Basically, the cloud hypervisor only passthrough the NVM - Non-Volatile Memory a.k.a. PMEM (e.g. AEP) to the virtual machines, then the cloud users are required to decide how to operate it. In order to deploy NVM, cloud users have to develop professional skills by refactoring their codes to explicitly save hot and cold data in DRAM and NVM and managing NVM as a NUMA node based on some special features from guest OS. Since refactoring code to deploy NVM is usually unacceptable for cloud customers and managing NUMA node based on guest OS is not efficient, there is critical disadvantage on the current NVM memory management for virtual machines.

Propose a hypervisor-based scheme to manage DRAM/NVM as tiered memory,

which can be seamlessly used by virtual machines.

1) An efficient solution to dynamically identify hot and cold memory pages for guest

workloads.

2) A method to support page migration between hot and cold memory pages in case

of passthrough devices.

3) A new memory controller prototype to accurately obtain information about hot

and cold pages.

• Tiered memory management in host OS
• Workload model of guest OS
• Scan mechanism for hot and cold memory page
• Page migration during passthrough devices' DMA operation
• Acceleration engine for page migration