CSR: Small: Towards Solutions of Maximizing Performance and Lifetime for Non-Volatile Main Memory Systems

University Of Texas At Dallas, Richardson TX

Investigators

Abstract

Non-volatile memories such as flash memory, Phase Change Memory and Magnetic Memory have the characteristics of low-cost, non-volatility, shock-resistivity and power-economy. Because non-volatile memory consumes less idle power than DRAM by orders of magnitude, achieves fast start-up time and can be twice as dense as DRAM, it is desirable for the computing system industry to use these memories not only as non-volatile storage, but also as main memory. In order to make it possible, however, we need answers to two challenges: The first one is the issue of lifetime and the second is the slowness of write activities. The objective of this project is to develop solutions to these problems by reducing write activities on non-volatile memory through software optimization and hardware support, which lead to the practical adoption of non-volatile main memory in mobile and embedded systems. This project, consisting of 3 groups of tasks, considers various types of platforms with single core or multiple cores. The type of cache on each platform can be either a hardware controlled cache or a software controlled one. Best cache consistency protocol is studied so that the required write activities to main memory are minimized. This project develops techniques including Write-Aware Scheduling, Recomputation, and Data Migration. These techniques will be combined to yield the best results on various platforms. Broader impact: The engineers need to learn how to design the new-generation of embedded systems with non-volatile memory. This project will provide effective tools and solutions to the general engineers.