Return to the Table of ContentsObjective: Centric Engineering Systems, as part of the HPCC Testbed project, will demonstrate the efficacy of the IBM's highly parallel SP2 system to simulate the aeroelastic behavior of real-life wings operating at transonic flight conditions and an operational automobile engine cooling fan. The first phase of this effort involves porting Centric's commercially-available multiphysics finite element simulation program to the SP2.
Approach: Centric's Spectrum product was designed to solve multiphysics problems ╤coupled interaction between fluid, structural and thermal physics. One aspect of the multiphysics approach is the capability to support multiple subdomains ╤a quantum of work comprising uniform physics. The first step in the parallelization effort is to modify the various Spectrum modules to use message-passing to communicate between the subdomains. The next effort is to incorporate a robust domain decomposition algorithm to further divide the natural subdomains to match the appropriate number of processing nodes and to assist in load balancing.
Accomplishment: Centric, as a proof-of-concept, has implemented a prototype message-passing version of Spectrum, first using the public-domain PVM library. This interface was then modified to use the IBM MPL library to make use of the high-speed switch on the SP2. Several real-world, industrial simulation problems have been run to verify the approach and to begin to characterize performance and scalability. For example, one customer-supplied problem simulated incompressible turbulent flow around a bluff body solving for pressure, velocity and intensity of turbulence. The problem uses about 160,000 3-D brick elements.
| System | Wall time | |||
|---|---|---|---|---|
| Speedup Efficiency | ||||
| ''Typical'' Engineering Workstation | 97.0 hours | |||
| IBM 560 4 WS cluster | 45.0 hours | |||
| IBM 590 Workstation | 38.6 hours | 1.0 | 1.00 | |
| IBM SP2 (4 node) | 9.9 hours | 3.9 | 0.97 | |
| IBM SP2 (8 node) | 5.5 hours | 7.0 | 0.88 | |
| IBM SP2 (16 node) | 4.2 hours | 9.2 | 0.57 | |
The tail-off for the 16-node run appears to be due to the small amount of work per node (~10,000 elements/node). Larger and more complex problems will be run to better determine both strengths and weaknesses of the prototype. However, the results at this point are very encouraging.
Significance: This work is significant since it demonstrates a commercial application solving real-world, industrial design problems on parallel computer systems. With this combination of parallel hardware and software, it is now possible to solve complex, non-linear, multiphysics design problems in ''engineering reasonable time''. Work will continue to be done solving larger and more complex problems, characterizing performance, automating the domain decomposition steps, examining I/O performance. However, the main tool is in place to begin investigation of aeroelastic design problems as defined in the HPCC CRA.
Point of Contact:
Return to the Table of Contents