Automatic Performance Estimation of SPMD Programs on MPP

ABSTRACT A methodology for the estimation of sequential and SPMD programs performance is presented. Performance metrics concerning the execution of a specific application on a target machine are derived using a machine parameterization and a symbolic application description. Both parameterizations can be obtained automatically using an integrated toolset: the Performance Estimation Toolkit. The toolset allows for the investigation of performance behaviour of applications and machines in hypothetical situations. A case study using the alpha-versions of the toolkit shows that the proposed performance model allows for prediction of tendencies of the execution time as well as absolute values.