1. Introduction
In recent years, 5G technology has led to the exploration of integrating ground networks and satellite communication networks [
1]. Satellite networks can provide global coverage and high-speed transmission, which can improve the performance of ground networks. However, the integration of 5G and satellite networks also has many challenges, such as spectrum, architecture, and resource challenges. With the research of 6G, satellite networks have become more attractive, as 6G aims to achieve higher performance than 5G, such as 1 Tbps rates and 0.1 ms latency [
2]. To meet these requirements, satellite networks need to be developed and improved in aspects such as number, orbit, antenna, frequency, and communication. Therefore, the integration of 5G and satellite networks could offer a promising network architecture that can provide better communication and more scenarios for users. With the development of 6G, satellite networks will also experience new changes and opportunities, bringing new possibilities for the future communication industry [
3].
There are three categories to classify satellite networks, which are based on the orbit altitude: low Earth orbit (LEO) satellite networks, medium Earth orbit (MEO) satellite networks, and geostationary Earth orbit (GEO) satellite networks [
4]. Recently, LEO satellite constellations have emerged as a significant technology capable of enhancing communication. Satellite networks have the ability to compete with ground networks owing to the higher speed of inter-satellite links (ISL), which surpass the capabilities of optical fibers used in ground networks by reaching the speed of light in a vacuum [
5]. In addition, operating at an altitude of 2000 km above the Earth’s surface, LEO satellites can significantly reduce the round-trip time (RTT) and waiting time, contributing to their advantage of low latency. However, the dynamic nature of LEO constellations presents new challenges that need to be overcome.
The next-generation Internet, based on terrestrial networks and mega-satellite networks, is rapidly growing as an up-and-coming future network with the development of commercial satellites such as Starlink, OneWeb, and Telesat [
6]. LSNs offer key advantages over terrestrial network solutions, including global network access, low deployment costs, and 24/7 availability. Moreover, LSNs provide lower communication latency and higher data rates compared to GEO and MEO satellite communication systems [
7]. The advantages of LSNs make them a crucial infrastructure in the development of the next generation of space–air–ground integrated networks. Given the LSN’s ability to support a range of applications, including telecommunication, remote sensing, location, and navigation, it is essential to enhance the performance of LSN facilities in response to their dynamic topology.
The issue of routing strategy and ground station geographic planning is inevitable [
8] in LSNs; the former controls the transmission of traffic and the latter improves LSN performance by ensuring satellite coverage [
9]. LSNs exhibit different behavior compared to terrestrial networks due to several factors. Firstly, the mobility of satellites causes frequent variations in the network topology. Secondly, the unstable relative position and severe space environment can potentially degrade the link performance. Thirdly, satellites in LSNs have a limited bandwidth and memory and need to minimize the communication overhead as much as possible. Lastly, the scale of LSNs makes it relatively challenging to maintain traffic forwarding [
10]. Therefore, routing algorithms suitable for LSNs need to prioritize latency performance and packet reception success rates [
11].
Traditional networks have undergone extensive research with regard to routing optimization [
12,
13,
14]. However, the algorithms developed for these networks cannot be directly used for LSNs due to their unique features. Therefore, this paper focuses on studying the routing optimization problem specifically in LSNs, taking into consideration their distinct features. The main contributions of this paper are as follows:
The low-orbit satellite network model is defined as an unsplittable multi-commodity flow problem (UMCF) [
15], which is then modeled as an MILP problem.
We propose a routing strategy called GAGR, which is based on GA assisted by ground access optimization; it incorporates GA to obtain an approximate optimal solution.
Before sending a flow from both the source and destination ground stations, we select three nearby satellites as candidate satellites, replacing the traditional strategy of selecting the nearest satellite.
We compare the performance of the GAGR routing strategy, virtual-topology-based shortest path (VTSR) routing strategy [
16], and Floyd–Warshall (FW) [
17] routing strategy under the UDP protocol and TCP protocol through experiments, and the experimental results demonstrate that our GAGR routing algorithm exhibits improved performance in terms of a reduced RTT, increased flow reception rate, and lower ISL occupancy rate.
Given the extensive use of abbreviations in this paper, we provide Abbreviations section containing their definitions for reference.
The rest of this article is organized as follows.
Section 2 introduces the related work on satellite network routing.
Section 3 introduces the network model and problem definition. In
Section 4, we introduce the GA solution algorithm for UMCF and discuss the reasons for introducing three candidate satellites for ground stations.
Section 5 describes the experiments and results. We summarize the paper in
Section 6.
