False Sharing Memory Bug Detection with LLM Static Code Analysis

Training LLM to detect and correct false sharing memory bugs in C code benchmarks

Navigation

• prompts/: contains the iterations of the prompt. Each prompt is stored in a separate file.
• src/: contains the prompt engineering notebook where working prompts are tested and the ts_fs_benchmarks notebook where all 12 benchmarks are with the current prompt.
• test_suites/: contains the true and false sharing benchmarks currently being used to access the model's performance.
  • shell_scripts/: scripts written to automate the runtime and memory overhead evaluation of LLM-corrected code and baseline benchmark code.

Technical Details:

• All benchmarks are written in C or C++.
• Groq is used to access the open source Meta LLama 3.1 Instruct 70B model for testing.

Resources:

• False sharing
• Meta Llama Cookbook
• Guiding paper
• Benchmarks source