Abstract |
||
|
Many computer programs use a large amount of data in changing patterns.
The question for our research is whether complex programs have a
predictable pattern of data access and, if so, to what degree that pattern
can be modeled, measured, and modified (improved). I will first motivate our work by examining the bandwidth bottleneck on modern processors using the balance model of Callahan, Cocke, and Kennedy. A compiler can significantly reduce the bandwidth demand of a wide range of programs. I will demonstrate the effect of global and dynamic transformations and present a technical survey. I will also introduce a data-locality model called group affinity and compare it with other locality models. In the second half, I will present a training-based analysis for general-purpose programs. Based on a few sample runs of a program, the second step constructs a parameterized model that predicts how reuse pattern changes in other program inputs. The analysis predicts the cache miss rate for all cache sizes and all program inputs, including those that are magnitudes larger than the sampled executions. |
||
Speaker's Biography |
||
|
Prof. Chen Ding is an assistant professor in the Department of Computer Science,
University of Rochester. He received Ph.D. from Rice University, M.S.
from Michigan Technological University, and B.S. from Beijing University,
all in computer science. He is a recipient of a young investigator award
from the Office of Science of the U.S. Department of Energy, a Career
award from National Science Foundation, and a best-paper award from IEEE
International Parallel and Distributed Processing Symposium. In 2002, he
co-organized the first ACM SIGPLAN Workshop on Memory System Performance.
He is currently a vice chair of the 2004 International Conference on
Parallel Processing and the general chair of the 2004 ACM SIGPLAN Workshop
on Memory System Performance. More information about this research is available on the Web at http://www.cs.rochester.edu/~cding/. |
||
|
Dirk Stroobandt, e-mail: dstr@elis.UGent.be |