Generative AI for Performance Engineering: Tailoring Llama-3 for Bottleneck Classification and Optimization Recommendations

This paper presents a novel approach to software performance analysis by integrating traditional profiling techniques with a fine-tuned large language model (LLM), based on the Llama-3 model.  Addressing the challenges of manual profiling – such as overwhelming data volumes and the high expertise required to interpret performance metrics – the study introduces a lightweight AI-powered profiler trained on structured JSON-based profiling logs and code samples.  The model is fine-tuned using parameter-efficient methods (LoRA and QLoRA) to classify performance bottlenecks (e.g., $N+1$ queries, blocking I/O, recursion) and generate actionable recommendations.  The article provides a comprehensive methodology, from dataset preparation and model training to inference and integration, enabling local deployment even without access to high-end GPUs.  This fusion of performance engineering and generative AI showcases a scalable, accessible solution for developers, enhancing the speed and precision of bottleneck detection and representing a step toward more automated software optimization.