Quoc Minh Le

Date of Award

2-22-2019

Document Type

Thesis

Publisher

Santa Clara : Santa Clara University, 2019.

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Engineering

First Advisor

Ahmed Amer

Abstract

Disk drives have seen a dramatic increase in storage density over the last five decades, but to continue the growth seems difficult if not impossible because of physical limitations. One way to increase storage density is using a shingled magnetic recording (SMR) disk. Shingled writing is a promising technique that trades off the inability to update in-place for narrower tracks and thus a much higher data density. It is particularly appealing as it can be adopted while utilizing essentially the same physical recording mechanisms currently in use. Because of its manner of writing, an SMR disk would be unable to update a written track without overwriting neighboring tracks, potentially requiring the rewrite of all the tracks to the end of a "band" where the end of a band is an area left unwritten to allow for a non-overlapped final track. Random reads are still possible on such devices, but the handling of writes becomes particularly critical.

In this manuscript, we first look at a variety of potential workloads, drawn from real-world traces, and evaluate their impact on SMR disk models. Later, we evaluate the behavior of SMR disks when used in an array configuration or when faced with heavily interleaved workloads. Specifically, we demonstrate the dramatically different effects that different workloads can have upon the opposing approaches of remapping and restoring blocks, and how write-heavy workloads can (under the right conditions, and contrary to intuition) result in a performance advantage for an SMR disk.

Download

Share

COinS