| Home > Publications database > PHIDE: A Parallel Hybrid Direct–Iterative Eigensolver for Hermitian Eigenvalue Problems |
| Journal Article | FZJ-2026-01049 |
; ; ; ; ; ; ; ;
2026
IEEE
New York, NY
This record in other databases:
Please use a persistent id in citations: doi:10.1109/TPDS.2025.3623188
Abstract: In this paper, we propose a Parallel Hybrid Direct–Iterative Eigensolver for Hermitian Eigenvalue Problems with-out tridiagonalization, denoted by PHIDE, which combines directand iterative methods. PHIDE first reduces a Hermitian matrixto banded form, then applies a spectrum slicing algorithm tothe banded matrix, and finally computes the eigenvectors of theoriginal matrix via backtransformation. Compared with conven-tional direct eigensolvers, PHIDE avoids tridiagonalization, whichinvolves many memory-bound operations. In PHIDE, the bandedeigenvalue problem is solved using the contour integral methodimplemented in FEAST, which may yield slightly lower accuracythan tridiagonalization-based approaches. For sequences of corre-lated Hermitian eigenvalue problems arising in density functionaltheory (DFT), PHIDE achieves an average speedup of $1.22×$ overthe state-of-the-art direct solver in ELPA when using 1024 pro-cesses. Numerical experiments are conducted on dense Hermitianmatrices from real applications as well as large sparse matricesfrom the SuiteSparse and ELSES collections.
|
The record appears in these collections: |