[PRL] Mary Lou Soffa @ Harvard 10/23

Mitchell Wand wand at ccs.neu.edu
Wed Oct 1 17:37:00 EDT 2003


Title: Mary Lou Soffa, Department of Computer Science, University of
Pittsburgh

Date:      October 23, 2003

Location: Maxwell Dworkin G125, (Tea at 3:30PM - Maxwell Dworkin 2nd
Floor Lounge Area) Time: 4 p.m. - 5 p.m.

Description
-----------
Harvard University Computer Science Colloquium Series
33 Oxford St., Cambridge, MA 02138 
tel: (617) 496-1440 fax: (617) 495-9837

Colloquium

Profit Driven Optimizations 

Abstract

Although optimizations have been applied by compilers for over 40
years, much of the research has been devoted to the development of
particular optimizations. Certain problems with code optimizations
have yet to be adequately addressed, including applying only
profitable optimizations, ordering optimizations, selecting
optimization configurations and combining optimizations. With the
rapidly growing use of cost-sensitive embedded systems and the
interest in dynamic optimization, handling these problems is becoming
all the more important.  In this talk, I will discuss our current work
in developing a framework for optimizations that can be used to
determine important properties of optimizations, and in particular,
the profitability of optimizations.  Our unifying framework includes
code, optimization and resource models for systematically exploring
the application of optimizations.  The framework will provide both
analytical and experimental models for understanding, predicting and
verifying the properties of optimizations (i.e., performance impact
and interactions). Also, practical and automatic strategies to drive
the application of optimizations based on the models are part of the
framework. By applying these model-based optimization strategies, the
goal is that optimizing compilers will be able to produce higher
quality code and use different paradigms than what is possible with
current approaches.

Host:    Professor Norman Ramsey


Calendar URL
- ------------
http://www.fas.harvard.edu/



More information about the PRL mailing list