|
Abstract : |
The Fortran D language extends Fortran by permitting the user to specify the distribution of array variables across the processors of a parallel computer system. This information can then be used by the compiler to derive a multidimensional parallelization. For programs in which multidimensional arrays have been linearized for optimal performance on vector processors, this strategy will not produce the best results because in those programs Fortran D is limited to a one-dimensional parallelization, which yields less efficient communication than multidimensional parallelization because of surface-to-volume effects. This paper proposes Fortran D extensions and associated compiler technology to support natural topology parallelization of applications ' codes with linearized arrays. Experimental results are presented to illustrate the improvement in performance of this new approach over the performance of currently available methods such as Fortran D and the PARTI inspector/executor strategy. 1, |
