Return to the Table of ContentsObjective: Evaluate structural analysis algorithms for large-scale HSCT configurations on new parallel HPCCP supercomputers.
Approach: Develop general-purpose structural analysis algorithms for efficient structural analysis which exploit the new parallel HPCCP computers. An equation solution algorithm, PVSOLVE, was developed on a shared-memory Cray computer to exploit multiple processors. Using loop unrolling with SAXPY operations and minimizing synchronization, the code was optimized for 1 to 16 Cray C-90 processors. The code was then modified for distributed-memory computers (i.e. Intel Paragon and IBM SP-2). For the Paragon, the code was optimized using the Intel BLAS1 DDOT (dot product) routine. The initial implementation on the IBM SP-2 is a direct port of the Paragon code with changes in the communications calls. Early access to the IBM SP-2 under the NASA-IBM Memorandum of Agreement have also allowed access to key IBM personnel. Accomplishment Description: The best performance in the solution of the 88,404 equation HSCT was achieved using on optimized version of PVSOLVE on the Cray C-90 (see figure). The performance on the Cray and IBM were both found to be scalable. However, due to the limited memory on each Paragon processor, excessive communications resulted in a lack of scalability (i.e 128, 256 and 512 processors gave the same performance). The most efficient operating system on Intel (SUNMOS) was used as even worse performance was experienced on the widely-used Paragon OSF 1.2 operating system shown in the figure.
Significance: These first results of PVSOLVE on the IBM look very promising when compared to more expensive Cray C-90 and the poor Paragon performance. The measured scalable performance of the IBM SP-2 for large-scale applications such as the High-Speed Civil Transport, combined with the reasonable cost make it attractive for structural analysis application in the future.
Status/Plans: The versions of PVSOLVE on the Cray and IBM have been optimized while considerable additional optimization is planned on the SP-2. The additional optimization (using the new ESSL Power2 library and loop unrolling both ways to level 8) may further reduce the analysis times by as much as a factor of 3.
Point of Contact:
Return to the Table of Contents