By

Compute intensive applications running on clusters of shared-memory computers are typically implemented using OpenMP and MPI. Applications are difficult to pro-gram, debug and maintain and performance portability is usually limited. Several program transformations have be to be applied at multiple levels of the software and hardware stack to expose parallelism, choose the adequate granularity and finally map it onto the target system in order to obtain high-performing code. This thesis focuses on different aspects of the task parallel data-flow model. We address performance issues at each of the three main levels of the cluster: single core level, single-node level and multi-node level. At the single core level, we effectively target a hardware layer, the SIMD-vector units available on modern processors, and attempt to maximize the vectorizability opportunities of the program by aggressively transforming it using a polyhedral compilation framework. At the medium level, on a single node, we address the issues of barrier removal and load balancing by introducing a combined compiler and run-time approach tha