Supervised by

A multiprocessor system with uniform memory access is difficult to scale due to the increasing contention on the memory bus and the complexity of the connections between CPUs and memory modules. Non-uniform memory access (NUMA) offers a solution for these problems by introducing so called nodes. Each node consists of a set of processors and local memory. Even though every processor can access memory on all nodes, the access delay differs for local and remote memory. Applications spanning over multiple cores might suffer from a high access delay for ill-placed memory. In this thesis, we introduce an extension for the Barrelfish operating system capable of detecting local and remote memory access and migrating memory between nodes. To achieve automatic NUMA optimization, we implement a migration policy which analyzes the access pattern and tries to find an optimal memory placement. Furthermore, we present a benchmark for measuring the impact of NUMA optimization on memory access performance. Acknowledgements First and foremost, I would like to thank Prof. Timothy Roscoe and Kornilios Kourtis for providing me with the opportunity to explore Barrelfish and contribute to it. This thesi