Performance Modeling and Composition: 
                 A Case Study in Cell Simulation

           Steve G. Steinberg, Jun Yang, and Katherine Yelick 
                      Computer Science Division 
                 University of California at Berkeley

We present a case study in the use of performance modeling for
parallel application development, with a biological cell simulation as
our target application.  We show that a simple performance model is
adequate for determining data layout for arrays and linked structures,
and validate our model against experimental results for some
application kernels.  We quantify the importance of optimizing across
program components using information about machine performance and
input characteristics.  The cell simulation application has two
phases, one regular and one irregular.  The model closely predicts
actual performance within the regular phase and allows for qualitative
design comparisons in the irregular one.  The resulting application is
written in Split-C and runs on multiple platforms.