OptiAdvisor Tool For OpenMP Performance Optimization Data
Hesham Youssef · 2024
Parallel programming performance optimization is often a difficult and time-consuming process that requires careful analysis of the subject code and requires a large amount of domain expertise and time for trial and error. Compiler optimizations tend to be more conservative as specific application contexts are missing for particular optimization needs. OptiAdvisor aims to address the problem of context-based optimization within the particular domain of parallel programming due to the lack of work published in this area. OptiAdvisor will attempt to leverage the computation power of high-performance computing (HPC) and large-language models (LLM) to autonomously offer parallel coding optimization with high degrees of accuracy. This work aims to contribute 4 key points; 1. Dataset scaling method to produce high-quality and accurate data for testing and evaluation. 2. Develop an evaluation pipeline to evaluate the model responses using different HPC evaluation metrics. 3. Develop a retrieval pipeline for offering additional context to the LLM at inference time to produce high-quality responses. 4. Formats the relevant documents and the original query to a prompt which is then used to obtain the LLM response.