A MACHINE LEARNING BSP-BASED APPROACH FOR AUTOMATIC PARALLEL ALGORITHM DESIGN

Dr. Mounir Kechid*

ABSTRACT

Current automatic parallel algorithm design tools often struggle with limitations like confined functionality, limited portability, and performance bias for specific architectures. This paper proposes a novel approach that leverages the Bulk Synchronous Parallel (BSP) model and machine learning to overcome these hurdles.We explore the feasibility of creating a comprehensive dataset of parallel algorithms using the BSP model to train a machine learning model. This model can then automatically generate efficient and portable parallel algorithms tailored to specific computational problems. Our multi-stage approach extracts parallelism and dependencies from code structure, enriches this information with target architecturefeatures, and feeds it to a GNN for parallel algorithm generation. Although still in theexploratory stage, this research has promising implications for expanding access toautomatic parallel algorithm design.

Keywords:

Parallel algorithm design, Bulk Synchronous Parallel (BSP) model,
machine learning, Big Data, Code performance optimization, Automatic parallelization.

Full Text Article