Award

Wide area network infrastructures, such as the Department of Energy’s (DOE) ESnet, are critical for moving large amounts of data between experimental facilities and data centers, presenting significant challenges related to resource allocation, provisioning, and performance optimization. Testbeds, like the ESnet testbed, are commonly used to evaluate and optimize network performance, but physical testbeds have inherent limitations in scalability, cost, and flexibility. A network digital twin, a virtual replica of a physical network, provides a powerful tool to overcome these limitations. It enables experimentation at scale, exploration of ”what-if” scenarios, and long-term simulations that would be difficult or impossible with physical testbeds alone. Network simulators have been widely used in network research and are well-suited for integration into a network digital twin. However, these tools are typically slow in rapid research and development environments, as even simulating a millisecond of network traffic can take hours when maintaining high fidelity. To address this, we propose developing a digital twin framework that guides users through the entire network simulation workflow. The tool will integrate several network simulators, each focusing on different aspects of network behavior, such as hardware and network protocols, to provide a comprehensive understanding of complex network infrastructures. To accelerate simulation, we will employ a hybrid modeling approach, combining high-fidelity simulation with machine learning models that can “fast-forward” through parts of the simulation. In Phase I, a prototype of the digital twin framework and an interactive visualization dashboard was developed to help users understand simulation results. Additionally, an initial machine learning model was created to predict switch behavior of the ESnet testbed. In Phase II, the framework will be expanded to include a user-friendly application that guides users through the process of configuring models, network topologies, simulation ensembles, and visualizing results. Machine learning models will also be enhanced to further speed up simulations, with the integration of foundation models and federated learning techniques to preserve data privacy. After Phase II, the framework will be expanded beyond research network providers, such as the DOE, to support cloud service providers, which manage applications requiring high throughput and low latency. These applications, such as streaming analytics services, are used for real-time decision making, where increased latency can result in financial consequences for users of cloud services. The proposed framework has the potential to improve the design and operation of these systems by providing insights into network performance optimization, planning, and resource management.