Scalable Block-Tridiagonal Eigensolvers in the Context of Electronic Structure Calculations

Alejandro Lamas Davinã, Xavier Cartoixà, José E. Román

Research output: Contribution to journalArticleResearchpeer-review

Abstract

© 2018 The authors and IOS Press. In computer-based numerical simulations, some methods to determine the electronic and optical properties of semiconductor nanostructures, require computing the energies that correspond to the interior eigenvalues of a Hamiltonian matrix. We study the case in which the Schrödinger equation is expanded into a matrix that has a block-tridiagonal structure. Additionally, the matrix can have two extra non zero blocks in the corners due to periodic boundary conditions. Given that not the whole eigenspectrum is required, we choose to use projection methods to compute the necessary set of eigenvalues. The shift-and-invert Lanczos method requires to solve a linear system at each iteration. We have developed a parallel code that improves the scalability of this step by exploiting the block structure. Results show that, to solve these specific cases, this method offers better scalability when compared to a general-purpose solver such as MUMPS.
Original languageEnglish
Pages (from-to)117-126
JournalAdvances in Parallel Computing
Volume32
DOIs
Publication statusPublished - 1 Jan 2018

Keywords

  • Electronic structure calculation
  • Schur complement
  • cyclic reduction
  • eigenvalue problem
  • shift-and-invert

Fingerprint

Dive into the research topics of 'Scalable Block-Tridiagonal Eigensolvers in the Context of Electronic Structure Calculations'. Together they form a unique fingerprint.

Cite this