A Trade-off Analysis of the Parallel Hybrid SPIKE Preconditioner in a Unique Multi-core Computer
In this paper we apply the parallel hybrid SPIKE algorithm as a preconditioner for a nonstationary iterative method to solve large sparse linear systems. In order to obtain a good preconditioner, we employ several strategies solving combinatorial problems such as matching, reordering, partitioning, and quadratic knapsack. Our SPIKE implementation combines MPI and OpenMP paradigms in a unique multi-core computer. The computational experiments show the influence of each strategy evaluating the number of iterations and CPU time of the iterative solver in a set of large systems from miscellaneous application areas. The experiments suggest that the SPIKE preconditioner is very advantageous when a suitable set of parameters is chosen. The choice of the number of MPI ranks and OpenMP threads is not an easy task, because the SPIKE algorithm can increase the number of iterations when the number of MPI ranks grows. Moreover, the increase in the number of threads does not ensure a better performance.
KeywordsParallel hybrid SPIKE preconditioner Combinatorial strategies Nonstationary iterative method
This work has been supported in part by CNPq, CAPES, and FAPES.
- 3.Cuthill, E., McKee, J.: Reducing the bandwidth of sparse symmetric matrices. In: Proceedings of the 1969 24th National Conference, pp. 157–172. ACM, New York (1969). http://doi.acm.org/10.1145/800195.805928
- 4.Davis, T.A., Hu, Y.: The university of Florida sparse matrix collection. ACM Trans. Math. Softw. 38(1), 1:1–1:25 (2011). http://www.cise.ufl.edu/research/sparse/matrices
- 7.Hestenes, M.R., Stiefel, E.: Methods of conjugate gradients for solving linear systems, vol. 49. NBS (1952)Google Scholar
- 8.Kouris, A., Sobczyk, A., Venetis, I., Gallopoulos, E., Sameh, A.: Revisiting the SPIKE-based framework for GPU banded solvers: a givens rotation approach for tridiagonal systems in CUDA (2014)Google Scholar
- 9.Li, A., Deshmukh, O., Serban, R., Negrut, D.: A comparison of the performance of SPIKE: GPU and intel’s math kernel library (MKL) for solving dense banded linear systems (2014)Google Scholar