Simply Reliable Storage Systems

Event details
Date | 21.06.2010 |
Hour | 11:15 |
Speaker | Prof. Andrea Arpaci-Dusseau, University of Wisconsin |
Location | |
Category | Conferences - Seminars |
File and storage systems contain design flaws, implementation problems, and software bugs that can lead to system crashes, data loss, and silent data corruption. In this talk, I describe two specific contributions of our group at Wisconsin in designing and building more reliable file and storage systems.
First, I describe SQCK, a new file system checker. File system checkers are necessary in order to fix problems that may occur in file system images. However, existing checkers (such as e2fsck) are overly complex and fail in significant ways. The key contribution of SQCK is that it is based on a declarative query language, which is a natural match for the cross-checking that must be performed across the many structures of a file system image. Thus, SQCK is able to perform more useful checks and repairs than existing checkers with surprisingly elegant and compact queries.
Second, I describe the I/O Shepherd. The main contribution of the I/O shepherd is to make the reliability policy of a file system a first-class concern. With the I/O shepherd, the reliability policy of the file system (e.g., retry, parity, mirrors, checksums, and/or sanity checks) can be cleanly specified and encapsulated. We again show that with the right abstraction, even complex policies can be specified in relatively few lines of code.
Prof. Arpaci-Dusseau's homepage
Practical information
- General public
- Free