Hierarchical application-oriented physical memory management

grantor: University of TorontoPage faults are becoming increasingly expensive in today's computer systems. Different applications present different access patterns, and therefore different page replacement policies are appropriate in each case. Thus, the operating system should support different policies for different parts of the system at the same time. This dissertation describes the design, implementation, and analysis of the physical memory management subsystem that we have developed for Tornado, a shared memory multiprocessor operating system. We developed a framework that defines a hierarchical distribution of memory management, in which each level in the hierarchy is responsible for certain portions of the system's memory. We developed mechanisms that allow policy enforcement at any of these levels, while at the same time providing for efficient communication between components of the hierarchy to allow memory to flow where it is needed most in the system. By managing the memory of different parts of the system independently, we allow applications to customize their memory management strategies to best suit their needs and access patterns.M.A.Sc