Design and evaluation of vertical scalability strategies for deploying large-scale co-simulation infrastructures

This thesis investigates the design and evaluation of vertical scaling techniques to enhance simulation performance within large-scale co-simulation environments for Multi-Energy Systems (MES). MES are systems where buildings, power grids, gas infrastructure, and heating systems interact optimally at various scales. Compared to traditional isolated energy systems, these systems represent a significant opportunity to enhance technical, economic, and environmental performance. As MES complexity increases, standalone simulations often fail to capture their interdisciplinary nature. In response, co-simulation frameworks, which integrate multiple simulators into a single coordinated environment, are used to model these systems holistically. Within a co-simulation framework, the orchestrator manages the interaction between simulators, each representing different model instances within a defined scenario. A significant bottleneck in this context arises from computationally heavy simulators that are often neither stable nor efficient, limiting scalability and computational efficiency. To address this, the thesis explores various vertical scalability techniques designed to optimize single-node performance, specifically improving scalability for individual simulators within co-simulation frameworks. The methods evaluated include standard multithreading, multithreading in a GIL-free environment, multiprocessing, and the Ray framework. This study focused on Mosaik, a flexible co-simulation framework that enables the integration of multiple simulation tools within a single orchestrated environment, and COESI, a co-simulation platform that implements Mosaik’s orchestrator. COESI offers a rich library of Mosaik-compatible simulators, integrating Functional Mock-up Units (FMUs) and various models that provide a solid foundation for realistic MES simulations. Two case studies were conducted. The first, a simplified traffic simulation, enabled an exploration of Mosaik’s core functionalities without COESI’s additional routines. The second case involved a complex MES scenario using the COESI platform, allowing a comprehensive examination of vertical scaling techniques under different loads. Tests were designed to scale the number of model entities managed by different simulators, allowing for controlled simulation complexity increases and systematic performance observation under varying loads. Results indicated that multithreading yielded some performance gains in I/O-bound tasks, but these improvements were limited in the broader simulation context. Testing in GIL-free environments revealed compatibility issues, as many Python libraries inherently depend on GIL-based designs, making GIL-free multithreading an inadequate solution for achieving significant performance improvements in this context. In CPU-intensive tasks, multiprocessing and Ray’s distributed architecture achieved more substantial performance gains, particularly with stateless simulators. Through these tailored techniques, significant reductions in execution time were achieved, confirming the value of adaptive vertical scaling of simulators for MES co-simulation. This work highlights the importance of selective parallelization in optimizing computational efficiency within COESI, providing insights and best practices for scaling MES simulations to meet the growing demands of complex energy system scenarios.