[Pw_forum] something wrong with vc-relax & relax on my clusters

[Paolo Giannozzi]

Axel Kohlmeyer wrote:

> btw: don't worry about MKL's FFT not being detected.
> there is no direct support anyways and both the fftw3
> and the fftw2 wrappers don't work reliably.
> in all my tests it turned out to be slower than the 
> bundled FFTW, except for power of two grids (which confirms
> results with BenchFFT). more importantly, you have to make
> certain that the OMP_NUM_THREADS environment variable is set 
> to 1 or else it will multi-thread across all local cpu cores,
> regardless of whether you parallelize with MPI (see the MKL
> documentation for details).

in serial execution (which uses 3d FFT) with no threading,
I notice a sizable speedup of MKL (FFTW3 interface) wrt
bundled FFTW. I instead never noticed any improvement in
FFTW v.3 wrt either FFTW v.2 or the bundled FFTW.

Disclaimer: there are few things I dislike more, and trust less,
than benchmarking.

Paolo
-- 
Paolo Giannozzi, Democritos and University of Udine, Italy