3. Problem Definition and Network Model
3.1. Problem Definition
We define the satellite routing problem as an UMCF problem; it involves a graph with undirected edges and capacities assigned to those edges, and each satellite has a maximum of four neighboring satellite nodes that communicate with each other through ISL. The number of neighboring nodes for each satellite depends on the number of visible nodes for that satellite in the current time step. Additionally, there is a set of commodities, each with a source node, a destination node, and a demand value. In order to describe the satellite routing problem as an UMCF problem, we treat the traffic forwarding requests made by each ground station as commodities, which include three attributes: source ground station, destination ground station, and traffic demand. The goal is to find a unique path for each commodity to route their traffic demand from the source node to the destination node, without exceeding the capacity limit on the link when passing through the path. This problem has been extensively studied in the field of network optimization.
The UMCF at hand is known to be NP-hard, which means that finding optimal solutions for instances with a large number of nodes and commodities is extremely challenging. However, there are heuristics and approximation algorithms available that can still provide high-quality solutions for such instances. In this paper, we will focus on the GA.
3.2. Network Model
To model an LEO satellite network, we utilize undirected graph , where N represents the set of satellites and E represents the set of all possible links connecting the satellites. Additionally, we define a set of commodities, denoted as , characterized by their source, destination, and demand. We summarize the main symbols and their explanations in Abbreviations section.
In order to ensure that each commodity can only choose one path to forward traffic in the end,
which
indicates whether there is traffic on link
e for commodity
i. In order to ensure that the traffic can pass through the inter satellite link normally, we set the optimization objective as a constraint:
where
represents the overflow on link
e. To guarantee that the incoming and outgoing traffic of each commodity flowing through each satellite node is the same, we have the following constraint:
where
represents the incoming traffic of node
n, and
represents the outgoing traffic of node
n. In order to ensure reasonable traffic overflow when traffic passes through link
e, we have the following constraint:
4. The Proposed Routing Strategy
In this section, we will introduce our inter-satellite routing algorithm and the first and last hop selection methods for ground stations to access and exit the constellation. Our inter-satellite routing algorithm is a heuristic method based on GA, which can effectively solve the UMCF problem and assign a unique path for each commodity. Our first and last hop selection methods for ground stations are simple methods based on the greedy strategy, which can select the optimal satellite node as the first and last hop according to the source and destination of the commodity. We will describe these two methods in detail in
Section 4.1 and
Section 4.2, respectively. Before introducing the GA to our problem, we will introduce linear relaxation for MILP, the constraint in our problem.
represents that each commodity can only select a single fixed path to send traffic. This is an NP-hard problem. By replacing this constraint with
, we obtain the linear relaxation form of the problem. In the linear relaxation problem, each commodity is allowed to split its traffic among multiple paths. This linear relaxation problem can reduce the complexity in solving the original problem.
4.1. Routing Algorithm
The GA offers a practical approach to performing a global search through multiple local searches. This algorithm relies on operations such as selection, crossover, and mutation to optimize the solution. In each loop, the GA generates a set of chromosomes, which represent candidate solutions for the optimization problem. The goal is to use the fitness value of each chromosome as a standard to find an approximate optimal solution over several generations.
The algorithm process is shown in
Figure 1. In each generation, the roulette wheel selection method is used to select elite chromosomes. The crossover operation from elite chromosomes then selects the optimal chromosomes as parents. In this operation, the genes of the two parent chromosomes are randomly exchanged with a predefined probability, giving rise to next-generation chromosomes. To maintain diversity and feasibility in the set of chromosomes, the genes in the next generation of chromosomes undergo mutation based on the following three principles:
Mutation operates on a randomly selected individual gene, i.e., an individual ISL;
The mutated gene is only allowed to be replaced with one of the three alternative ISLs on the same path;
We ensure that this hop after mutation is feasible.
Once the crossover and mutation operations are completed, the next-generation chromosomes undergo evaluation based on the fitness value and problem constraints. This evaluation determines the fitness values of the chromosomes and prepares them for the next loop of operations. This process continues until the optimal chromosome remains unchanged for a specified number of generations.
Table 1 summarizes the GA terminology that we will use in the algorithm.
The specific execution process of the routing algorithm is displayed in Algorithm 1, and we can see that it is generally executed alternately between two steps:
Solve the linear relaxation problem, where each product can have multiple paths that can forward the flow;
Utilize GA to ensure the uniqueness of the solution after linear relaxation, so that each commodity has a unique path.
The algorithm continues to execute until all commodities have a unique traffic forwarding path. Due to the fact that solving relaxation is easier than solving the satellite routing problem, our algorithm can achieve faster convergence times. Referring to the worst-fit memory allocation concept, we rank the commodities in descending order of demand. Studies have demonstrated the significant influence of this order on the heuristic’s solution quality. After obtaining a linear relaxation solution, each commodity has several candidate paths.
