The size and complexity of Low Earth Orbit (LEO) SATCOM constellations introduces substantial new challenges for network traffic engineering. As the number of companies offering low latency SATCOM connectivity via LEO increases in number, the cost and performance will be key to achieving market share.
Delivering the best service to customers in this increasingly competitive marketplace is driving operators to consider new technologies to scale-up capacity while ensuring service performance and controlling expenditure Managing a LEO SATCOM network is challenging because of the continually changing network geometry that is unlike anything found closer to the ground.
Historically, LEO networks have reduced management complexity by reducing network flexibility and over-allocating capacity. Finding a way to make the complexity manageable, while fully exploiting the network capabilities, allows operators to simultaneously increase capacity and performance.
To address this challenge, CGI has collaborated with TESAT-Spacecom, a renowned leader in telecommunication payloads and laser communications, and has developed the Dynamic Predictive Routing (DPR) network traffic engineering solution.
Built for the next-generations of mega-constellations, DPR demonstrates communications performance improvements by using a new type of networking solution which is proven on space-qualified hardware operating under realistic network conditions.
DPR blends CGI’s comprehensive expertise in Artificial Intelligence (AI) and leverages CGI’s Machine Learning (ML) accelerator platform AccelerateAI360, and a 50 year history of delivering complex and successful projects and solutions within the European and North American Aerospace markets.
The vision at the heart of all mega-constellation systems is a Very Wide Area Network (VWAN) with low latency that can outcompete land-based fiber systems.
One area where mega-constellation systems shine in comparison to Earth-based competition is delivering lower latency and higher throughput communications over long distances with fewer gateways. Optimizing the network routing to reduce the number of gateways used substantially drives down the total cost of system ownership.
Delivering an efficient service over long distances requires a routing concept that manages multi-orbit meshed networks with no prior knowledge of the traffic profile. Using traditional Open Shortest Path First (OSPF) routing to, for
example, send a message from London to New York always privileges the same connection path crossing the Atlantic, thereby creating congestion bottlenecks and neglecting other potentially available links.
The dynamic nature of these links is also quite challenging for traditional routing algorithms to manage. Heuristics such as the Dijkstra or Bellman-Ford algorithms only work with a network as it is now; if conditions change due to delay or fault, then the entire routing table must be recalculated and the forwarding elements updated, adding overhead and delay which accumulate in a negative feedback cycle.
Many operators choose to solve this problem by modifying their chosen routing heuristic to accommodate these factors; mostly, this involves simply increasing the amount of raw compute used, either in terms of the number of factors being considered or the frequency at which updated routing tables are generated.
Neither of these approaches is a cost-effective solution to the problem. Sustaining ever greater constellation numbers only increases the compute demands and cost, so this is certainly not an effective way to compete in the long term in what are increasingly established and highly efficient markets.
CGI’s Dynamic Predictive Routing is a traffic engineering solution based on a centralized routing concept, implementing an SDN (Software Defined Networking) and using a Graph Neural Network (GNN) to model the dynamic properties of the network.
Graph neural network model architectures are highly applicable to network graph formatted data and benefit from the cross-application of a decade of research in convolutional neural networks (CNN) and ordinary, deep neural networks (DNN).
Many optimizations developed for CNNs and DNN can immediately be applied and tested on the more recent GNN architectures.
GNNs are highly attractive for this problem as they can maintain a representation of the network as it changes over time and that resolves the limitation of applying traditional routing heuristics to highly dynamic network topologies.
While the computational cost of training the model is high, the cost is borne once during the training process which can be performed on a specialist system — such as CGI’s AccelerateAI360 platform — away from the mission system.
Delivering service operations using the trained model is low compute and it remains stable for as long as the network is in a stable configuration. By including realistic faults in the model training, CGI can also accommodate fault scenarios without the need for retraining.
CGI’s Dynamic Predictive Routing framework supplies a lightweight and self-contained solution designed to integrate with existing mission systems as cost-effectively as possible.
The model can be optimized to any mission profile and is able to accommodate business needs, with the overall aim of minimizing the service cost per megabyte delivered to the end customer.
Succeeding in this highly competitive future communications marketplace will require innovation to be applied at every level of the business and operational support systems — CGI is here to deliver the core technologies that will keep the forward-thinking innovators prospering in this industry.
www.cgi.com/uk
