[CPMD-list] NCACHE size

[Philippe Blaise]

Hello,

just to say that a value of ncache = 4 * 1024 is optimal on my opteron 
machine,
but the difference between the default value 2 * 1024 and 4 * 1024 is 
below 10% in time.
More generally, it seems to me that when you play with the ncache value, 
you don't always obtain
huge differences, it's probably due to compiler optimizations and large 
sizes of nowaday caches -
so, does the ncache value effect was larger some years ago ?
Anyway, the fft that comes with cpmd is very good, and I would be 
suprised if you
obtain better performances with another one  - for example on a NEC 
vector machine and
on a alpha superscalar machine, the vendor implementations are not better.

Good luck,

  Philippe Blaise
 

Axel Kohlmeyer wrote:

>>>>"IK" == I Kozin <Kozin> writes:
>>>>        
>>>>
>
>hello,
>
>IK> Hello,
>IK> I'd appreciate to hear any comments on varing NCACHE size 
>IK> (as given in mltfft.F) on various platforms.
>IK> Particularly Xeon, Itanium2, Opteron, Power4,5, NEC sx6.
>
>IK> In the file above NCACHE is given as 1024*N where N varies.
>IK> So I'd guess it has to do with L1 cache.
>IK> For any i386 NCACHE=1024*10.
>IK> Xeon's L1 cache is 8 KB. So should N be 8?
>
>well. i don't know whether there was historically an explicit 
>relation to the L1 cache. but due to way modern cpus and compilers 
>work, (e.g. by automatic prefetching) there is none anymore.
>for instance the i386 value came about by running a series of
>calculations with different values for NCACHE and picking the
>value that gave the on average best performance. since the various
>x86 platforms have quite different characteristics internally,
>this always has to be a compromise. if you want to squeeze out
>the last bit of performance, you may want to tune it for your
>specific machine and your specific example. 
>
>btw: the best way to optimize the fft is use it as little as 
>possible (e.g. by using the REAL SPACE WFN KEEP keyword, provided
>there is enough memory on your machine).
>
>IK> Itanium2: NCACHE=1024*8 but L1 cache size is 16 KB.
>IK> This is the only machine I've experimented with so far.
>IK> Taking N = 8, 16, 32 I found that N = 16 is marginally quicker
>IK> (wat32 benchmark, run 1: 1372 s, 1363 s, 1388 s).
>IK> So basically no need to bother.
>
>you should also try 'in-between' values.
>
>IK> More interesting cases:
>IK> Itanium2 + HPUX: NCACHE=1024*64 
>IK> (obviously the L1 cache is the same as above)
>
>IK> Operon: NCACHE=1024*2 (default) but L1 cache is 64 KB.
>
>this has not been tuned yet. if you find an optimal value,
>please let us know.
>
>IK> BTW, are there any efforts in attempting to use ACML or MKL
>IK> on AMD and Intel respectively for FFT instead of default FFT?
>
>not that i know of. feel free to try. the default fft is quite
>competitive, considering that it is portable fortran code.
>IMO it would be more interesting to see, how well a port 
>to fftw3 performs.
>
>regards,
>        axel kohlmeyer.
>
>IK> Thanks,
>
>IK> Igor Kozin
>IK> Computational Science & Engineering Dept.
>IK> CCLRC Daresbury Laboratory
>IK> Keckwick Lane
>IK> Warrington
>IK> WA4 4AD
>IK> UK
>
>IK> i. kozin at dl.ac.uk
>IK> +44 (0) 1925 603308
>IK> http://www.cse.clrc.ac.uk/disco
>IK> _______________________________________________
>IK> CPMD-list mailing list
>IK> CPMD-list at cpmd.org
>IK> http://cpmd.org/mailman/listinfo/cpmd-list
>
>
>
>--
>
>=======================================================================
>Axel Kohlmeyer       e-mail: axel.kohlmeyer at theochem.ruhr-uni-bochum.de
>Lehrstuhl fuer Theoretische Chemie          Phone: ++49 (0)234/32-26673
>Ruhr-Universitaet Bochum - NC 03/53         Fax:   ++49 (0)234/32-14045
>D-44780 Bochum  http://www.theochem.ruhr-uni-bochum.de/~axel.kohlmeyer/
>=======================================================================
>If you make something idiot-proof, the universe creates a better idiot.
>_______________________________________________
>CPMD-list mailing list
>CPMD-list at cpmd.org
>http://cpmd.org/mailman/listinfo/cpmd-list
>
>  
>