PARALLEL DATA LAB 

PDL Abstract

Metadata Optimization for Shingled Disks

Information Networking Institute Master of Science Thesis. CMU-PDL-13-108, May, 2013.

Pavan K Alampalli

Information Networking Institute
Carnegie Mellon University

http://www.pdl.cmu.edu/

Continued growth in demand for storage capacity is driving the disk drive industry to increase areal density. Shingled Magnetic Recording (SMR) increases areal density by up to a factor of 2.5x by partially overlapping previously written tracks. To avoid multi-track read-modify-write penalties known as write amplification, data should only be appended and no random write should be allowed. To deal with this, previous work implemented a SMR-aware File System called ShingledFS. This work builds on the ShingledFS with the aim of optimizing metadata storage. An SMR disk contains two partitions, shingled and unshingled. A shingled partition has tracks shingled on one another and thus there can be write-amplification. An unshingled partition has tracks laid out with a track gap between adjacent tracks, just as in a traditional disk. The unshingled partition trades off the data density of SMR disk in order to retain a random-write. The unshingled partition can be used to store constantly changing metadata in the filesystem. The larger the need for unshingled capacity, the lower is the overall increase in data density achieved by SMR disks over traditional disks. This research explores the use of an embedded key-value store to optimize filesystem metadata storage. We pack frequently changing metadata, such as filesystem inodes, directories, file attributes and small files into a few sequentially written large files, thereby reducing the need for unshingled space. We report on the ways in which metadata can be packed into LevelDB and the resultant savings of unshingled disk space.

FULL TR: pdf