|
Tuesday, April 19, 2005 |
|
|
|
|
Semantically-Smart Disk Systems |
|
|
Robust and efficient storage of data is a prerequisite of current and future computer systems. To keep pace with both rapid changes in technology as well as increasing demands from applications and users alike, storage systems must evolve in new and interesting ways -- innovation in storage is therefore of paramount importance. Unfortunately, storage systems today have a problem: the range of functionality they can provide is fundamentally limited, despite the presence of significant processing power within them. The main reason for this limitation is that storage systems today lack higher-level "semantic" understanding about how they are being used. Simply put, today's disk systems are "dumb", but they need to be smarter to keep pace with the growing demands placed upon them. I propose a novel solution to this fundamental problem. Specifically, we have developed a a new class of storage systems called "semantically-smart disk systems" (SDS's); such disk systems are capable of providing entirely new classes of functionality by exploiting information about the system above (e.g., a file system or a database management system). An SDS does so by carefully monitoring the low-level stream of block reads and block writes that a storage system normally sees, and then inferring higher-level behaviors of the system above. Importantly, an SDS does so without any changes to the existing block-level storage interface, taking a pragmatic approach that enables ready deployment in existing computing environments. In this talk, I will first describe various techniques used by an SDS to track semantic information, demonstrating how to transform an I/O request stream into a source of useful high-level information for the underlying disk system. I will then demonstrate the utility of semantic information by presenting new improvements to the availability, security, and performance of storage. Such innovations are impossible to implement in the current storage infrastructure, but become possible with the acquisition and careful use of semantic information. Bio: Muthian Sivathanu is a graduate student at the University of Wisconsin-Madison, under advisors Andrea and Remzi Arpaci-Dusseau. He received a B.E. in Computer Science from Anna University India in 2000, and an M.S. in computer science from the University of Wisconsin, Madison in 2001; he plans to complete his dissertation in May 2005. His interests lie broadly in the design, construction, and evaluation of computer systems, with a recent focus on file systems and storage.
|
|