High Performance with Prescriptive Optimization and Debugging

Parallel programming is the dominant approach to achieve high performance in computing today. Correctly writing efficient and fast parallel programs is a big challenge mostly carried out by experts. We investigate optimization and debugging of parallel programs. We argue that automatic parallelization and automatic vectorization is attractive as it transparently optimizes programs.

The thesis contributes an improved dependence analysis for explicitly parallel programs. These improvements lead to more loops being vectorized, on average we achieve a speedup of 1.46 over the existing dependence analysis and vectorizer in GCC. Automatic optimizations often fail for theoretical and practical reasons.

When they fail we argue that a hybrid approach can be effective. Using compiler feedback, we propose to use the programmer’s intuition and insight to achieve high performance. Compiler feedback enlightens the programmer why a given optimization was not applied, and suggest how to change the source code to make it more amenable to optimizations.

We show how this can yield significant speedups and achieve 2.4 faster execution on a real industrial use case. To aid in parallel debugging we propose the prescriptive debugging model, which is a user-guided model that allows the programmer to use his intuition to diagnose bugs in parallel programs.

The model is scalable, yet capable enough, to be general-purpose. In our evaluation we demonstrate low run time overhead and logarithmic scalability. This enable the model to be used on extremely large parallel systems.