It has been long recognized that numerical simulations hold the key
to further progress in various scientific fields, but are hindered
by the inability to obtain the necessary compute power.
It is well known that detailed performance models aid
in understanding the performance relationship between
the parallel system and the application; this understanding is necessary
to reduce the large performance gap. The CELERO Group has completed
significant work in this area that includes developing and
validating performance models for scientific application, using these
models to identify future system performance requirements, using the
models to develop more efficient algorithms and developing techniques
for efficient load balancing. The target scientific application are
graph-based application which include molecular dynamics, finite
element and transportation.
Our current work in this area is focused on researching methods, for
extracting performance information, that can be automated.
The Performance Modeling project is composed of several sub-projects:
