[gmx-users] [Fwd: [FFTW-announce] FFTW 3.2 is released]

[Carsten Kutzner]

On Nov 16, 2008, at 12:47 PM, Mark Abraham wrote:

> Apologies to people receiving this from the FFTW list too, but it  
> should be of interest to a large slice of people using GROMACS. In  
> particular, the improvements for SIMD and size of multiples of 5/10  
> should be of value for lots of people.
>
> Does anyone (Carsten?) have an off-the-cuff opinion on whether the  
> latter means that FFT grids with dimensions that are multiples of 5  
> might now be more desirable? How might that affect PME load-splitting?

Hi,

I fear that the performance gain in mdrun due to a faster FFT will not  
be too large - of course that
depends on how 'bad' the performance was before for multiples of 5.  
The reason is that a significant
amount of the time used for transforming to and from reciprocal space  
is actually communication
time: MPI transposes of the FFT data, while the calculation itself is  
not so time consuming for a
normal-sized MD system. Since these transposes are done in Gromacs  
anyway, the only thing that
could be faster with the new FFTW library is the calculation itself -  
which should not make a big
difference in the overall mdrun performance. But why not compare the  
old and new FFTW performance
with a test MD system?  I would choose zero PME nodes for that test,  
just to be shure that a
possible FFTW performance gain does not get hidden behind a PME/PP  
load imbalance.

Carsten

>
>
> Mark
>
> -------- Original Message --------
> Subject: [FFTW-announce] FFTW 3.2 is released
> Date: Sat, 15 Nov 2008 22:01:44 -0500 (EST)
> From: Steven G. Johnson <stevenj at ab-initio.mit.edu>
> Reply-To: Steven G. Johnson <stevenj at alum.mit.edu>
> To: fftw-announce at fftw.org
>
> Dear FFTW users,
>
> The official release of FFTW 3.2 is now available from the FFTW home  
> page:
>
> 	http://www.fftw.org/
>
> This release does *not* contain the experimental MPI transfroms from  
> FFTW
> 3.2alpha; as those routines are still in alpha, they have been moved  
> to FFTW
> 3.3alpha1 (also available from the FFTW web page).
>
> Compared to version 3.1.3 of FFTW, highlights of the FFTW 3.2  
> release include:
>
> * Significantly faster r2c/c2r transforms, especially on machines
>  (e.g. x86) with SIMD instruction.
>
> * Performance improvements for sizes with factors of 5 and 10.
>
> * Rewritten threading support for improved performance.
>
> * Support for the Cell processor, donated by IBM Research; see
>  README.Cell and the Cell section of the manual.
>
> * Support for MIPS paired-single SIMD instructions, donated by
>  Codesourcery.
>
> The complete release notes can be found in the NEWS and ChangeLog  
> files of the
> release, and at:
>
> 	http://fftw.org/release-notes
>
> Specific changes compared to FFTW 3.2alpha3 are listed below.
>
> Regards,
> Steven G. Johnson and Matteo Frigo
>
> FFTW 3.2 changes from 3.2alpha3
>
> * Worked around apparent glibc bug that leads to rare hangs when  
> freeing
>  semaphores.
>
> * Fixed segfault due to unaligned access in certain obscure problems
>  that use SSE and multiple threads.
>
> * MPI transforms not included, as they are still in alpha; the alpha
>  versions of the MPI transforms have been moved to FFTW 3.3alpha1.
>
> _______________________________________________
> FFTW-announce mailing list
> FFTW-announce at fftw.org
> http://fftw.org/cgi-bin/mailman/listinfo/fftw-announce
>
> _______________________________________________
> gmx-users mailing list    gmx-users at gromacs.org
> http://www.gromacs.org/mailman/listinfo/gmx-users
> Please search the archive at http://www.gromacs.org/search before  
> posting!
> Please don't post (un)subscribe requests to the list. Use the www  
> interface or send it to gmx-users-request at gromacs.org.
> Can't post? Read http://www.gromacs.org/mailing_lists/users.php

Carsten Kutzner
ckutzne at gwdg.de