FaultLLAMA2: A Fine Tuned Large Language Model for Code Recommendations to Reduce Java Lock Contention

Lock contention occurs when multiple threads or processes try to access the same lock simultaneously, often causing performance degradation in software that runs concurrent threads. In this paper, we propose a new recommendation model that leverages Large Language Models (LLMs) and suggests refactored source code to alleviate lock contention faults. Our proposed FaultLLAMA2 model is based on a fine-tuned customization of the generalized LLAMA2 LLM model. FaultLLAMA2 is fine-tuned on examples of lock contention anti-patterns and their associated recommendations. Our results show that the accuracy of the FaultLLAMA2 model was on average 2.7 times higher than the LLAMA2 baseline model based on the BLED and Pass@K metrics. We also compared the FaultLLAMA2 model recommendations with a CodeBERT code generation model fine-tuned for lock contention recommendations and its accuracy was 4 times better.