HC-CS-TR-2011-01: Experiences with expressing and optimizing dense numerical algorithms in AlphaZ

As multi-core processor architectures and parallel programming paradigms become more and more commonplace, there are those who concentrate on producing parallel code by hand and those who advocate the development of tools to parallelize code automatically. The AlphaZ system represents a middle road: the user provides a high-level description to the system about how the parallelization of the problem is to be accomplished, while the system generates the actual C code. In our experiences in using the AlphaZ system to express time-skewing of dense matrix codes that are not ‘embarrassingly parallel’, such as Jacobi stencils, we achieved qualified success in producing speedup across a cluster of computers. We are optimistic about the future use of AlphaZ in similar projects.