PEXSI

## General questions

Q: How do I know whether PEXSI works for my application?

A: PEXSI may not necessarily be faster than the diagonalization method or other competitive methods. The simplest way to see whether PEXSI brings acceleration for your applications is to use PEXSI to compute the selected elements of the inverse for a typical matrix from your applications. See pagePselinvReal and driver_pselinv_real.c for how to do this.

Q: Can I just use the selected inversion routine?

A: The parallel selected inversion (PSelInv) is a standalone routine. See pagePselinvReal for an example.

Q: Does PEXSI accelerate dense matrix computation?

A: No. The acceleration is based on the sparsity of the LU factor or the Cholesky factor. PEXSI should not be fast if the matrix is dense or nearly dense.

Q: Does PEXSI work for asymmetric matrices?

A: Currently PEXSI only works for (real or complex) symmetric matrices. We plan to support asymmetric matrices in the future. If you have applications in mind and PEXSI may bring significant acceleration to your application, please contact us as in Troubleshooting, with some description of your matrix and application.

## Installation

A: We have observed that when ParMETIS/PT-Scotch is used, the number of processors for symbolic factorization (npSymbFact) cannot be larger than a magic number depending on the matrix and the machine. This may to be related to the parallel symbolic factorization routine in SuperLU_DIST. If this problem happens, try to reduce npSymbFact to a small number (such as 4 or 16), or even try to use the sequential symbolic factorization if feasible.