Carnegie Mellon University, Dept. ECE Ph.D. Dissertation CMU-PDL-04-104. May 2004.
Steven W. Schlosser
Electrical and Computer Engineering
Carnegie Mellon University
Pittsburgh, PA 15213
http://www.pdl.cmu.edu
MEMS-based storage is an interesting new technology that promises to bring fast, non-volatile, mass data storage to computer systems. MEMS-based storage devices (MEMStores) themselves consist of several thousand read/write tips, analogous to the read/write heads of a disk drive, which read and write data in a recording medium. This medium is coated on a moving rectangular surface that is positioned by a set of MEMS actuators. Access times are expected to be less than a millisecond with energy consumption 10-100X less than a low-power disk drive, while streaming bandwidth and volumetric density are expected to be around that of disk drives.
This dissertation explores the use of MEMStores in computer systems, with a focus on whether systems can use existing abstractions and interfaces to incorporate MEMStores effectively, or if they will have to change the way they access storage to benefit from MEMStores. If systems can use MEMStores in the same way that they use disk drives, it will be more likely that MEMStores will be adopted when they do become available.
Since real MEMStores do not yet exist, I present a detailed software model that allows their use to be explored under a variety of workloads. To answer the question of whether a new type of device requires changes to systems, I present a methodology that includes two objective tests for determining whether the benefit from a device is due to a specific difference in how that device accesses data or is just due to the fact that that device is faster, smaller, or uses less energy than current devices. I present a range of potential uses of MEMStores in computer systems, examining each under a number of user workloads, using the two objective tests to evaluate their efficacy.
Using the evidence presented and the two objective tests, I show that systems can incorporate MEMStores easily and employ the same standard abstractions and interfaces used with disk systems. At a high level, the intuition is that MEMStores are mechanical storage devices, just like disk drives, only faster, smaller, and requiring less energy to operate. Accessing data requires an initial seek time that is distance-dependent, and, once access has begun, sequential access is the most efficient. This intuition is described in more detail, and the result is shown to hold for the range of uses presented.
KEYWORDS: MEMS-based storage, storage systems performance, database systems, disk arrays
FULLTHESIS: pdf