Improving Memory Management with Hardware-Generated Memory Access Profiles

With the rise of new memory technologies such as Flash we expect future systems to rely on a flatter memory hierarchy using multiple heterogeneous memory units in parallel. Operating systems face new challenges in managing these memories in a costand energy-efficient way while maximizing performance. We believe that the key to meet these challenges is to leverage the full potential of each memory technology by migrating pages between the different memories according to their current usage. In this thesis we introduce a new memory profiling architecture that provides the operating system with hardware generated sets of promising candidate pages. The candidates are selected according to a customizable strategy (e.g., least frequently used) and can be employed by the operating system to make a migration decision.