Architectural performance models, such as the Palladio Component Model, can support early design decisions for microservice systems by enabling performance simulation. However, early stage models often lack the service resource demand specifications (e.g., 100 ms CPU demand) required for such performance simulations. Existing approaches to build performance models often depend on late-stage running prototypes and intrusive profiling, or otherwise, on manual expert estimation, which is costly and hard to reproduce. We present GARMA, an LLM-based workflow that processes early design artifacts and automatically generates behavioral microservice models with bounded best-case and worst-case resource demand estimates. In a microservice test scenario (TeaStore), GARMA generated 150 behavioral models that closely matched referencecstructures (average Jaccard similarity 0.97; perfect matches at 84%). The predicted CPU resource demand intervals aligned well with measurements, capturing most (85%) user-facing interactions within the predicted intervals. In addition, GARMA produced consistently narrow CPU-demand intervals for the main user-facing steps. Additionally, we compared GARMA against a naive LLM baseline that estimates CPU demands without performing the presented LLM workflow. GARMA consistently outperformed that baseline, achieving lower absolute errors, higher coverage, and narrower intervals.
Hummel et al. (2026) studied this question.