Thursday, April 13, 2006
10:30 a.m. – 11:300 a.m. - NOTE SPECIAL TIME
Intel Seminar (CIC Suite 410)
EVENTS PAGE: http://www.intel-research.net/pittsburgh/events.htm
University of Wisconsin-Madison
Data Driven Models in Storage System Design
As storage components become more complex and powerful, it becomes increasingly
challenging to take full advantage of their potential. Because of ease
of composition and development, these systems are built in layers that
have little or no information about the others. For good performance,
decisions taken at one layer need to have knowledge of the behavior of
other layers. In order to achieve this, we propose the use of data driven
empirical models. For example, a disk scheduler at the kernel level should
have an explicit model of the disk, and thus, the information it needs
in order to decide how to order I/O requests.
We have built such models for disks and RAIDs and used them in the context
of I/O request scheduling. The models are constructed such that they are
efficient to access, need no configuration, and are portable across a
wide range of devices. With simulations and experiments conducted under
Linux, we show how our empirically-driven disk scheduler noticeably outperforms
standard algorithms such as C-LOOK, and achieves the same level of performance
as more complex rotationally-aware schedulers.
Florentina Popovici is a doctoral student at University of Wisconsin-Madison,
where she is advised by Andrea C. Arpaci-Dusseau and Remzi H. Arpaci-Dusseau.
She is interested in operating systems, with a focus on storage systems.
She is one of the holders of the Datamation world sort record.
Contact Kim Kaan, 412-605-1203,
or visit http://www.intel-research.net.
SDI Home: http://www.pdl.cmu.edu/SDI/