacy concerns related to traditional machine learning methods. Satellite federated learning is a novel approach that combines the power of satellite technology and machine learning to enable efficient and privacy-preserving data analysis across geographically distributed data sources. For instance, in environmental monitoring, satellite federated learning can analyze multiple sources, such as satellite imagery, ground sensors, and weather stations, to monitor air and water quality, track deforestation, and predict natural disasters. By federating the data and applying machine learning algorithms, accurate models for environmental prediction and monitoring can be developed, while protecting data privacy. In this paper, we summarize three typical architectures of satellite federated learning, as shown in Fig. 9, and discuss their characteristics. Moreover, we provide a comparison of selected articles that may facilitate further study in Table V.

Razmi et al. [85] proposed using satellite federated learning (FL) based on LEO Mega-Constellation. This technique enables data to stay in the satellite instead of being downloaded to the ground station, saving time and communication energy. The FL-Based Architecture typically consists of two planes: satellite and ground, and Chen et al. [86] explored their roles in FL. It shows that using a ground station as the server is more feasible than LEO satellites. However, using LEO satellites as the server results in less delay and faster convergence. Jing et al. [87] introduces a new FL architecture that consists of UAV-LEO-MEO. Fig. 9(a) shows the architecture of the GS-LEO without ISLs, which is the basis of most studies [85, 88, 89]. The most significant drawback of this architecture is the issue with satellite idleness and model staleness because the global model update needs to wait for all of the client satellites to upload their local model.

Fig. 9 (b) illustrates the GS-LEO architecture with ISLs. In this architecture, client satellites can update the model directly if at least one of them has a connection to the GS [90, 84]. Specifically, more research has been conducted on intra-orbit ISLs, rather than inter-orbit ISLs because the latter are highly dynamic, making them vulnerable to Doppler shift and requiring more cautious connection planning [91].

LEO satellites have a short duration of visiting a GS, which is typically in the range of tens of minutes. Consequently, Jing et al. [87] replace GS with MEO to provide more visit time and reduce transmission delay and energy consumption. Additionally, they utilize high attitude platforms (HAP) and UAVs as an integral part of the server-side. This paradigm is illustrated in Fig. 9 (c). Similarly, Cheng et al. [79] focuse on a space/aerial-assisted computation offloading algorithm based on unmanned aerial vehicle (UAV)-satellite-ground architecture.

Fig. 9(a) and (b) are summarized as space/ground-assisted architecture, while Fig. 9(c) is categorized as space/aerial-assisted architecture. The former case is suitable for various earth observation applications, including urban planning, disaster management, and climate change mitigation [92]. It helps to avoid the transmission of large volumes of data to ground stations. The latter case provides several advantages, including being less expensive, offering better visibility, allowing for better communication, and easy relocation [93]. Additionally, it can facilitate edge computing for unmanned aerial vehicles (UAVs) and airplanes. The rest of Fig. 9(c) comprises the LEO-GEO architecture that eliminates the aerial network segment. The LEO-GEO configuration is useful for autonomous mega LEO satellite constellation management [84]. Besides, Tang et al. [94] introduced the potential advantages and applications of FL in SAGIN. They discussed how FL can assist in solving resource scheduling problems in SAGIN. In summary, all these architectures aim to address the challenge of transmitting huge volumes of data from LEO satellite constellations, which leads to excessive spectrum occupancy.

Resource optimization in satellite edge computing aims to minimize latency and energy consumption. As an example, Cui et al. [105] simulated a satellite equipped with a CPU capable of 1000 cycles/bit, an energy consumption rate of $3\times 10^{-10}$ J/cycle, and a communication capacity of 10Mbps. Through their experiments, they explore the correlation between a satellite’s computational capability and its task execution time. Moreover, [100] examines the commercial benefits of utilizing advanced resource allocation algorithms in OEC.

