Non-local Data Reuse in Data Parallel Compiler

Abstract

Data parallel language is treated as a promising paradigm to describe a parallel program. In most of data parallel languge, non-local data is disposable. Every time there need non-local data, communication should be done. However some non-local data are used more than twice before they are redefined. To eliminate redundant communication for them, they must be reused. CHAOS and some communication optimization try to reuse that butt their efforts are limited because of the difficulty of coherence between original data and copied one. We propose a new non-local data reuse method. Our method uses data flow dependence to know the redefinition time of non-local data. Our method is composed of two stages. In the first stage, it detects redefinition time and reusable time of non-local data, and in the second stage, it transforms original code to reuse non-local data. We solve the coherence problem and reuse more non-local data. We make an experiment to know the effect of our method in networked workstation environment.