Predicting memory page stability and its application to memory deduplication and live migration

摘要

There are various applications and operations in virtualized environments that rely on memory page stability to achieve satisfactory performance. These applications include VM live migration and memory deduplication. Unfortunately, there is a large gap between existing prediction mechanisms and actual behavior. This is the gap we hope to narrow. We make the following contributions: (1) We characterize the behavior of memory pages based on the page flags available through the linux kernel's proc file system. (2) We propose a prediction framework that can be used to predict memory pages that are highly likely to be relatively stable. (3) We study two applications that may benefit from the memory characterization and the proposed prediction framework: memory deduplication and VM live migration.