Thursday, October 9, 2003
Noon - 1 pm
Wean Hall 8220
A Two-tiered Learning Architecture for Automated Tuning of Disk Layouts
The way data is laid out on disk has a huge impact on the performance
seen in accessing that data. Many heuristics have been developed for adapting
on-disk data layouts to expected and observed workload
characteristics. In this talk I will describe a two-tiered software architecture
for cleanly and extensibly combining such heuristics. In this architecture,
each heuristic is implemented independently and an adaptive combiner merges
their suggestions based on how well they work in the given environment.
The result is a simpler and more robust system for automated tuning of
disk layouts, and a useful blueprint for other complex tuning problems
such as cache management, scheduling, data migration, and so forth. I
will also present some evaluations of a prototype system.
The utilization of freebandwidth provides an efficient way to perform
this layout reorganization on line without impacting foreground traffic.
This allows reorganization to be performed frequently and without interruption
of service. I will present a prototype system for doing reorganization
using freebandwidth, and evaluations of its performance.
Brandon Salmon is a second year graduate student in Electrical and Computer
Engineering at Carnegie Mellon University. He received his B.S. in Computer
Science from Stanford University in 2002. His research focuses on building
systems that can adapt to their surroundings, allowing them to self tune
and self heal. His current projects include Continuous Reorganization
and Self-* systems.
Further Seminar Info:
Contact Linda Whipkey, Karen
Lindenfelser or visit http://www.pdl.cmu.edu/SDI/