Comparing Performance of Different Cleaning Algorithms for SMR Disks
M.S. Thesis: Master of Science in Information Networking, April 2014..
Mukul Kumar SinghInformation Networking Institute
Carnegie Mellon University
Shingled Magnetic Recording (SMR) promises to sustain current growth in disk drive capacities with minimal change in the current disk drive technology. Shingling implies overlapping of tracks in a hard drive. Shingling would cause overwrites on down-track sectors with each sector write, hence new interfaces are being proposed to allow host software to exploit SMR with minimal change. An obvious interface is a Shingled Translation Layer which is akin to a Flash Translation Layer. Here the disk can completely hide the layer of remapping and background cleaning, but this comes at the cost of complexity in the disk processor and hard-to-predict performance changes. Other interfaces which enable the host application to handle shingling have been proposed as well. In a strict append model, the disk is divided into fixed sized bands and data is written to a particular band in a strict append order, with cleaning done by resetting the write cursor to the beginning of a band. Another promising interface, Caveat Scriptor, gives the host an address space of all possible sectors. In- order to handle shingling, this interface exposes two drive parameters to determine which sectors may or will not be damaged because of a certain write. These parame- ters are Drive No Overlap Range (DNOR) and Drive Isolation Distance (DID). This paper will explain these parameters, explain the design of a filesystem designed for this extreme interface, caveat scriptor, and compare the cleaning performance of a lesystem designed for the Caveat Scriptor interface to one designed for the Strict Append interface.
FULL PAPER: pdf