An a posteriori verification method for generalized real-symmetric eigenvalue problems in large-scale electronic state calculations (Q1987430)

The physical background of the paper is computational material science. Specifically, the authors want to study the solution of fundamental Schrödinger-type equations for the determination of the energy of electrons and the wave functions in order to characterize the properties of the material. The numerical solution of the treated linear partial differential equation gives a generalized eigenvalue problem (GEP) \(A x_k = {\lambda}_k B x_k\) under the orthogonality condition \({x_i}^T B {x_j} = 1, \text{ if } i = j, 0 \text{ otherwise}\), where both \(A\) and \(B\) are real symmetric \(n \times n\) matrices, with \(B\) being positive definite, and assuming \({\lambda}_1 \leq {\lambda}_2 \leq \ldots \leq {\lambda}_n\). An eigenvalue represents the energy of an electron in the material and the corresponding eigenvector the wave function. The derivation of the GEP for the linear Schrödinger equation is presented in the appendix with references to textbooks. The matrix dimension is approximately proportional to the number of molecules, atoms or electrons in the different materials. Thus, large-scale generalized eigenvalue problems have to be solved using massively parallel supercomputers for industrial applications. The eigenvalues of these problems are mostly densely clustered or almost degenerate with difficulties to distinguish them. The difference of sequential eigenvalues tends to be proportional to \(\frac{1}{n}\). Furthermore, in order to compute efficiently such large-scale calculations, lower-precision arithmetic, such as single-precision or even half-precision arithmetic, are used. Many well-known solvers for the generalized eigenvalue problem are parallelized and show difficulties to overcome the mentioned problems. Thus, the authors develop an a posteriori verification procedure to give an impression of the difference between the approximate solution of a GEP-solver and the exact solution. In order to verify the algorithm, the authors consider the following four parts of the GEP: (i) the Cholesky decomposition of \(B\), (ii) the reduction to the standard eigenvalue problem (SEP), (iii) the solution of the SEP, (iv) the transformation of the eigenvectors. The steps (I), (ii), and (iv) are referred as reducer, and step (iii) is the SEP solver. The authors point to the standard parallel numerical library ScaLAPACK, developed in the 1990s, that exhibit severe bottlenecks on modern massively parallel supercomputers. The authors present results of the created verification procedure using the corresponding ScaLATACK routines as eigenvalue solver on the K computer, a Japanese flagship supercomputer. But, they mention that there are also novel solver libraries of ELPA (developed in Europe) and EigenExa (developed in Japan) which should overcome the mentioned bottlenecks of ScaLAPACK. A schematic diagram of hybrid workflows is presented for a future version, which also includes the new SEP-solvers. The a posteriori verification method is based on Wilkinson's bound for the standard eigenvalue problem, the verification of the error bounds of all computed eigenvalues, Gershgorin's circle theorem, and a variant of Yamamoto's theorem. The corresponding code is demonstrated in detail. The verifier procedure uses primarily matrix multiplications. Thus, the computational time of the verification algorithm is moderate in comparison to the solver procedure with the more time-consuming Cholesky decomposition and the tridiagonalization. The test data stem from the ELSES matrix library, mainly for systems of organic polymers. The largest matrix problem belongs to a nano-composite carbon solid with a matrix dimension of \(n = 430.080\) using 6400 processor nodes. Tables and Figures show that the verification algorithm delivers the intervals that contain the exact eigenvalues and show the computational times and number of used processor nodes. Future issues are announced for the use of the verification technique if lower-precision arithmetic is applied as an initial guess, in order to compute the large-scale problems efficiently, and refinements in a mixed-precision calculation are needed.