Optimizations for parallel object oriented frameworks

Application codes reliably under perform the advertised performance of existing architectures, compilers have only limited mechanisms with which to effect sophisticated transformations to arrest this trend. Compilers are forced to work within the broad semantics of the complete language specification and thus can not guarantee correctness of more sophisticated transformations. Object-oriented frameworks provide a level of tailoring of the C++ language to specific, albeit often restricted contexts. But such frameworks traditionally rely upon the compiler for most performance level optimization, often with disappointing results since the compiler must work within the context of the full language rather than the restricted semantics of abstractions introduced within the class library. No mechanism exists to express the restricted semantics of a class library to the compiler and effect correspondingly more sophisticated optimizations. In this paper, the authors explore both a family of transformations/optimizations appropriate to object-oriented frameworks for scientific computing and present a preprocessor mechanism, ROSE, which delivers the more sophisticated transformations automatically from the use of abstractions represented within high level object-oriented frameworks. They have found that these optimizations permit improved performance over FORTRAN 77 by factors of three to four, sufficiently interesting to suggest that higher level abstractions can contain greater semantics and that the greater semantics can be used to drive more sophisticated optimizations than are possible within lower level languages