|
|
Blitz Support : |
From: Todd Veldhuizen (tveldhui_at_[hidden])
Date: 2004-02-26 13:16:56
Hi Hendrik,
It might help if you gave specifics. Can you show the code you use
to create the blitz arrays vs. the non-blitz version arrays?
e.g.
Array<3,TinyVector<double,3> > X(256,256,256) ?
try also outputting sizeof(TinyVector<double,3>) etc. if this is
what you are using. If the TinyVector is being padded (say from
sizeof(double)*3 = 24 bytes up to 32 bytes) this might explain
your problem.
Cheers,
Todd
On Thu, 26 Feb 2004, Hendrik Belitz wrote:
> Hi,
>
> after I have implemented a finite difference Solver for a biiig PDE, I
> encountered the following problem.
>
> When I work on large arrays ( 256^3, containing 3D vectors, and I need 4 of
> them), my system begins to swap out memory. Interestingly, the swapping also
> seems to slow down my computations a lot (I measured a decrease about 30% for
> the pure process run time). I also implemented the same solver without blitz.
> That version normally is about 10% slower than the blitz version, but does
> not suffer from the performance decrease for large data sets.
>
> Can somebody explain me this rather strange behaviour?
>
> PS:
> I'm working with GCC 3.3, SuSE Linux 8.2 on a P4 2,6 GhZ, 1 GByte Ram
>
> ----------------------------------------------------------------
> Dipl.-Inform. Hendrik Belitz
> Zentralinstitut für Elektronik
> Forschungszentrum Jülich GmbH
> D-52428 Jülich, Germany
> Tel.: (++49)2461 61 4578
> Fax: (++49)2461 61 3990
> email: h.belitz_at_[hidden]
>
>
> _______________________________________________
> Blitz-support mailing list
> Blitz-support_at_[hidden]
> http://www.oonumerics.org/mailman/listinfo.cgi/blitz-support
>
-- Todd Veldhuizen / tveldhui_at_[hidden] / Indiana University Computer Science