Memory hot-zone and its application to accelerate the convergence of KSM for VM consolidation

Abstract

Memory deduplication, which detects and removes redundant memory pages, can efficiently increase the efficiency of memory usage. It is often used with virtualization technologies, such as virtual machines (VMs), to improve their memory utilization. However, current memory deduplication systems, such as KSM (Kernel Samepage Merging), suffer the problem of either slow convergence or high CPU usage. In this paper, we proposed an algorithm, called hotscan, to accelerate the convergence of KSM for VM consolidation. We have observed that the pages have higher chance to be merged among different VMs are clustered in certain memory regions, called memory hot-zone. Based on this observation, hot-scan changes the scanning rule of KSM to check the pages in hot-zones of all VMs first. A novel data structure, called transpose list, is proposed to efficiently carry out the new scanning pattern. Experiments show that our method can accelerate the speed of deduplication 20% to 30% faster than the vanilla KSM without consuming too much additional CPU resource.