Database system performance heavily depends on data access efficiency on all levels of the memory hierarchy. CPU cache performance depends on in-memory data layout, whereas I/O performance depends on data placement on the disks. Throughout the memory hierarchy, however, the workload behavior is a function of the access patterns the application dictates. On-line transaction processing applications, for instance, tend to access full data records in a random fashion, whereas decision-support systems typically scan large tables sequentially, reading small fractions of the records. Unfortunately, current database storage systems are static and fragile. Low-level system parameters are fully configured by human administrators, whereas there is typically a single data organization used at all levels of the memory hierarchy, and data organization is specialized to a single workload. This results in frequent errors, suboptimal performance, and inability to take advantage of the storage technology at its full potential.

This talk describes Fates, a dynamic, robust, and automated system for database storage management. Fates decouples in-memory data layout from on-disk storage layout, providing the opportunity to design efficient data placement at each storage level separately. In addition, it provides device-specific hints and automatically specializes the system to extracted device characteristics. To efficiently place data on disk, we implement logical-to-physical volume mapping and arrange storage to offer robust performance across workloads, even when accessing multi-dimensional data. The talk will describe the Fates system design and implementation, and will demonstrate its efficiency through experimental results with application benchmarks.

Thursday, October 27, 2005

IRCS Large Conference Room 470,

3401 Walnut Street

3:00 p.m. - 4:30 p.m.