Composable Digital Twins for Science Network Infrastructures using Parallel Discrete Event Simulation

Research wide area network (R-WAN) infrastructures, such as the DOE ESnet, are used to move large amounts of data between experimental facilities and data centers, which imposes challenges on the network infrastructure and requires proactive strategies for resource allocation and provisioning methods. Parallel discrete event simulators (PDES), such as ROSS and ns-3, can be used to perform studies on network performance and evaluate new network configurations. However, these tools are too slow to use by network researchers in a rapid research and development environment, as even a millisecond of network traffic can take hours to be simulated in high fidelity. We propose to develop a digital twin framework built on top of a PDES network simulator that would be a safe and cost-effective tool for use by network researchers as well as network providers and maintainers. We will design the framework to be easy to use without requiring knowledge of PDES. To speed up the network simulator, we will follow a hybrid modeling approach that combines high-fidelity PDES models and coarser-grained machine learning (ML) models to fast-forward through parts of the simulation. Our digital twin framework will be built on top of the ROSS simulator. The framework com- prises 1) a Component Model Library that contains ML models of network components such as routers, and 2) an Orchestrator that will compose PDES models with ML models from the library based on the network configuration provided by the user. For Phase I, we will focus on training models for the components of the ESnet 400G testbed, but the library can be extended as necessary to model other networks. We will also create a visualization dashboard that can be used to analyze the simulated network performance, as well as the performance of the underlying network simula- tor, which will be useful in ensuring the digital twin meets performance requirements. Finally, we will evaluate our digital twin with an application middleware experiment that we can perform on the physical ESnet testbed as well. After Phase II, we plan to expand beyond research network providers such as the DOE to cloud service providers. Cloud service providers support applications, such as streaming analytics services, that require high throughput and low latency. These applications are used to make decisions using real-time information, so increased latency can have financial consequences for the users of cloud services. As an established open-source technical computing software company, Kitware, Inc. is deeply familiar with the needs of commercial, scientific software and how to fulfill those needs within a profitable and sustainable business model. In addition, the team from Kitware, along with our collaborators at Rensselaer Polytechnic Institute and Argonne National Laboratory have extensive collective experience in network simulation and PDES, making us uniquely suited to develop and commercialize the digital twin infrastructure.