Block diagonal decompositions for parallel computations of large power systems. Final report
- Santa Clara Univ., CA (United States)
In this report we present the algorithm and C code for balanced bordered block diagonal (BBD) decompositions of large sparse matrices, as well as a variety of experimental results relating to the algorithm`s performance. The software has been tested on a number of large matrices, including models of the West Coast power network (1,993 x 1,993 matrix). The algorithm was found to compare very well with the symmetric minimal degree ordering in terms of sparsity preservation-in the test cases considered, the BBD decomposition produced only up to 15% more fill-in. This is more than satisfactory considering that BBD structures are far better suited for parallel computing than the scattered and unpredictable element patterns obtained by minimal degree ordering. For some denser matrices, the BBD decomposition was actually seen to produce lower fill-in than the minimal degree. In applications to power systems the execution time for the BBD decomposition was found to have a quadratic upper bound on its complexity, which is comparable to a number of other sparse matrix orderings. Simulation results indicate that the actual execution time is similar to the execution time of the symmetric minimal degree ordering in Matlab 4.0. The special structural advantages of balanced BBD decompositions have been utilized to parallelize the process of LU factorization. The speedups obtained with respect to solutions using symmetric minimal degree ordering on a single processor have confirmed the significant potential of BBD decomposition in parallel computing. For the 1,993 bus power system, a speedup of 11.2 times was obtained using 14 processors on a PVM 2.4.
- Research Organization:
- Electric Power Research Inst., Palo Alto, CA (United States); Santa Clara Univ., CA (United States)
- Sponsoring Organization:
- Electric Power Research Inst., Palo Alto, CA (United States)
- OSTI ID:
- 88612
- Report Number(s):
- EPRI-TR--105083
- Country of Publication:
- United States
- Language:
- English
Similar Records
A balanced decomposition algorithm for parallel solutions of very large sparse systems
A parallel Gauss-Seidel algorithm for sparse power system matrices