Data-intensive applications tend to use a large quantity of dynamic random-access memory (DRAM) using DRAM caching. This is relatively expensive, but using insufficient DRAM can lead to paging some data to secondary storage. That may result in unacceptable levels of performance. The authors suggest the use of less expensive, but still relatively fast storage as a substitute for part of DRAM. They suggest the use of non-volatile memory (NVM) devices, such as magnetic RAM (MRAM), as it has lower latency than flash-based storage devices.
Efficient paging to fast storage requires the use of techniques such as memory-mapped input/output (MMIO), which in Unix-like operating systems is a special system call named mmap. The paper proposes a new concept for page reclamation that improves the efficiency of mmap. To demonstrate, the authors modify the Linux kernel to introduce a new page recycling technique and introduce other new approaches, such as extended vector I/O to reduce write operation overhead.
Experimental results compare a full DRAM configuration versus an insufficient DRAM configuration using the new page recycling technique and other changes. This benchmark test shows that the modified approach achieves 92 percent of the performance of the full DRAM system. All in all, the paper seems to make a very useful contribution for efficient memory management where cost versus performance may play a key role.