RECOVERY ACCELERATOR TECHNOLOGIES AND EARTH SYSTEMMODELS

NASA Goddard Space Flight Centers (GSFC's) negotiated a sub-task order withNorthrop Grumman Information Technologies an effort to test software to meet the needs ofour particular use implementing the American Recovery and Reinvestment Act of 2009. Underexisting indefinite-delivery, indefinite-quantity Task Order NNG06HN01D, NG will provideall necessary support services to perform an analysis entitled, "Accelerator Technologiesand Earth System Models". The goal of this Order is to have NG evaluate the suitabilityof the current generation of hardware "accelerators" (e.g. commercial products such asthe NVIDIA Tesla series of graphical processing units (GPU's)). This class ofcommercially available processors can potentially improve performance of key aspects ofNASA's weather and climate models. The Contractor's evaluation should primarily analyzethe potential performance benefits against the labor costs for adapting NASA's softwareto exploit this technology. In addition to the software costs, the Contractor willanalyze costs associated with integrating accelerators within a computing environmentincluding power and cooling considerations. The computational kernels of NASA'sweather/climate models to be considered within this task are: 1)Solar radiation (SORAD)component from the GEOS-5 model; 2) Cubed-sphere Finite Volume Dynamical Core (CSFVCore);and 3) Sparse Matrix Vector Gear Code (SMVGEAR) from the Global Modeling Initiative (GMI)model.

For each of these kernels, NG shall perform an initial evaluation to determine the beststrategy for studying the potential performance impact of accelerator technology. When afull port for a given kernel is not feasible, NG shall provide a written explanation ofboth the nature of the difficulty and a rationale for a proposed alternativeimplementation where appropriate. This explanation shall also estimate the possible errorin the performance estimation due to such simplifications. NG shall commit the finalimplementation prototypes to the sourcemotel.gsfc.nasa.gov repository.

NG shall deliver a draft and final version of a comprehensive report which summarizes thepotential performance impact of accelerator technologies for NASA Earth system modeling.This document must address the degree to which the selected kernels are representative ofthe computational work on NASA's high-end computers. The document must also estimate theconversion/maintenance costs associated with migrating full models to exploit acceleratortechnologies. The report must estimate the potential for future support for compilerand/or Fortran support (e.g. Portland Group GPU Fortran directives) for using thistechnology within scientific models. GSFC will make the final report available to thepublic. The following lists the deliverables under this Sub-Task Order: 1)Weekly: oralstatus reports to the Contracting Officer Technical Representative (COTR); 2) Start + 4weeks: Written evaluation strategy for all 3 computational kernels delivered to COTR andContracting Officer (CO); 3) Start + 20 weeks: Prototype implementation of all 3computational kernels delivered to COTR; 4)Start + 22 weeks: Draft version of acomprehensive report described above delivered to the COTR. GSFC will return it'spreliminary 2 business weeks from receipt for further processing to NG; and 5) Start + 24weeks: Final version of a comprehensive report described above delivered to the COTR andthe CO.

The existing Task Order is a Cost-Plus-Award-Fee arrangement awarded to NG 10/2005 on acompetitive basis. CPAF was determined appropriate because the contract is for theperformance of research and development effort, and uncertainties involved in contractperformance do not permit costs to be estimated with sufficient accuracy to any type offixed-price contract.