|
|
|
|
|
The majority of the studies utilize wall clock timings combined with careful experimental design. The studies are "vertically integrated", using low-level tests to help interpret or design the experiments using kernel codes, which are themselves used to design experiments using the production codes. The studies look at issues of performance sensitivity, varying algorithm, problem, or machine parameters to determine good (and bad) ways of running the codes. Some special studies also use profiling or tracing and visualization to analyze the performance of the codes. For sanity's sake, since we run on many platforms, we use the portable tracing and performance visualization packages MPICL and ParaGraph.
The impetus behind the initial studies was the design of a message-passing parallel implementation of the NCAR Community Climate Model , as part of the Department of Energy CHAMMP research program. Performance studies have continued as the climate model components and the computing platforms evolve.
For these web pages, the exposition is organized by the codes used:
However, these links do not represent different studies, just different facets of a long-term examination of the (message-passing) performance of parallel spectral AGCMs. The studies are also continuing, and the presentation in these web pages is (and will continue to be) in flux. In particular, we have begun looking at the performance of hybrid parallel algorithms (message-passing and shared memory) on clusters of shared memory parallel architectures, and will be incorporating these results in the near future. Note that much of the description of metholodgy and experimental design is repeated (ad nauseum) in these and subordinate pages, but the intent is that individual web pages be complete unto themselves to the extent possible.The primary motivation for these web pages is to organize the results for easy interpretation by the authors. Visitors are welcome to rummage around, and we welcome questions or suggestions. Please contact us before using these results in publications, so that we can indicate the most relevant citation to use.
Publications, presentations, and tools related to this research can be found here.
Patrick H. Worley / (
worleyph@ornl.gov)