ParEVO: Synthesizing Code for Irregular Data: High-Performance Parallelism through Agentic Evolution

A research team led by Quanquan C. Liu has introduced ParEVO, a novel framework that addresses one of the most challenging problems in high-performance computing: automatically generating correct and efficient parallel code for irregular data structures. Traditional large language models (LLMs) like GPT-4 and Claude frequently fail on these tasks, producing code with subtle race conditions, deadlocks, and poor scaling when dealing with sparse graphs, unbalanced trees, and non-uniform meshes. ParEVO bridges this gap through a multi-component approach that includes the Parlay-Instruct Corpus—a curated dataset of 13,820 tasks—and specialized models fine-tuned to align with the rigorous semantics of the ParlayLib parallel programming library.

The system's core innovation is its Evolutionary Coding Agent (ECA), which performs 'last mile' correctness improvements by iteratively repairing generated code using feedback from compilers, dynamic race detectors, and performance profilers. This agentic evolution process enables ParEVO to achieve remarkable results: an average 106x speedup across the ParEval benchmark suite, with a maximum speedup of 1103x on specific tasks. On complex irregular graph problems, it maintains a robust 13.6x speedup, significantly outperforming state-of-the-art commercial models. Perhaps most impressively, ParEVO matches expert human baselines, achieving up to 4.1x speedup on highly-irregular kernels where traditional automated approaches typically fail. The framework represents a significant advance toward automating the parallelization of algorithms that have resisted previous AI-assisted coding solutions.