| Abstract |
Irregular applications such as partial differential equation solvers and
molecular dynamics simulations frequently exhibit poor performance on contemporary computer architectures in large part due to their
inefficient use of the memory hierarchy. Run-time data and iteration reordering transformations have been shown to improve the locality and
therefore the performance of irregular benchmarks. In this talk, I will review my contributions to run-time data and iteration reordering
transformations, which include the development of a framework for the compile-time composition of such transformations. I will then introduce
the spatial and temporal locality hypergraphs, which we have developed to predict the combination of run-time data and iteration reordering
heuristics that will result in the best performance for a given dataset. By developing heuristics that focus on improving the
metrics based on these models, we have developed iteration reordering heuristics that
result in better performance than existing techniques. Finally, I will overview the OpenAnalysis toolkit and describe how OpenAnalysis will be
used to automate the compile-time composition and code generation of run-time reordering transformations.
|
