The current challenge in performance analysis is to give the
software
developer understandable results with a minimum of learning overhead
and effort. We advocate a simple performance model, based on the lost cycle approach of Crovella [Crovella,
M. E. and Leblanc, T.J.: Parallel Performance Prediction using Lost
Cycles Analysis. In: Proc. of Supercomputing 1994, IEEE Computer
Society (1994)]. It offers insight in the performance of
applications using a terminology close to that of the user. This
research resulted in EPPA,
a tool that analyzes the performance of parallel applications, which is
currently used by our students in the practicum of the Parallel
Systems course.
Next ,we developed a novel approach to study the influence on the
performance of application and system parameters, design decisions and
other variables. It is based on learning
causal models from experimental data. The EPDA
tool was developed for collecting the data from experiments, store them
in a database and present then to the user in a graphical way.
An old overview about this research: Automated
Experimental Parallel Performance Analysis (2002)
More information & publications: Jan
Lemeire (jan.lemeire@vub.ac.be)
