CSR: Small: High-Performance and Energy-Efficient Single-Level Stores: Efficient Coordinated Management of Storage and Memory

Carnegie Mellon University, Pittsburgh PA

Investigators

Abstract

Modern computing systems require manipulation of persistent data. However, the mechanisms to access persistent data are separated from the mechanisms to manipulate such data in a traditional two-level storage model: persistent data is stored in non-volatile storage, such as hard disks or solid-state disks, whereas it is manipulated in volatile memory, such as DRAM. Unfortunately, this decoupled model of system design leads to large amounts of inefficiency in terms of both performance and energy. This research undertakes a rethinking of the memory and storage systems in the presence of new and emerging non-volatile memory technologies such as phase-change memory , spin-transfer torque RAM, and memristors. These technologies promise large capacities at low latency, making them prime candidates for the design of a single-level coordinated memory and storage system where the manipulation and storage of all data are unified. Yet, there are significant application-level and system-level efficiency and correctness challenges in devising such a single-level store with emerging technologies. The central goal of this project is to develop hardware/software cooperative techniques to enable high-performance and energy-efficient single-level stores in computer systems targeting a wide range of applications in cloud computing, client systems, and mobile systems. It is expected that research performed in this project will enable new applications and much more robust and efficient data storage systems, making our daily lives that increasingly depend on data processing better and more productive. Enabling unified memory and storage for modern computing systems can largely reduce wasted work and take advantage of emerging low-power technologies, thereby taking a large step in making computing more energy efficient.