[petsc-dev] P3DFFT is an open-source numerical library providing highly scalable implementation of 3D spectral transforms

[Barry Smith]

On Aug 20, 2014, at 4:08 PM, Matthew Knepley <knepley at gmail.com> wrote:

> On Wed, Aug 20, 2014 at 4:05 PM, Barry Smith <bsmith at mcs.anl.gov> wrote:
> 
> On Aug 20, 2014, at 4:00 PM, Matthew Knepley <knepley at gmail.com> wrote:
> 
> > On Wed, Aug 20, 2014 at 3:52 PM, Barry Smith <bsmith at mcs.anl.gov> wrote:
> >
> > On Aug 20, 2014, at 3:34 PM, Matthew Knepley <knepley at gmail.com> wrote:
> >
> > > The blurb should also answer the question, Why do we need another FFT library?
> >
> >   Because it actually does multidimensional FFTs in parallel?
> >
> > For something this old and established, it might be important to say that it does blah
> > blah blah that Spiral, FFTW, etc.
> >
> >   In theory, with this one could write very efficient parallel 3d Poisson solvers in PETSc for boxes, which is an important special case that PETSc does not currently support.
> >
> > Wouldn't you just use MG?
> 
>    FFT when done properly is much faster!  Yes it is a specialized case but an important one.
> 
> Are we calculating the following way:
> 
>   1) MF MG for the Laplacian is about 50 F/dof
> 
>   2) An FFT is about 6 F/dof, and we use 1 transform, 1 divide, and 1 inverse transform, so we get 13 F/dof
> 
> so you get 3x or so?

   My statement is based on experience, not models. I stand by it and yes it can be factors of that magnitude.  Now one could argue 3 times faster so what, but if you are doing this solve millions of times and it is the most time consuming part of the simulation (by far) then it adds up. 
> 
>   Matt
>  
> > I have other uses for FFT, including DFT codes where it makes a lot of sense.
> >
> >    Matt
> >
> >
> >   Barry
> >
> > >
> > >    Matt
> > >
> > >
> > > On Wed, Aug 20, 2014 at 1:53 PM, Barry Smith <bsmith at mcs.anl.gov> wrote:
> > >
> > >
> > > From: Dmitry Pekurovsky dmitry at sdsc.edu
> > > Date: August 12, 2014
> > > Subject: Library for spectral transforms in 3D for parallel machines
> > >
> > > P3DFFT is an open-source numerical library providing highly
> > > scalable implementation of 3D spectral transforms namely Fast
> > > Fourier Transform (FFT), with an option to combine it with
> > > cosine/sine/Chebyshev/empty transform in the third dimension. (The
> > > empty transform allows the user to substitute their own custom
> > > transform in the third dimension. This can be useful in
> > > applications such as inhomogeneous wall bounded turbulence.) P3DFFT
> > > implements 2D domain decomposition which allows it to overcome a
> > > scalability restriction inherent in 1D decomposition. This approach
> > > has shown good scalability up to 131,072 cores.
> > >
> > > A new version of P3DFFT 2.7.1 is now available. The project Home
> > > Page is http://code/google.com/p/p3dfft where instructions for
> > > obtaining the source code are provided. Installation instructions
> > > and a User Guide are also available.
> > >
> > > P3DFFT features include real-to-complex and complex-to-real
> > > transforms, in-place transforms, pruned transforms (with less than
> > > full input or output), and multi-variable transforms. The package
> > > includes example programs in Fortran and C. This is a project in
> > > active development, with a user mailing list, a wiki page and a
> > > version control system. P3DFFT is considered community software and
> > > is being installed in public space at many supercomputer centers.
> > > Contributions and feedback from users are welcome.
> > >
> > >
> > >
> > > --
> > > What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.
> > > -- Norbert Wiener
> >
> >
> >
> >
> > --
> > What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.
> > -- Norbert Wiener
> 
> 
> 
> 
> -- 
> What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.
> -- Norbert Wiener