Multi-core platforms with non-uniform memory access (NUMA) design are now a common resource in High Performance Computing. In such platforms, the shared memory is organized in an hierarchical memory subsystem in which the main memory is physically distributed into several memory banks. Additionally, the hierarchical memory subsystem of these platforms feature several levels of cache memories. Because of such hierarchy, memory access costs may vary depending on the distance between tasks and data. Furthermore, since the number of cores is considerably high in such machines, concurrent accesses to the same distributed shared memory are performed. These accesses produce more stress on the memory banks, generating load-balancing issues, memory contention and remote accesses. Therefore, the main challenge on a NUMA platform is to reduce memory access latency and memory contention. In this context, the main objective of this thesis is to attain scalable performances on multi-core NUMA machines by controlling memory affinity. The first goal of this thesis is to investigate which characteristics of the NUMA platform and the application have an important impact on the memory affinity control and propose mechanisms to deal with them on multi-core machines with NUMA design. We focus on High Performance Scientific Numerical workloads with regular and irregular memory access characteristics. The study of memory affinity aims at the proposal of an environment to manage memory affinity on Multi-core Platforms with NUMA design. This environment provides fine grained mechanisms to manage data placement for an application by using compilation time and architecture information. The second goal is to provide solutions that show performance portability. By performance portability, we mean solutions that are capable of providing similar performances improvements on different NUMA platforms. In order to do so, we propose mechanisms that are independent of machine architecture and compiler. The portability of the proposed environment is evaluated through the performance analysis of several benchmarks and applications over different platforms. Last, the third goal of this thesis is to design memory affinity mechanisms that can be easily adapted and used in different parallel systems. Our approach takes into account the different data structures used in High Performance Scientific Numerical workloads, in order to propose solutions that can be used in different contexts. We evaluate the adaptability of such mechanisms in two parallel programming systems. All the ideas developed in this research work are implemented in a Framework named Minas (Memory affInity maNAgement Software). Several OpenMP benchmarks and two real world applications from geophysics are used to evaluate its performance. Additionally, Minas integration on Charm++ (Parallel Programming System) and OpenSkel (Skeleton Pattern System for Software Transactional Memory) is also evaluated.
