An empirical study of memory sharing in virtual machines

Content-based page sharing is a technique often used in virtualized environments to reduce server memory re-quirements. Many systems have been proposed to cap-ture the benefits of page sharing. However, there have been few analyses of page sharing in general, both con-sidering its real-world utility and typical sources of shar-ing potential. We provide insight into this issue through an exploration and analysis of memory traces captured from real user machines and controlled virtual machines. First, we observe that absolute sharing levels (exclud-ing zero pages) generally remain under 15%, contrast-ing with prior work that has often reported savings of 30 % or more. Second, we find that sharing within in-dividual machines often accounts for nearly all (>90%) of the sharing potential within a set of machines, with inter-machine sharing contributing only a small amount. Moreover, even small differences between machines sig-nificantly reduce what little inter-machine sharing might otherwise be possible. Third, we find that OS features like address space layout randomization can further di-minish sharing potential. These findings both temper ex-pectations of real-world sharing gains and suggest that sharing efforts may be equally effective if employed within the operating system of a single machine, rather than exclusively targeting groups of virtual machines.