Firstly, we use a uniform-length integer-coded method to encode the chromosomes. The number of genes on each chromosome corresponds to the number of ISLs present on the candidate paths. The gene located at each position within the chromosome represents a specific ISL.
Secondly, the algorithm maintains high-quality genes by performing selection operations on the chromosomes. The fitness value serves as the basis for the selection process. We define the fitness equation as follows:
where
represents the minimum remaining capacity among all candidate paths, and
represents the minimum delay among all candidate paths. Meanwhile, the algorithm employs the roulette wheel selection method for the chromosome selection operation. We define the probability that the candidate path
p is selected, which is
Based on Equation (6), the operation of roulette wheel selection prioritizes chromosomes with higher fitness values and eliminates unreachable chromosomes. In this way, the operation chooses chromosomes with better genes to be carried over to the next loop. Furthermore, our algorithm employs the best chromosome retention strategy. This strategy involves directly including the chromosome with the highest fitness value in the next population. Subsequently, the chromosome with the lowest fitness value is discarded to preserve the elite chromosomes.
Finally, we generate the next generation through crossover and mutation operations, which are important components of GA and help to maintain population diversity. For crossover operations, we employ the loop crossover method. This method involves selecting two low fitness chromosomes from the parents and performing a crossover operation on randomly selected genes at the same position on their chromosomes, resulting in the generation of new chromosomes. For mutation, we use single point mutation, i.e., when a candidate path is selected for the mutation operation, the chromosome is a candidate path, selecting one link on this path and randomly changing to one of the other three links on the forwarding node of this link.
4.2. Ground Station Access
When a GS accesses the constellation, the satellite selection is typically based on the minimum distance. If multiple satellites are visible and accessible to a GS at a given moment, the one with the closest distance is directly chosen for data transmission. The satellites within a constellation are interconnected in a grid pattern, which allows them to maintain connectivity with several satellites consistently. Nevertheless, if satellites are moving in reverse directions, it is impossible to establish stable communication, even if they are neighbors. This means that two GSs that are geographically close may need to use a longer satellite path for communication. On the contrary, considering several adjacent satellites to the GS, different paths may be found, such as the source and destination satellites being in the same direction of motion. We present these two possible scenarios in
Figure 2.
To enhance the performance of existing routing algorithms, we propose conducting additional computations on the allowed incoming and outgoing hops within the constellation. The specific process of the algorithm is shown in Algorithm 4. The specified three nearest candidate satellites to all ground stations are stored in the memory at each time step. By comparing the three nearest satellites to the GS for all considered pairs, this process defines two scenarios as follows.
- (i)
If the source GS and destination GS are located within the range of the three nearest candidate satellites, only one satellite is needed to establish the connection. In such cases, we can optimize the routing by directly utilizing this satellite node for forwarding.
- (ii)
If two of the candidate satellites from two ground stations have direct inter-satellite links (ISLs), it means that these candidate satellites are neighboring each other in the constellation. In such scenarios, we can designate this pair of satellites as the source node and destination node.
Algorithm 4: GS Access/Exit Algorithm |
|
5. Performance Evaluation
For LSNs, the communication delay is a key optimization indicator. Compared to the link capacity, reducing the delay is more important in improving the network performance. Therefore, we assigned weight to the delay, and weight to the remaining capacity in the simulation, giving more importance to the delay.
The performance of the proposed GAGR algorithm is assessed using several performance indicators, such as RTT, ISL utilization, the ratio of received/sent traffic, and the hop count, and compared to the VTSR algorithm [
16] and FW routing algorithm [
17]. VTSR is a DFS algorithm based on an SDN and virtual topology; the FW algorithm utilizes the idea of dynamic programming to calculate the shortest path between any two satellite nodes whenever the topology state changes.
To evaluate its performance in terms of the link capacity and flow distribution within the network, we utilize the UDP protocol and TCP protocol. We selected 100 pairs of cities from the set of the 100 most populous cities to represent 100 commodities, and these three routing algorithms were simulated for 60 s in the Telesat simulation constellation provided by Hypatia v1. Telesat’s satellites use laser ISLs to communicate with each other. A radio up-link from the source GS to the ingress satellite, some laser ISLs, and a radio down-link from the egress satellite to the destination GS form an end-to-end path between two GSs. In order to simulate a case where the maximum capacity is insufficient, we set the flow size of each commodity to about 100 Mbit, which is much higher than the maximum capacity of the link, 32 Mbit/s. In this way, we can observe the behavior and bottlenecks of the link under a high load. Its other attributes are shown in
Table 2.
