A novel approach to loop parallelization

Parallelization environments still have limited usability for the large number of scientists and industrial users wishing to use high performance computing. This is due to software development being geared towards the expert user with knowledge of parallelization and the small percentage of real code that can be analysed because of limitations with symbolic data dependence analysis. There is a need for software that will convert sequential programs to parallel code without concerning the user with parallel programming issues in the main. During parallelization, the sequential program graph is restructured by the application of a number of transformations that remove and reduce loop carried dependences. Selecting the optimum sequence of transformations is a problem that has not been solved satisfactorily by traditional programming techniques. The paper outlines an investigation into the usefulness of neural networks in offering transformation selection guidance. The key issues of code characterization, knowledge acquisition and two neural network based parallelization tools are addressed.