Title: Performance Enhancements in MSC'Nastran for Large Scale Design Optimization on Cray SV1 Computers
1Performance Enhancements in MSC.Nastran for Large
Scale Design Optimization on Cray SV1 Computers
Dr. D. Obrist, Dr. H. Misra, Cray Inc. Dr. S.
Zhang, Dr. D. Chou, MSC.Software Corp.
2Large Scale Design Optimization
- joint project between MSC.Software Corp. and
Cray Inc. (October 2000 - July 2001)
? 2-4x shorter turnaround time
3Characteristics of Large Scale Optimization
Problems
- millions of degrees of freedom
- hundreds of design variables and responses
- hundreds of modes
? prohibitive turnaround times for simulations ?
excessively large I/O
4List of enhancements
- exploit the sparsity of the design model
- improved data management to process DSADJ in a
single pass - improved vectorization - highly optimized matrix-matrix multiplications
from the Cray Scientific Library - optimized sparse matrix I/O
- parallelization of DSADJ and DSVG1
- misc. improvements (GP5, EMG, MPYAD, PARTN,
MERGE, SADD5, etc.)
5Sparsity of the design model
In many design optimization tasks only a small
number of elements are modified during the design
process (sparse design set)
Example data recovery sub-dmap DISPRS
6Data recovery sub-dmap DISPRS
- Industry example
- design model is 25 sparse
- 2091102 DOF
- 251 modes
- 128 design variables
- 2931 retained responses
- amount of I/O is reduced by 4x
- CPU time is reduced by 5x
7Improved data management in DSADJ - single pass
single pass ? reduced scalar overhead
8Industry Example IDSADJ performance
7x improved!
9Industry Example IIDSADJ performance
9x improved!
10Industry Example IIIDSADJ performance
13x improved!
11Industry Example IVDSVG1 performance
10x improved!
12Parallelization of DSVG1 and DSADJ
Parallel runs of the Industry Example III (2
million DOF) with 1, 2, and 4 processors.
13Total Improvements over one Design Cycle
Overall improvement 2-4x !
14Conclusions
The turnaround time for a large design
optimization task is dramatically reduced (2-4x)
...
The performance is independent of the open core
memory size ...