The simulation environment is created using Hypatia on a personal computer configured as follows: Core i5-8300H 2.3 GHz CPU, 16GB RAM, and Ubuntu 20.04. Hypatia converts the continuous motion of satellites and the consequent changes in their paths into discrete processes. Although the delay on the path is constantly updated, the forwarding status is only recalculated at a fixed time step. Moreover, we use commercial solvers such as Gurobi 10.0.3 to solve the linear relaxation problems.
Firstly, we calculated the RTTs of all commodities under the TCP protocol under the three routing algorithms, as shown in
Figure 3. The simulation time is within [0, 45] s; our proposed GAGR algorithm has an average RTT of 404.02 ms, while the FW algorithm has an average RTT of 469.57 ms, and the VTSR algorithm has an average RTT of 453.93 ms. The average RTT ratio of the GAGR algorithm to the FW algorithm is 0.86, and the average RTT ratio to VTSR is 0.89. The average RTT of our proposed routing algorithm is better after 45 s; our proposed GAGR algorithm has an average round-trip time of 334.93 ms, while the FW algorithm has an average round-trip time of 482.01 ms, and the VTSR algorithm has an average round-trip time of 438.84 s. Compared with the FW algorithm, the average RTT ratio of GAGR is 0.69, and, compared with the VTSR algorithm, it is 0.76.
In
Figure 4, we show the average round-trip times of 100 commodities within the simulation time based on the UDP protocol and sort them in ascending order. During the entire simulation time, the average RTT of the GAGR algorithm is 378.51 ms, the average RTT of the FW algorithm is 472.64 ms, and the average RTT of the VTSR algorithm is 448.42 ms.
In
Figure 5, we show the average round-trip times of 100 commodities within the simulation time based on the TCP protocol and sort them in ascending order. Among the first 87 commodities, the ratio of the average delay of the GAGR algorithm to the average delay of the FW algorithm is 0.8, and the ratio of the average delay of the VTSR algorithm is 0.84. However, both the FW algorithm and VTSR algorithm have 13 commodities that do not arrive at the ground station, resulting in packet loss, while the 100 commodities of the GAGR algorithm do not experience packet loss.
Secondly, we calculate the ISL utilization of all commodities under the TCP protocol under the three routing algorithms, as shown in
Figure 6. We divide the simulation time into 10 s intervals, which can capture the dynamic changes in the network performance and the influence of our algorithm over time, showing the advantages of the GAGR algorithm in reducing the link utilization. Obviously, in all simulation intervals, the link utilization of the GAGR algorithm is lower than that of the FW algorithm and VTSR algorithm. The average link utilization of the FW algorithm is the highest, followed by that of the VTSR algorithm. Because the paths found by these two algorithms are not satisfactory, the number of forwarding hops is greater and more ISLs need to be passed. Similar to the situation under TCP, the monitoring link utilization under UDP is as shown in
Figure 7. The link utilization of the GAGR algorithm is still the lowest among the three, because the approximate optimal solution is found in the optimization and iteration process of the candidate path, which reduces the number of forwarding hops.
Furthermore, we sort each commodity based on its receive–send comparison according to the TCP protocol, as shown in
Figure 8. Among the receiving performance of 100 commodities, the average receive–send ratio of our proposed GAGR algorithm is 0.49, with that of the FW algorithm being 0.38 and the VTSR algorithm being 0.42. In the UDP protocol, as shown in
Figure 9, the average receive–send ratio of our proposed GAGR algorithm is 0.89, with that of the FW algorithm being 0.69 and the VTSR algorithm being 0.72. It can be seen that the GAGR algorithm has better throughput and provides more reliable network services.
Finally, we compare the three routing strategies in terms of the number of hops for each commodity, as shown in
Figure 10. FW requires the highest number of forwarding steps to complete a request, while GAGR has an average of 9.35 hops per commodity request and VTSR has an average of 9.44 hops per commodity request.
6. Conclusions
In this paper, we propose a routing strategy based on GA assisted by ground access optimization for LEO satellite constellations. The proposed algorithm aims to optimize the routing strategy for LSN systems with ISLs, which can reduce the RTT and improve the packet reception rate. The experimental results show that the proposed algorithm outperforms the traditional satellite routing algorithm in terms of these metrics. Moreover, the proposed algorithm can adapt to different network topologies and traffic patterns, which makes it suitable for various LSN systems.
In summary, the proposed GAGR routing strategy provides a promising solution for the optimization of the routing strategy in LSN systems with ISLs. Although our proposed GAGR routing strategy demonstrates significant performance improvements in our experiments, it still inherently suffers from the drawbacks of heuristic algorithms, making it prone to falling into local optima. Future work will consider payload reconfiguration, introduce onboard processing techniques [
36], incorporate machine learning techniques for predictive routing optimization to further improve the routing algorithm performance, strengthen the load balancing effect, increase the realism of experiments, and extend the proposal to other LEO satellite constellations.