Access Type

Open Access Dissertation

Date of Award

January 2014

Degree Type


Degree Name



Computer Science

First Advisor

Lihao Xu


Flash memory promises to revolutionize storage systems because of its massive performance gains, ruggedness, large decrease in power usage and physical space requirements, but it is not a direct replacement for magnetic hard disks. Flash memory possesses fundamentally different characteristics and in order to fully utilize the positive aspects of flash memory, we must engineer around its unique limitations. The primary limitations are lack of in-place updates, the asymmetry between the sizes of the write and erase operations, and the limited endurance of flash memory cells. This leads to the need for efficient methods for block cleaning, combating write amplification and performing wear leveling. These are fundamental attributes of flash memory and will always need to be understood and efficiently managed to produce an efficient and high performance storage system.

Our goal in this work is to provide analysis and algorithms for efficiently managing data storage for endurance in flash memory. We present update codes, a class of floating codes, which encodes data updates as flash memory cell increments that results in reduced block erases and longer lifespan of flash memory, and provides a new algorithm for constructing optimal floating codes. We also analyze the theoretically possible limits of write amplification reduction and minimization by using offline workloads. We give an estimation of the minimal write amplification by a workload decomposition algorithm and find that write amplification can be pushed to zero with relatively low over-provisioning. Additionally, we give simple, efficient and practical algorithms that are effective in reducing write amplification and performing wear leveling. Finally, we present a quantitative model of wear levels in flash memory by constructing a difference equation that gives erase counts of a block with workload, wear leveling strategy and SSD configuration as parameters.