Performance prediction with benchmaps

Abstract

Benchmapping is a performance prediction method for data-parallel programs that is based on modeling the performance of runtime systems. This paper describes a benchmapping system, called BenchCVL, that predicts the running time of data-parallel programs written in the NESL language on several computer systems. BenchCVL predicts performance using a set of more than 200 parameterized models. The models quantify the cost of moving data between processors, as well as the cost of moving data within the local memory hierarchy of each processor. The parameters for the models are automatically estimated from measurements of the execution times of runtime system calls on each computer system.