J. K. Dewhurst

Dear Cuiju, That's correct: you take the square-root of the sum of squares of the real and imaginary parts. See Fig. 4 in Hughes and Sipe, Phys. Rev. B 53, 10751 (1996). Regards, Kay.

A dense k-point set is required to converge the non-linear optical response. Here is a converged second-order susceptibility of GaAs using a 56x56x56 k-point set: Compare this to Hughes and Sipe, Phys. Rev. B 53, 10751 (1996). Input file attached. (The conversion factor of electric field from atomic units to c.g.s-Gaussian units (e.s.u) is 1.71525554062 × 10^7 statV/cm.) Regards, Kay.

A dense k-point set is required to converge the non-linear optical response. Here is a converged second-order susceptibility of GaAs using a 56x56x56 k-point set: Compare this to Hughes and Sipe, Phys. Rev. B 53, 10751 (1996). Input file attached. (The conversion factor of electric field from atomic units to c.g.s-Gaussian units (e.s.u) is 1.71525554062 × 10^7 statV/cm.) Regards, Kay.

Elk version 10.7.8 released

Elk version 10.6.11 released

Hi All, I don't think there was a bug -- the DOS code hasn't changed in a long time. The IDOS is defined to be the TDOS minus the sum of the PDOS components. Regards, Kay.

Hybrid OpenMP+MPI is usually faster than MPI alone for large systems. For small systems, I think thread creation and cache coherency can cause too much overhead for nested OpenMP to be worth it. Hopefully this will improve as compilers, kernels and architectures advance. Elk uses its own thread management scheme which limits the number threads to the total number of cores. This seems to be more efficient than using OMP_DYNAMIC=true. If your calculations are for large systems, then the compute time...

Are you using -fcheck=all ? If so, then remove this option because it won't allow passing of unallocated arrays. You may be reaching the limit of 4-byte integers for addressing arrays. The maximum address is 2,147,483,647. For a real(8) array this would correspond to 16GB. You may also reach the limit of the size of arrays that MPI routines can transfer. We had this problem with BSE matrices and had to transfer them in smaller chunks. Some things you can try: a different compiler set lowq=.true.,...