Integrating Satellites and Mobile Edge Computing for 6G Wide-Area Edge Intelligence: Minimal Structures and Systematic Thinking

In terms of achieving edge intelligence, it is widely acknowledged that a complementary relationship between terrestrial-MEC and satellite-MEC networks is necessary. In contrast to the terrestrial network, the integrated satellite-MEC network mainly aims at application scenarios in remote or disastrous areas, as discussed before. The service requirements in these application scenarios have these unique characteristics. First, MTDs are expected to be dispersed over extensively wide geographical areas. The spatial distribution of their service requirements is remarkably sparser than that in the terrestrial network. A typical example is the buoys that collect hydrological information for ocean monitoring [5]. Moreover, the temporal and spatial distributions of the service requirements can be heterogeneous and highly dynamic. For instance, industrial automation and efficient disaster relief applications require aggregation of MTDs, and the positions of these aggregates vary over time. These unique characteristics render the system design of terrestrial networks inappropriate due to the low resource efficiency. Therefore, new challenges concerning the integrated satellite-MEC networks’ system designs require careful consideration.

In addition to the service requirement characteristics, the inherent limitation in communication and computing capability also poses challenges to the integrated satellite-MEC network. On the one hand, the communication capability of the satellite network is inherently limited. The transmission rate of satellite-ground links is relatively low due to the long propagation distance and limited bandwidth resources. In addition, the number of satellites in the network is restricted. This is because of the high costs of space launch and the limited orbit resources. Specifically, although a single low earth orbit (LEO) satellite is expected to have a similar level of throughput as a 5G base station, the number of satellites in a constellation is remarkably smaller [7]. This makes it difficult for the satellites to satisfy the global service demands. On the other hand, due to the limited capability of satellite communication, MEC servers on satellites and aerial platforms (APs) are necessary in the integrated network for delay-sensitive services. These MEC servers, especially the satellite servers, are under stringent restrictions in terms of size, weight, and power. Moreover, because of the severe electromagnetic radiation in space, the satellite servers need to be radiation-hardened [8]. This requires special design and incurs extra costs. These factors account for the limited computing capability in integrated satellite-MEC networks.

Fig. 2 shows a schematic diagram of the overall latency of a single offloading task, varying with the communication and computing resources provided. The resources are normalized, and the latency is obtained by linearly combining the reverse of both resources¹¹1In different use cases, the specific relationship between the overall latency and the network resources can be different.. We observe that the latency decreases quite slowly with both resources in the resource-adequate scenario. Therefore, configuring proper resources to match the service requirements of an offloading task can considerably improve resource efficiency. For a practical system of multiple offloading tasks, such a resource configuration evolves to a complex high-dimensional backpack problem. For the integrated satellite-MEC network, the hierarchical communication and computing resources, such as the diverse transmission links and the multiple tiers of MEC servers, render it difficult to match the resources with the service demands. Moreover, LEO satellites and aerial platforms are highly mobile in the network, which makes the network resources vary over time. The resource configuration problem becomes even more complex when the time dimension is accounted for.

As shown in Fig. 1(b), this minimal integrating structure consists of an AP equipped with an MEC server, a satellite, a gateway and multiple MTDs. The AP can be a UAV or a high altitude platform (HAP). With MEC deployment, the AP provides local computation services for MTDs. Due to the limitations of APs in terms of size and energy, the central processing unit (CPU) capability of an on-board MEC server is restricted. In addition, the AP can also relay the task data to the satellite, and the task data are further transmitted to a cloud server for computation. This minimal structure can be extended by considering multiple APs.

Compared to satellites, APs can provide high-speed communications, which leads to lower transmission delay. However, the coverage area of an AP is smaller than that of a satellite. From these perspectives, it can be deduced that the computing-in-forward-link structure is suitable for local-area computation services with ultra-low latency requirements. Examples of these services include robotic machine controls in industrial automation and real-time decision making in efficient disaster relief.

