Openmp Extensions To Support Dependent Work Distributions

Abstract

OpenMP is an Application Program Interface (API), which was first introduced in 1996 and since then it was adopted as the most popular standard for shared-memory parallel programing.It provides a portable and scalable model for the developers of multi-threaded applications. OpenMP is still in the process of being extended, which means that in order to support each new feature changes should be done both in the front and back ends. These changes have to do with the new syntax that appears as well as the code generation that targets the runtime system. For this reason and for the needs of this thesis we use the Mercurium compiler and the Nanos runtime system which are designed to accommodate new concepts and extensions that we need to add in OpenMP. In this thesis we propose extensions to allow the runtime detection of dependencies between for loops. With these extensions we are able to reduce the delays that are potentially caused from unbalanced workloads in the for loops and improve performance.