Defensa Tesis Licenciatura Luis Mastrangelo

• Título: "Dynamic Analysis and Transformation of Programs for Automatic and Speculative Parallelization: Contributions and Perspectives"
• Directores: Alexandra Jamborean (Université de Strasbourg, Francia), Sergio Yovine
• Jurados: Hernán Melgratti, Vincent Loechner (Université de Strasbourg, Francia)
• Resumen:

With the advent of multi-core processors, automatically parallelizing sequential code became increasingly important. Particularly, it is a challenging task to parallelize code at runtime, if the information available at compile time is not sufficient. Runtime parallelization techniques are usually based on thread level speculation (TLS), where straightforward parallelization transformations are optimistically applied on the original sequential code. In most TLS proposals, modest performance gains were obtained, since the unique parallelization strategy being applied generally yields many rollbacks. To be more efficient, TLS systems must handle more complex code transformations that can be profitably selected at runtime, depending on the current execution context. To our knowledge, no attempt on applying various parallelizing transformations using the TLS systems have been reported yet.

This thesis contributes to the development of a TLS framework whose main feature is to be able to speculatively parallelize a sequential loop nest in various ways, by re-scheduling its iterations. The transformation to be applied is selected at runtime with the goal of minimizing the number of rollbacks and maximizing performance. Experimental validation is performed through significant case studies. Our experiments yielded promising outcomes on the applications we proposed.