Memory performance in a computer system that implements large page mappings is improved by dynamically tuning the page scan rate at which a memory sharing module (e.g., in a hypervisor) performs small page scanning operations that identify and exploit potential small page sharing opportunities within large pages. In operation, when free memory is relatively low, the hypervisor adjusts the page scan rate based on a statistical estimate of the percentage of virtual small pages that are mapped to physical large pages that are shareable. In this fashion the hypervisor dynamically tunes the sharing rate to reflect memory usage of applications. Further, unlike conventional approach to page sharing, the hypervisor proactively breaks large pages before resorting to more expensive memory reclamation techniques, such as ballooning and host swapping.