SITIS Topic Details |
||||||
| Proposals Accepted: | |
| Program: | STTR |
| Topic Number: | AF10-BT13 (AirForce) |
| Title: | Highly-Scalable Computational-Based Engineering Algorithms for Emerging Parallel Machine Architectures | Research & Technical Areas: | Air Platform |
| Objective: | Develop domain partitioning and/or execution algorithms for computing aero-loads and aero-elastic phenomena by way of highly-parallel solutions to the equations of fluid and structural dynamics.
| Description: | Modern, production-quality solvers for the governing equations of fluid dynamics (CFD) and structural dynamics (CSD) are generally based upon solving discretized versions of the governing equations on a domain that has been partitioned over a specified number of compute cores. CFD solutions are amenable to iterative techniques running on distributed memory architectures, and state-of-the-art solvers have demonstrated linear scalability up to about 4,000 processors. On the other hand, CSD solutions are better suited for large, shared-memory machines. Computational tools are rapidly moving into the areas of multi-disciplinary analyses such that these solver types must execute together on machines where core counts are quickly approaching 105 and where the individual nodes contain upwards of 32 compute cores with a corresponding amount of shared memory on the node. Future computational-based engineering tools must be able to operate efficiently in these environments and possess linear scalability such that high-fidelity simulations are completed within the time required to impact acquisition and certification activities. To date, ad-hoc attempts have been made to apply hybrid MPI/OpenMP techniques to take advantage of the many-core processors with limited success. However, formal investigations into novel looping strategies, cache memory usage techniques, and other areas in the context of emerging parallel machine architectures have been very limited. The validity and consistency of the solved equations should obviously be retained in any implementation, but consolidated efforts to couple transformative partitioning techniques with hybrid parallel execution models and scalable numerical algorithms should yield fertile ground such that CFD/CSD solvers may take advantage of the massively-parallel machines currently going into general production (upwards of 50,000 compute cores).
| PHASE I: Survey the state-of-the-art, highly-scalable CFD/CSD algorithms, develop/prototype execution strategies/execution techniques that enables these types of solvers to scale to tens-of-thousands of processors on emerging HPC platforms while preserving the physical consistency/validity of the equations.
| PHASE II: Apply the theory and prototype methods developed in Phase I to a production-quality CFD/CSD solver and demonstrate linear speedup on computations spanning tens-of-thousands of processors.
| PHASE III | DUAL USE COMMERCIALIZATION:
Military Application: The computational algorithms developed will be directly applicable to computational-based engineering tools used by all military branches as well as many other non-DoD government organizations.
Commercial Application: Commercial air and space companies will benefit from the developed products since the results are equally applicable to in-house computational tools.
| References: |
1. Post, D.E., Arevalo, S., Atwood, C., Bell, P., Blacker, T.D., Dey, S., Fisher, D., Fisher, D.A., Genalis, P., Gorski, J., Harris, A., Hill, K., Hurwitz, M., Kendall, R.P., Meakin, R.L., Morton, S.A., Moyer, E.T., Strawn, R., van Veldhuizen, D., Votta, L.G., Wynn, S. and Zelinski, G., Journal of Physics: Conference Series vol. 125 (2008) 012090. 2. Morton, S.A., McDaniel, D.R., Sears, D.R., Tuckey, T.R., Tillman, B., “Rigid, Maneuvering, and Aeroelastic Results for Kestrel - A CREATE Simulation Tool,” 48th AIAA Aerospace Sciences Meeting, 5 - 8 January 2009, Orlando, Florida. 3. Mohagna, P., Frink, N., Khaled, A.H., Chung, J., “Recent Enhancements to USM3D Unstructured Flow Solver for Unsteady Flows,” AIAA-2004-5201, 22nd Applied Aerodynamics Conference and Exhibit, Providence, Rhode Island, Aug. 16-19, 2004. 4. Grismer, M.J., Strang, W.Z., Tomaro, R.F., and Witzemman, F.C., “Cobalt: A Parallel, Implicit, Unstructured Euler/Navier-Stokes Solver,” Adv. Eng. Software, Vol. 29, No. 3-6, pp. 365-373, 1998. 5. Mavriplis, D. J. and Venkatakrishnan, V., “A Unified Multigrid Solver for the Navier-Stokes Equations on Mixed Element Meshes,” International Journal for Computational Fluid Dynamics, Vol. 8, 1997, pp. 247–263. |
| Keywords: | algorithms, aero-loads, aero-elastic phenomena, structural dynamics |
Questions and Answers: |
Q: Do the target "emerging HPC platforms" include GPU enabled systems, or is this solicitation focusing mainly on multi-tiered architectures that use multi-core CPUs on the individual nodes? |
A: . . . response pending . . . |
Q: Q1. Is there a preference for application codes in the "CREATE" family of codes |
A: A1: The Kestrel and Helios solvers both use unstructured explicit/point implicit solvers parallelized using the PARMETIS utility code. This scalability is limited by the ratio of mesh cells/cores on the order of 10,000 to 100,000 which is reached by this class of solvers when the overall core count of a single simulation is in the 1,000's. For this reason the government is interested in improving a class of codes on machine architectures of the future that will use multi-core technology to increase the overall machine core count by 10 to 1000X. The class of solvers for CFD are the unstructured explicit or point implicit Navier-Stokes solvers such as AFRL/RB's AVUS, Cobalt LLC's Cobalt, NASA's USM3D, University of Alabama Birmingham's Hyb3D. This class of code has demonstrated scalability to 1000's of processors but the need for the future is 10,000's of cores for a single simulation. The improvement to scalability should be due to algorithmic improvements not simply a specialized hardware solution. The class of codes for CSD should be Finite element Solvers similar to MSC NASTRAN, ABAQUS, ANSYS. |
Q: Is it possible to submit an SBIR-based proposal for this topic? |
A: The main difference between an SBIR and an STTR is that an STTR requires a University partner to participate in the project. Also, given the nature of the work AFOSR does, we are generally looking for a balance between basic and applied research in the project. An SBIR does not necessarily cover basic research. There is a financial minimum of 30% that must go to the university partner and 40% to the small business. |
Q: If "CREATE" codes are of interest, would it be feasible to obtain either the AVUS, kAVUS, or Cobalt60 source code during Phase I? Is it possible to get access (to the actual code, or a similar code/matrix) during proposal preparation? |
A: It would be very useful to use AVUS or Cobalt60 as the baseline codes for the study. It is the responsibility of the respondent to obtain a code in the class of codes of interest from the originators of the code. The government is interested in improving a class of codes on machine architectures of the future that will use multi-core technology to increase the overall machine core count by 10 to 1000X. The class of solvers for CFD are the unstructured explicit or point implicit Navier-Stokes solvers such as AFRL/RB's AVUS, Cobalt LLC's Cobalt, NASA's USM3D, University of Alabama Birmingham's Hyb3D. This class of code has demonstrated scalability to 1000's of processors but the need for the future is 10,000's of cores for a single simulation. The improvement to scalability should be due to algorithmic improvements not simply a specialized hardware solution. The class of codes for CSD should be Finite element Solvers similar to MSC NASTRAN, ABAQUS, ANSYS. |
As of midnight September 1, questions for solicitations SBIR 10.3 and STTR 10.B will no longer be accepted.
To read the solicitation for full proposal preparation and submission details click here. |