Principles of Operation for Shingled Disk Devices

Abstract

A leading strategy for driving the areal density of magnetic disk drives through 1 – 10 terabit/inch 2 (the coming decade) is to shingle (partially overlap) adjacent tracks, imposing significant restrictions on where data can be written without incurring multi-track read-modify-write penalties. These restrictions and penalties can be 1) fully hidden from system software using techniques familiar in NAND Flash disks; 2) minimally exposed as multi-track, shingled bands of predetermined size that can be read normally, but only appended to or trimmed (erased); or 3) maximally exposed as dynamically sized bands of shingles separated by guard regions of previously erased tracks, allowing maximal capacity to be obtained by the most sophisticated system software. While the latter options require significant changes in system software, there is a rich history of demonstrations of log-structured file systems that should be able to do this, and a profusion of write-once cloud storage systems that could provide the economic “killer application ” [Kleiman11]. Now is a very good time for systems software experts to take interest and weigh in as magnetic disk technologists are experimenting and prototyping shingled disks. Experience shows that changes in the interface model for magnetic disks can take decades to change (for example, 512B to 4096B sectors) unless device vendors and systems software developers work together toward mutually desired principles of operation.