Improving the memory affinity of parallel applications via optimized thread and data mappings is the key to
improve their performance and energy consumption on sharedmemory architectures. In this paper, we
proposed kMAF, a kernel-based framework to improve affinity. kMAF uses page faults of parallel applications
to determine their memory access behavior and to optimize the mapping online. In an evaluation with two
parallel benchmark suites on three different NUMA systems, kMAF showed substantial performance and
energy efficiency improvements with a low overhead, with gains close to an Oracle mechanism. For the future,
we will evaluate the impact of the hierarchical structure of modern NUMA systems on the data mapping.