Revisiting Virtual Memory for the Next Epoch of Memory

Virtual memory is an essential component of the operating system that provides various helpful features used by all programs running on computers.

Over the years, Virtual memory has progressed through two prior epochs with very different requirements, shifting from offering flexibility and a functioning system, to performance in order to support latency critical services.

Now, computing is entering the third epoch of Virtual memory due to the explosion of data, where both performance (latency is critical) and efficient use of memory capacity (DRAM is too expensive) is required. However, unlike the first epoch that also lacked memory, storage devices are now much quicker and provide parallelism.

This project will re-design the HW/SW interface for virtual memory to unlock the parallelism of modern storage devices to improve performance.

Second, the operating system needs to identify and move less frequently accessed data (or cold data) onto cheaper/slower memory.

However, the large page mappings introduced in the second epoch of Virtual memory make it difficult to accurately identify hot and cold data and cannot be used after hot/cold data has been migrated.

This project will enable larger mappings with accurate detection and after moving out cold data to improve performance and save memory.

This work will identify and evaluate the fundamental changes required for Virtual memory to deliver for modern, third epoch applications