High-End Computer System Performance: Science and Engineering ISIC
PI: David Bailey (LBNL), TSTT Point of Contact: Dan Quinlan (LLNL)

The High-End Computer System Performance Science and Engineering (HCSP) ISIC proposal will focus Performance Analysis and Optimization for complex memory-hierarchies. Such architectures are common to all modern parallel computers and present difficult challenges for scientific computing generally. The tools being developed within the performance ETC include tools to both analyze and report existing performance and explicitly optimize application source using source-to-source transformations. The mechanisms developed for source-to-source transformations automate the introduction of optimizations that address memory-hierarchies (particularly cache). This approach is specifically directed to the compile-time optimization of libraries such as the Discretization Library that we propose.

Our development of the Discretization Library and its use of predefined mathematical operators to simplify the development of applications using complex computational meshes will have direct benefits from the research work on ROSE proposed within the HCSP ETC. Since the operators defined within the Discretization Library are used within user defined applications, the library cannot see the context of their use with one another. Only a compile-time mechanism can see the sufficient context of the use of multiple operators within the application source (within and across program statements). The work on ROSE by the HCSP ETC is explicitly a mechanism to recognize the use of high-level abstractions and automate the source-to-source transformations required to make them efficient. We are working closely with researchers there in the development and testing of the ROSE compile-time approach to automate optimization of application codes using our Discretization Library. Current work has demonstrated initial parts of this work using relatively simple operators and improved the performance by a factor of 4-6 on structured grid computations. Additional work will define benchmarks specific to evaluating the performance of the Discretization Library. This proposal shares some of the same personnel with the HCSP ETC. We expect that the Discretization Library itself will be influenced by the development of this optimization research within the HCSP ETC.