Computation offloading is an important strategy for improving the performance and energy efficiency of satellite communication systems. To evaluate the effectiveness of offloading algorithms, researchers often use metrics such as convergence speed, average latency, and energy consumption. For instance, previous work has explored the impact of task quantity and satellite computational ability on latency, and has evaluated the performance of reinforcement learning algorithms in this context [106, 79, 107]. Other studies have focused on the energy consumption of ground users in scenarios where offloading involves interactions with them [7]. By considering factors such as the number of users, the computing and communication capabilities of edge devices, and the types of tasks being offloaded, we can gain insights into the time and energy costs of different offloading strategies.

In satellite federated learning, algorithms are designed with a focus on convergence rate, accuracy, and the effect of dataset distribution, similar to its ground counterpart [89]. Currently, popular datasets used in FL are MNIST and CIFAR-10, so as the satellite FL. Researchers use these datasets to ensure comparability and referenceability in comparing the performance of different algorithms. Moreover, to better meet the requirements set by operational environment characteristics, FMoW [108], a remote sensing image dataset was utilized by Jinhyun et al. [89]. Typical network models employed in satellite federated learning consist of mainstream image classification network such as DenseNet-161 and ResNet-18, as well as simpler neural network like the convolutional neural network (CNN) and multilayer perceptron (MLP).

The simulation of satellite network plays a pivotal role in the development of edge computing. For instance, the simulation tool, SNS3 [109], has been developed for this purpose by leveraging the widely used network simulation tool, NS-3. SNS3 is proficient in performing simulation tasks with various satellite link configurations, and analyzing the network at the physical layer. Furthermore, Kssing et al. proposed Hypatia [110], which is a network simulation tool specifically designed for simulating large-scale low earth orbit constellations based on NS-3. Hypatia incorporates unique characteristics of low earth orbit constellations, such as high-velocity orbital motion, to simulate and visualize their network behavior. Likewise, SILLEO-SCNS [111] concentrates on designing inter-satellite and satellite-ground station network and implements network visualization based on VTK.

At the time of writing, Celestial [18] is currently the only testbed available for satellite edge computing. Celestial uses the SILLEO-SCNS network simulator to accomplish network simulation tasks for large-scale low earth orbit constellations, while carrying out edge computing simulation based on microVMs. One unique aspect of the testbed is its ability to execute a high-definition video conference task based on the Starlink satellite constellation, with servers deployed across South Africa, Cameroon, Nigeria, and Ghana. The results show that OEC effectively lowers the latency in the meeting. Celestial also provides support for simulation related to OEC challenges, such as resource allocation, computation offloading, and satellite routing.

Traditional satellite payloads are typically tailored to specific missions, resulting in platform and system incompatibilities when implementing edge computing. Furthermore, updating these payloads’ systems and functions poses a challenge. Solar power serves as the primary source of energy for satellites, and proper energy management is necessary given the variance in light exposure and duration. To reduce mission failure rates, satellite energy management must not only minimize consumption, but also consider whether the remaining energy can support mission completion. Additionally, the complex space environment presents challenges for chip fabrication and heat dissipation, both of which impact satellite computing performance.

Recently, large-scale low orbit satellite constellations have become a prominent trend in intelligent satellite network. The massive scale of satellite network presents challenges to optimization algorithms. As mentioned earlier, typical NP-hard optimization problems exhibit exponential growth in solution complexity as the number of nodes increases, presenting significant challenges for optimization algorithms in large network. Furthermore, mobility management costs increase significantly within these large-scale network. Using one or a few ground stations to manage the trajectory information of all nodes is no longer practical.

In recent years, all work related to OEC assumes that all the nodes of the satellite are idle and available, but this is a simplification of actual application scenarios, which may be much more complex. Firstly, satellite computing power is varied, and resource distribution is unbalanced. Secondly, the possibility of satellite malfunctions or other factors leading to nodes going offline complicates computations. Finally, in actual applications, a satellite network undertakes diverse services, meaning not all accessible satellites can participate in any given task. Therefore, future research should approach computation offloading and experiments in a manner that accounts for these complexities.

