Workshop on Geoscience Application Requirements for Petascale Systems

The geosciences community needs to collaborate to formulate a clearly defined set of computational requirements that vendors and potential resource providers can incorporate into the design processes for their next-generation systems. At the same time, the geoscience modeling community must collectively understand general high performance computing (HPC) technology trends to effectively design and implement applications for this new class of equipment. These circumstances suggest the reason for organizing this workshop: it is beneficial to establish an in-depth, long-term dialog between the scientific researchers, computational scientists, application domain experts, and vendors of HPC system components. The Workshop on Geoscience Application Requirements for Petascale Systems (GARPA) addressed the challenges of preparing geoscience applications to run on petascale (and beyond) computing systems.

The NSF-funded GARPA workshop was aimed at developing a set of specific recommendations for creating a better dialog between all parties involved in high-end geoscience computing. There were clear social, technical, and political obstacles to establishing this dialog. The workshop, held in Arlington, Virginia near NSF headquarters, was charged with overcoming these and other obstacles in group discussions that included NSF staff.

GARPA produced a number of positive outcomes. Community-building mechanisms were established between the modelers and scientists in the different branches of the geosciences. The geosciences application developers learned which aspects of supercomputer design they can influence, and they learned how they could coordinate with each other to strengthen their influence on vendors and other resource providers. Groundwork was established for creating a common geoscience benchmark suite. Finally, benchmarking the largest supercomputers will be problematic, so all groups agreed that predictive models of system performance be created based on key architectural parameters, and that I/O subsystem performance prediction models be created for large-scale applications.

For the future, GARPA attendees recommended merging the GARPA workshops with the sister Petascale Computing and the Geosciences Workshop organized by Allan Snavely of SDSC. The merged workshop will be called GARPA-2 and will be held January 7-9, 2007 in Boulder, Colorado. Attendees will debate and refine a draft set of recommendations and findings from the first workshop, and this will serve as a guide to further action for all parties.

Benchmark results will be presented for two preliminary benchmarks, the WRF mesoscale meteorological model and the POP 2 global ocean model. Since GARPA-1, the Southern California Earthquake Center (SCEC) PetaShake application has been put forward as a benchmark from the Earth science domain, and the parallel NetCDF I/O benchmark IOR has been suggested as the first I/O code for the GARPA benchmark suite. We anticipate that other candidate benchmarks will emerge as a result of the first workshop.

The outcome of GARPA-2 is expected to be a set of recommendations for further actions to establish a sustainable effort in the area of geoscience application characterization, to be provided to the sponsoring NSF program officers.