AutoPCM: Automated Generation of Architectural Performance Models from Early Design Artifacts using LLM Agents

Architectural performance models, such as the Palladio Component Model (PCM), allow architects to evaluate design alternatives for software systems through performance simulation. The dominant barrier to such early-design simulation, however, is the construction of a valid performance model in the first place: building a complete PCM instance requires considerable performance engineering expertise and manual effort, and a metamodel-valid instance is a necessary precondition for any subsequent simulation.

Existing approaches to (semi-)automatically construct such models either require a running system or its source code, depend on mature annotated UML models, or produce only partial models that still need manual completion. We present AutoPCM, an LLM-based multi-agent workflow that processes heterogeneous early-design artifacts and automatically constructs complete, metamodel-valid PCM instances.

The conceptual contribution is a role-based agentic decomposition that generalizes to other component-based performance formalisms (e.g., Layered Queueing Networks); in this paper, we instantiate and evaluate it for PCM using GPT-5.4. In an evaluation on two subject systems, Corona-Warn-App (CWA) (11 use cases) and TeaStore (9 use cases), 205/220 generated PCM instances pass PCM metamodel validation. ... mehrThe generated models achieve a mean structural similarity of 0.86 (Jaccard) compared to manually authored reference models.