At present, research on OEC is restricted to ground simulation that might not represent the space environment’s actual scenario. The mission of “TianSuan” series satellites [15] is to create a computing platform in space, providing an opportunity for OEC researchers. Nevertheless, performing experiments on actual satellite equipment continues to be expensive. Moreover, achieving OEC under the envisioned “large-scale constellation” seems like a distant undertaking.

The utilization of software defined satellite technology represents a significant advancement in satellite technology. This technology allows for the installation of plug-and-play payloads, as well as the ability to load software on demand. Additionally, satellite functions can be conveniently redefined through software updates to adapt flexibly to various tasks and user needs. Furthermore, this suggests that future satellite network will evolve towards a software-defined architecture. Within this context, satellite resource management by OEC will become even more flexible.

In recent years, free-space optical communication has garnered significant interest as it greatly enhances the communication efficiency of satellite network. These types of satellites provide higher bandwidth and lower latency through laser link communication, making them suitable for scenarios which require high communication needs like large-scale satellite distributed computing and collaborative tasks that demand prompt response times.

Unlike ground facilities, satellite payloads’ hardware must withstand extreme temperature fluctuations and strong radiation. In addition, volume, weight, mass and mass are subject to strict requirements since they ultimately determine the cost of manufacturing and launching a satellite. With the growing demand for computing power, it is imperative to study and develop efficient hardware for satellite payloads.

Due to the aforementioned limitations on satellite hardware, there is a particular need for lightweight AI models in satellites. The design of such models should consider task types, data, and communication resources required for OEC. Additionally, AI tasks in remote sensing differ significantly from those on the ground. Satellite data is sampled without identification, meaning some observations are of high-value to an application while others are of low-value. The OEC system Kodan, which overcomes computational bottlenecks by maximizing data value density [73], inspired us to explore the design space of data selection in remote sensing tasks.

Resource optimization is a popular research topic in terrestrial network. Satellites have more constraints compared to terrestrial equipment. As a result, researchers are encouraged to investigate advanced and practical optimization algorithms. Existing research uses deep reinforcement learning to solve these problems, but no related work has experimented on larger constellations. Increasing the number of nodes may hinder further applications due to the parameter and convergence speed limitations of deep reinforcement learning network. Therefore, it is imperative to conduct experiments on larger satellite network to align with future practical industrial development trends, requiring better performing optimization algorithms.

The use of intelligent satellite network is being actively explored for potential application services, including advanced business needs on the ground, such as AR/VR, UHD vedio. OEC offers remote areas the possibility to apply these services. In addition, OEC breaks the inefficient “bent-pipe” architecture, where satellites collect data and then transmit it to the ground for processing. OEC provides the possibility for autonomous processing for satellites, which saves satellite-ground communication costs and improves mission execution efficiency. In the future, when large-scale LEO satellites are commercialized, satellites can access computing power services for ground users. Furthermore, the application of satellite federated learning can support satellites to update AI models autonomously. Users can implement online update of the model according to their needs.

At present, OEC is still a concept, and only one comprehensive testbed work on OEC has been proposed [18]. However, due to the large number of nodes in satellite constellations and the complex network topology, developing OEC models on real devices are not feasible. Hence, with the diversification of application scenarios and the introduction of new algorithms, there is an urgent need to develop the more testbeds and improve their performance. To achieve a test-friendly OEC testbed model, it should have the following characteristics:

• •

  Accurate satellite constellation simulation. Satellite network topology and satellite mobility models are one of the most important features of OEC tasks. Accurate satellite constellation simulation can realistically simulate the movement of nodes in space. Typically, it can support professional satellite simulation tools such as STK, or develop independent satellite constellation simulation tools.

• •

  Excellent platform compatibility. The testbed should be deployable on actual physical devices to achieve a more realistic simulation. Therefore, the testbed needs to be compatible with various platforms, such as personal computers and various embedded devices.

• •

  Customizable resource management. Limited satellite resources have become a recognized challenge. The testbed should support simulation of satellite edge computing performance under various resource conditions.

• •

  Support for actual task execution. OEC is designed for a variety of services, such as deep learning training and inference, channel encoding and decoding, intelligent task scheduling, etc. The testbed should support simulations of these actual applications.