Sparsity-oriented sparse solver design for circuit simulation

The sparse solver is a critical component in circuit simulators. The widely used solver KLU is based on a pure column-level algorithm. In this paper, we point out that KLU is not always the best algorithm for circuit matrices by experiments. We also demonstrate that the optimal algorithm strongly depends on the sparsity of the matrix. Two sparse LU factorization algorithms are proposed for extremely sparse matrices and dense matrices. A simple but effective strategy is proposed to select the optimal algorithm according to the sparsity. By combining the two new algorithms and the selection method together, the proposed solver achieves much higher performance than both KLU and PARDISO.