A fresh look at memory hierarchy management

Abstract

The memory systems of modern computers are much more complicated than the ones for which the operating system memory management algorithms in current widespread use were developed. Additionally, the way in which computer systems are used has changed substantially. This paper argues that these changes are sufficient to require reevaluating some of the fundamental assumptions made in operating system memory management. One obvious problem motivating this reevaluation is that current systems do a terrible job of maintaining performance for high-priority processes when the system comes under memory pressure due to the behavior of low-priority processes. This paper discusses potential operating system modifications and proposes a new metric, the system memory cycles per instruction (system MCPI) of a process, which can be used to evaluate these modifications.