[an error occurred while processing this directive] [an error occurred while processing this directive]


University of Saskatchewan, Department of Computer Science


2001-2002 Seminar Series

Mammoth: A Robust and Confederated Global File System

Mike Feeley
Assistant Professor
Department of Computer Science<> University of British Columbia

DEPARTMENT SEMINAR
DATE: Monday, March 18, 2002
TIME: 3:30pm
PLACE: Anthropology 132
*** Everyone is welcome ***

Abstract

Modern file systems don't work. Not really. File data is accessible from only a small set of computers (perhaps one) that have been carefully configured by a human administrator; this data can be accidentally deleted or overwritten by careless users and malicious programs; and it may be lost when disks, systems, networks or physical facilities fail. While most of us accept these core weaknesses without much thought, they are poorly suited for the modern world in which people generate and manipulate an increasingly wide variety of digital content such as home movies, photographs, financial data, business data, journals, cookbooks, web pages and on and on. It would be perfectly reasonable for users to expect their valuable content to be stored safely and to be accessible any time from any where. This, however, is almost never the case. In this talk I will describe a file system we are building at UBC, called Mammoth, that is designed to really work. Our goal is to provide scalable, globally-accessible storage that automatically protects data from accidental or malicious manipulation and from all forms of software, hardware and physical failure. Scalability and accessibility are achieved by storing data on a loosely-connected, peer-to-peer, confederation of PC nodes. This structure allows the entire system to be comprised of perhaps many thousands of nodes, but with each node aware of only a few of its neighbors. Data is protected by retaining older versions of files and directories and by automatically replicating some versions to other nodes. User actions can thus be easily undone by restoring an older version and failures can be easily recovered from by accessing versions stored on nodes that have not failed. Updates to replicated data are handled in an optimistic manner to allow nodes to read and write whatever version of data they currently have access to, even if failures prevent the system from guaranteeing that these are the current versions. Inconsistencies that arise due to network partitions are automatically detected and corrected when communication is re-established. In the talk, I will describe our design and provide some details of our prototype implementation, which is still being developed.
About the speaker

Mike Feeley has been an Assistant Professor of Computer Science at the University of British Columbia since 1997. His research interests span a wide range of topics related to operating systems and distributed systems, including file systems, the web and PC clusters. Mike received his Ph.D. from the University of Washington in 1996.

[an error occurred while processing this directive]