A Performance-Oriented Technique for Hybrid Application Development

摘要

In SMP clusters it is not always convenient to switch from pure message-passing code to hybrid software designs that exploit shared memory. This paper tackles the problem of restructuring an existing MPI code through the insertion of OpenMP directives. The choice of the best code is carried out with a performance-oriented approach, predicting the effect of application hybridization in the MetaPL/HeSSE simulation environment, without writing and running any hybrid software. The technique is validated by applying the devised changes to the code, and comparing the predicted results to actual running time measurements.