For such a minimal structure and its extensions, multiple challenges exist in terms of system design. First, it is of great importance to determine the on-board CPU capability, as well as the capacities of the satellite-AP link and the AP-user link. It should be noted that the capacities of the two links need to achieve a proper fit to improve resource efficiency. Moreover, in the multi-AP case, transmission links might exist among the APs, which further adds to the difficulty of configuring communication and computing resources.

Another challenge is that proper user association plays an important role in the system design of the multi-AP structure. One typical user association scheme is that each AP is equipped with an MEC server, and each user is associated with one AP. Ding et al. [9] considered such a scheme and investigated the joint optimization problem of user association, resource allocation, offloading assignment, and transmit precoding. In addition, we can also assume that each user has access to multiple APs and their on-board servers, forming a coordinated multi-point network. Moreover, some of the APs may not carry MEC servers and only serve as relay nodes.

Due to the mobility of APs, the trajectory of APs and the maximum coverage time also require careful consideration in system design.

This minimal integrating structure is composed of a satellite equipped with an MEC server, a gateway and multiple MTDs, as shown in Fig. 1(c). The MTDs’ computation tasks can be directly executed at the satellite or further uploaded for cloud computing. The satellite is highly limited in terms of size, weight and energy supply, which places stringent restrictions on the CPU capability of the on-board MEC server. There are several derivatives from this minimal structure. For instance, the space segment can be a constellation of LEO satellites. In addition, a geostationary earth orbit (GEO) satellite and a LEO constellation can coordinately provide edge computing services.

Compared with APs, satellites can provide a much larger coverage to serve more devices. In addition, satellites are in proximity to ground MTDs topologically compared with gateway stations. Therefore, the computing-on-orbit structure can well serve the requirements of computation tasks that span a wide area and require relatively low latency. For instance, path planning for large-area vehicles can utilize this structure to execute their computation tasks.

Multiple challenges need to be handled for this minimal structure and its extended structures. First, it is not as easy to place MEC servers on satellites as on UAVs or at gateways. On the one hand, to implement edge computing on orbit, on-board processing (OBP) capabilities (e.g., modulation/demodulation, encoding/decoding) are necessary on satellites. On the other hand, the severe electromagnetic radiation in space can induce bit-flips and even cause damage to the satellite payloads. This requires that the servers and the OBP modules should be radiation-hardened [8]. These factors add to the complexity of system design, especially when the limited size, weight and energy of satellites are considered.

With multiple satellites, the placement method of MEC servers is another important challenge. For instance, we consider the scenario of a LEO constellation providing services. Radiation-hardened MEC servers can be placed on all LEO satellites, which leads to high costs of MEC hardening. Another scheme is that the servers are placed on part of the satellites, while the other satellites merely relay data through inter-satellite links (ISLs). Adopting this scheme can reduce the MEC hardening costs, but the ISL costs will be higher. For the MEC server placement problem, this tradeoff requires careful consideration.

In the extended structure where the space segment is a LEO constellation, there are often multiple satellites in the MTD’s line of sight. For more computation-intensive applications, it is an intriguing problem to achieve simultaneous offloading to multiple satellites. Song et al. [10] investigated this problem and jointly optimized the task division and the transmission power to each satellite. From another angle, Wang et al. [11] proposed a game-theoretic approach to optimize the offloading strategy of multiple users.

Finally, LEO satellites are highly mobile and thus their coverage time can be quite limited. Therefore, handover of MEC servers may be needed, which is also a problem worth investigating.

As shown in Fig. 1(d), this minimal integrating structure consists of a satellite, a gateway equipped with an MEC server and multiple MTDs. MTDs offload their computation tasks to the gateway through the satellite relay. The computation tasks are executed in the gateway server or they are transmitted to the remote cloud. It should be noted that the MEC server in this minimal structure possesses higher CPU capability compared with the above two minimal structures, since the gateway station allows a larger server size and higher energy consumption. The minimal structure can be extended by considering multiple gateways.

With satellites as relays, the MEC server deployed at the gateway station can meet the demands of large-area computation services. In addition, the CPU capability of the MEC server is higher. However, the communication latency is relatively high since the data may traverse several ISLs before arriving at the gateway. Therefore, this minimum structure is most suitable for wide-area computation-intensive but delay-tolerant services, such as collecting and processing massive sensor data in environmental monitoring.

