Might be worth creating a plan with FFTW.MEASURE or FFTW.EXHAUSTIVE. See help 
for `plan_fft`.

FYI, this isn't really a "Julia issue" since Julia is just calling the FFTW 
library.

Best,
--Tim

On Sunday, December 20, 2015 03:40:41 PM Pontus Stenetorp wrote:
> On 19 December 2015 at 20:02,  <[email protected]> wrote:
> >> As can be seen, all the 8 cores of my pc is being used by the julia
> >> program, however, only 35% of the system resource is covered. My julia
> >> code
> >> mainly does the fft operates in a loop. I use fftw.set_num_threads(8)
> >> outside the loop.  In my opinion, the top command should exhibit nearly
> >> 100% use of the system resource, but that's not true.
> > 
> > I'm not quite sure where you get 35% from, as I read your `top` the top
> > three users are Julia using 6.89 cpus, QQ.exe 0.45 cpu and X 0.11 cpu.
> > Julia is getting the vast majority of the cpu resources, nearly 7 out of 8
> > cores.
> 
> He is reading time spent in user space which is given as 35.0%, the
> problem however is that 60.6% of the time is spent on kernel processes
> which is a really bad sign.  Without seeing the code it is difficult
> to know what is going on, especially since I am not familiar with
> FFTW.  My best guess is that the job is I/O bound or that the data
> given to FFTW is so small that the thread overhead is eating up
> performance.  For the latter, set the number of threads to one and see
> if the performance improves.
> 
>     Pontus

Reply via email to