In high-performance computing symbolic performance estimation is
an important engineering tool that allows rapid assessment of
application scalability and the selection of coding or data
partitioning alternatives. In this paper we present an automatic
cost estimator that compiles data parallel programs into symbolic
performance models of very low solution complexity. With minimal
program annotation by the user, symbolic cost models are generated
in a matter of seconds, while the evaluation time of the models
ranges in the milliseconds. Experimental results on four data
parallel programs show that the average prediction error is less
than 15%. Apart from providing a good scalability assessment,
the models correctly predict the best design choice in all cases.
|