The computing-after-feeder-link structure and its extensions also raise problems in terms of system design. For instance, the MEC server placement problem is an important challenge. It is typically assumed that multiple gateways work separately. In this case, the CPU capability of each gateway server needs to be properly configured. In addition, the computing resource allocation of each server requires optimization. For instance, Zhang et al. [12] proposed a game-theoretic scheme to solve the joint offloading decision and resource allocation problem. Cui et al. [13] jointly optimized user association, task scheduling, and resource allocation. Another assumption is that the gateways are interconnected through high-speed fiber links. In this case, the transmission latency among gateways can be ignored, and the problem of server placement degenerates to deciding the total CPU capability of all servers.

On-demand adjustments of the integrated satellite-MEC network are of great importance for providing wide-area customized services [1]. As illustrated before, the temporal and spatial distributions of MTD users are sparse. In addition, the communication and computing capability of the integrated network are rather limited. Through dynamically adjusting the network to match the changing service requirements, the resource efficiency can be substantially improved. Because of this, the network can fully utilize its limited resources to provide ubiquitous services.

The current 5G network is mainly adjusted at two different timescales. First, mobile network operators conduct network planning at the timescale of months or years. Network planning adjusts communication and computing infrastructures, such as the number of satellites and the CPU capability of satellite servers. The second is the timescale of milliseconds or less, which is close to the channel coherence time. At such a timescale, the transmission parameters are adjusted, such as channel estimation and beamforming.

However, the service demands often change at a medium timescale (e.g., minutes or hours), in terms of number, spatial distribution and service type. Both network adjustments in the 5G network are incapable of matching these changes. Therefore, novel medium-timescale network adjustments need to be introduced in the integrated satellite-MEC network. Moreover, the network can be viewed as a nonlinear orchestration of the three minimal structures, as illustrated above. Therefore, we will establish an on-demand network orchestration framework, which adjusts the minimal structure orchestration at medium timescales to match the varying service requirements. Specifically, this can be achieved by adjusting the activation ratios of certain transmission links and MEC servers for instance. Table 1 summarizes the network adjustments, as well as their corresponding timescales and adjusted parameters.

Fig. 3 depicts an instance of medium-timescale minimal structure orchestration for four constant time periods. As shown in Fig. 3(a), multiple environmental sensors require updating their data for analysis in the first period. In this case, the activation ratios of gateway servers are high, while the satellite servers are turned off to save energy. Such a network adopts the computing-after-feeder-link structure. The second period sees a decrease in the number of transmitting sensors, as shown in Fig. 3(b). Therefore, the activation ratios of gateway servers are lowered. Fig. 3(c) shows that some mineral occurrences start exploiting resources in the third period. To provide industrial automation services, some satellite servers are turned on, and a UAV is dispatched for delay-sensitive computation offloading. The network is then adjusted to a combination of the computing-in-forward-link structure and the computing-on-orbit structure. In the fourth period, the activation ratios of UAV servers and satellite servers are further raised to adapt to the increase in service demands, as depicted in Fig. 3(d).

For medium-timescale network orchestration, some network parameters are adjusted over different time periods, which matches the change of service requirements. However, these parameters, such as the MEC activation ratio in Fig. 3, remain fixed during a time period. This requires that the parameters should be optimally designed to maximize the system performance of the entire period. To address this problem, we propose a process-oriented optimization method, where the process refers to a medium-timescale period. The design of this type of method is challenging, since the accurate process information can be difficult, if not impossible, to acquire. For instance, the process duration is much larger than the channel coherence time. Therefore, the full channel state information (CSI) during the process is considered random, and thus cannot be obtained accurately. To address this challenge, one promising direction is to introduce external information to assist the optimization. For instance, the large-scale CSI of the whole process can be conveniently acquired from a pre-established database, referred to as a radio map [14]. Process-oriented optimization can be conducted based on the large-scale CSI. Likewise, we also need a large-scale model that characterizes the service requirements, which is worth further investigation.