Systems Seminar: Remzi Arpaci-Dusseau, University of Wisconsin

IRON File Systems

Abstract

Disks fail -- but not in the way most commodity file systems expect. For many years, file system and storage system designers have assumed that disks operate in a "fail-stop" manner, where disks either work perfectly or fail absolutely in an easily detectable manner. Unfortunately, modern drives exhibit much more complex failure modes, including latent sector faults and block corruption.

In this talk, I will present a detailed examination of disk failure, describing a new "fail-partial" failure model that incorporates these more realistic faults. I will then show what happens to commodity file systems when such failures occur, demonstrating numerous odd inconsistencies and serious bugs in many popular file systems. Finally, I will discuss a new class of file system we are building at the University of Wisconsin; such "IRON" file systems have built-in mechanisms to detect and recover from a broad range of modern disk failures.

Speaker Bio

Remzi H. Arpaci-Dusseau is an assistant professor at the University of Wisconsin, Madison. He received his B.S.C.E. from the University of Michigan, Ann Arbor (summa cum laude) and his Masters and Ph.D. from the University of California, Berkeley, under the supervision of David Patterson. Remzi's research interests center around file and storage systems, with a current focus on techniques to design and implement more robust, fault-tolerant file systems.