CIF:Small:High Performance Memories that Integrate Coding and Computer Architecture

Duke University, Durham NC

Investigators

Abstract

Part 1: Researchers continue to develope new memory and storage technologies for computer systems, including improvements to Flash memory in solid-state drives, as well as novel technologies that have not yet reached the market. These technologies offer new features---including greater storage density, or the ability to retain their data even without power---yet they also introduce new challenges, such as reduced lifetimes or the possibility that writing data into one part of the memory could disrupt values in other parts. Historically, imperfections in memory technologies have been overcome by encoding the information such that disrupted or lost data can be recovered. However, existing techniques do not suffice for the newly emerging memory technologies that have different imperfections and divergent usage models. This project is a collaboration between researchers in information theory (who study how to encode information for different purposes) and researchers in computer architecture (who study how to design computer systems, including memory and storage). The architects will identify exciting technologies and the problems they introduce, and the information theorists will develop new ways of encoding information to overcome these challenges. Whereas much research in coding is theoretical and lacks connection to the reality of computer systems, the collaboration with computer architects---who focus on implementation issues and costs---will ground the coding work and enhance its impact. To foster more interdisciplinary work in this research area, the research team will present half-day tutorials at conferences, including a coding tutorial at an architecture conference, and an architecture tutorial at an information theory conference. The investigators will continue to work with undergraduate research assistants, including students from an established summer outreach program at Duke?s Pratt School of Engineering. The investigators will also continue to recruit female research assistants alongside research assistants from under-represented populations; both investigators have extensive track records in this area and are committed to cultivating diversity in the computing research community. Part 2: The proposed research program seeks to improve the lifetime and fault tolerance of existing and emerging storage technologies---specifically Flash, 3D DRAM, and racetrack memory---by developing practical coding techniques that can be implemented in real-world systems. These storage technologies feature increased storage density, yet suffer the particular challenge of limited lifetime, that might otherwise limit their potential in microarchitectures. Historically, improvements in memory and storage have been due not only to advances in the memory technologies themselves, but also to innovations by computer architects who design memory and storage systems and by coding theorists who devise codes for data storage. The proposed research program is a collaboration between a coding theory group led by PI Calderbank and a computer architecture group led by co-PI Sorin. In their preliminary work, architecture problems have driven development of new theory, and advances in coding theory have driven development of new systems that can take advantage of them. The research thrusts can be classified by the memory technology, and all include specific challenges as well as more open-ended research directions: Specific objectives include 1. Development of new coding solutions for racetrack memory that employ delimiter bits to identify single shift errors and practically implementable codes to correct the error, once identified. Introduction of coding across multiple racetracks (spatial diversity) to correct multiple shift errors. 2. Development of virtual multilevel Flash cells that connect computer architecture with coding theoretic innovations, and evaluation of the tradeoff between host-visible capacity and lifetime. 3. Development of new coding solutions that protect stacked DRAM designs from failures in bits, rows, banks, channels, dies and through silicon vias. Development of fundamental limits on how redundancy trades off against read/write latency, bandwidth requirements, and energy consumption in stacked designs.