Metadata for Balanced Performance

Paul Brown, Richard Troy
University of California at Berkeley
Computer Science Division, Berkeley, CA 94720
{pbrown, rtroy}@postgres.berkeley.edu

and

Dave Fisher, Steve Louis, James R. McGraw, Ron Musick
Lawrence Livermore National Laboratory
P.O. Box 808, Livermore, CA 94551
{dsf, stlouis, jmcgraw, rmusick}@llnl.gov

Abstract

Data and information intensive industries require advanced data management capabilities incorporated with large capacity storage. Performance in this environment is, in part, a function of individual storage and data management system performance, but most importantly a function of the level of their integration. This paper focuses on integration, in particular on the issue of how to use shared metadata to facilitate high performance interfaces between Mass Storage Systems (MSS) and advanced data management clients. Current MSS interfaces are based on traditional file system interaction. Increasing functionality at the interface can enhance performance by permitting clients to influence data placement, generate accurate cost estimates of I/O, and describe impending I/O activity. Flexible mechanisms are needed for providing this functionality without compromising the generality of the interface; we are proposing active metadata sharing. We present an architecture that details how the shared metadata fits into the overall system architecture and control structure, along with a first cut at what the metadata model should look like.

Appeared in

IEEE Metadata 1996

Look at the Paper (ps.gz, pdf.gz)