Seminars

DATE: Thursday , October 23, 2002
TIME: Noon - 1 pm
PLACE: Wean Hall 8220

SPEAKER:
Jiri Schindler
Graduate Student, ECE, CMU

TITLE:
Lachesis: A storage device access patterns advisor for database systems

ABSTRACT:
While the efficiency, and thus the response time, of a single I/O operation is highly variable, the notion of a variable-cost I/O is not taken into account by today's database systems. Current optimizers choose a query plan that minimizes the total number of I/O operations using a fixed-cost model for a single I/O. The resulting I/O access pattern, governed by the chosen query plan, is then optimized by the storage manager using implicit assumptions about storage device characteristics and a series of configuration parameters. However, there is no guarantee that it will be most efficient (i.e., having the shortest execution time) from the storage device's point of view.

To minimize the total I/O cost of a query, I propose to provide explicit information about individual I/O costs that can be used to choose access patterns that are most efficient. In this talk, I will describe a storage device access patterns advisor, called Lachesis, which allows a database storage manager to choose the most optimal access patterns and automatically adapt to the underlying storage device characteristics. It does not require any changes to the existing data page layout or algorithms inside the storage manager. It merely provides more detailed, and accurate, information to the storage manager which are given in a storage device-independent format. Our preliminary results show a 1.2-2x speed up for the majority of TPC-H queries in environments with competing traffic.

BIO:
Jiri Schindler is fifth-year graduate student at Carnegie Mellon's Parallel Data Laboratory working under Prof. Greg Ganger. His research areas include file and storage systems. Other projects he has been working on include disk characterization, timing-accurate storage emulation, and freeblock scheduling inside device driver.

For Further Seminar Info:
Visit http://www.pdl.cmu.edu/SDI/