PPT - Performance Prediction Tools for Data Redistribution in HPF and MPI Irregular computations
Period : 06 / 1998 - 06 / 2000
Responsible : Francisco Fernández Rivera. Depto. de Electrónica e Computación. Facultade de Física. USC.
Team:
Tomás Fernández Pena, José Carlos Cabaleiro Domínguez, Vicente Blanco Pérez, Dora Blanco Heras, Juan José Pombo García, Patricia González Gómez, Marcos Boullón Magán. Electronic and Computation Deparment. Physics Faculty. USC.
Description :
A performance prediction tool is proposed in this project to establish the most efficient way to solve irregular computations on distributed memory systems, and in particular on the AP3000. We focus on HPF codes that call MPI routines to solve sparse matrix computations. Our tool will decide if this routine can be used efficiently in terms of runtime and memory overhead or not. In order to do this, a prediction tool of the sequential, the HPF and the MPI versions of the matix computations is needed, as well as the overhead associated to the redistribution of data among them.