Algorithms and Data Representations for Emerging Non-Volatile Memories

The evolution of data storage technologies has been extraordinary. Hard disk drives that fit in current personal computers have the capacity that requires tons of transistors to achieve in 1970s. Today, we are at the beginning of the era of non-volatile memory (NVM). NVMs provide excellent performance such as random access, high I/O speed, low power consumption, and so on. The storage density of NVMs keeps increasing following Moore’s law. However, higher storage density also brings significant data reliability issues. When chip geometries scale down, memory cells (e.g. transistors) are aligned much closer to each other, and noise in the devices will become no longer negligible. Consequently, data will be more prone to errors and devices will have much shorter longevity. This dissertation focuses on mitigating the reliability and the endurance issues for two major NVMs, namely, NAND flash memory and phase-change memory (PCM). Our main research tools include a set of coding techniques for the communication channels implied by flash memory and PCM. To approach the problems, at bit level we design error correcting codes tailored for the asymmetric errors in flash and PCM, we propose joint coding scheme for endurance and reliability, error scrubbing methods for controlling storage channel quality, and study codes that are inherently resisting to typical errors in flash and PCM; at higher levels, we are interested in analyzing the structures and the meanings of the stored data, and propose methods that pass such metadata to help further improve the coding performance at bit level. The highlights of this dissertation include the first set of write-once memory code constructions which correct a significant number of errors, a practical framework which corrects errors utilizing the redundancies in texts, the first report of the performance of polar codes for flash memories, and the emulation of rank modulation codes in NAND flash chips