next up previous contents
Next: Execution Up: Programs for running an Previous: Input

    
LAPW1 (generates eigenvalues and eigenvectors)

lapw1 sets up the Hamiltonian and the overlap matrix (Koelling and Arbman 75) and finds by diagonalization eigenvalues and eigenvectors which are written to case.vector. If the file case.vns exists (i.e. non-spherical terms in the potential), a full-potential calculation is performed.

For structures without inversion symmetry, where the hamilton and overlap matrix elements are complex numbers, the corresponding program version lapw1c must be used in connection with lapw2c.

Since usually the diagonalization is the most time consuming part of the calculations, several options exist here. In we include highly optimized modifications of LAPACK routines, which, depending on your hardware, may speed up the code significantly (e.g. on SGI R10000 processores a speedup by a factor of 2-3 can be achieved with respect to the ``standard'' LAPACK). We call all these routines ``full diagonalization'', since we provide also an option to do an ``iterative diagonalization'' using a block-Davidson method (see Singh 94). This scheme starts from an old eigenvector (previous scf-iteration), needs additional memory and produces only approximate eigenvalues/vectors, but can be significantly faster than LAPACK, in particular if the ratio of matrix size to number of relevant(e.g. occupied) eigenvalues is large. In any case, it is recommended that convergence is checked by another scf-iteration with ``full diagonalization''. Often the best performance can be obtained by first running to a crude convergence, next doing a ``full diagonalization'' and finally continuing iteratively to self consistency. (use first: run_lapw -it -fc 10 and then run_lapw -it -fc 1)

Parallel execution (on the k-point level) is also possible and is described in detail in Sec. 5.4.



 
next up previous contents
Next: Execution Up: Programs for running an Previous: Input

2000-04-11