Design and Implementation of a

Configurable Mixed-Media Filesystem

Extended Abstract

Silvano Maffeis?

Computer Science Department, University of Zurich, Switzerland

[email protected]

IFI TR 94:03

June 1993

Abstract
Traditional UNIX1 filesystems are not adequate enough for directly supporting database systems, multimedia systems or applications requiring high I/O performance. In this paper we describe the design and implementation of a configurable mixed-media filesystem. The attribute configurable means that a filesystem serving a specific application area can be realized with less effort out of a library of reusable filesystem classes. The attribute mixed-media stands for the filesystem's ability to integrate different media types (RAM, harddisks, WORM optical disks, CDROMs, tape devices, RAIDs etc.) into a virtual storage and making applications unaware of this aggregation. A prototype C++ implementation of the proposed design, called the Vanilla filesystem, is presented and its performance assessed. Raw-write performance is up to 4:5 times higher than in a standard Sun OS filesystem. We will also demonstrate how various storage organization forms, especially hierarchies, arrays, and mirrors of both local and remote storages, can be realized using an expressive syntax.

1 Problem Statement

Traditional UNIX filesystems [8] were designed to support computers with little RAM and small disk devices. Files are split into fixed size logical blocks and sequentially accessing a whole file induces several movements of the read/write head often causing long delays. This scheme is inefficient, since, as the well-known study conducted by Ousterhout and others [10] reports, more than two thirds of the file accesses in an academic UNIX environment are whole-file transfers. Better performance can be achieved by maintaining files contigously on storage [15, 13, 4]. Other drawbacks of the traditional UNIX filesystem design are:

ffl the difficulty to integrate different media types, like RAM, harddisks and WORM optical disks to a single, virtual filesystem. A novel approach to such a filesystem is described in [12].

ffl its inadequateness for database applications [9, 3]. Database systems for UNIX, like Oracle or Exodus, circumvent the UNIX filesystem by directly implementing their own storage system on the raw devices, which makes UNIX database systems more expensive.

ffl its inadequateness for multimedia applications. Multimedia applications require high I/O throughput rates and quality of service guarantees like constant minimum data rates [1].

?This work is supported by Siemens AG ZFE, Germany, by the Union Bank of Switzerland, and by the Eidgenossisches Volkswirtschaftsdepartement, Kommission zur Forderung der wissenschaftlichen Forschung (KWF), Switzerland, Grants No. 2255:1, 2554:1
1UNIX is a registered trademark of Unix Systems Laboratories, Inc.