Re: [boost] uBLAS formal review

From: Martin Weiser (weiser_at_[hidden])
Date: 2002-06-28 06:48:01

On Freitag, 28. Juni 2002 00:24, Joerg Walter wrote:
> ----- Original Message -----
> From: "Martin Weiser" <weiser_at_[hidden]>

> > 3. Large matrices (N=40,..,250). The operands don't fit any longer
> > into the 1st level cache. As expected, the lower memory bandwidth of
> > the 2nd level cache decreases the performance of both uBLAS and the
> > naive implementation. Due to blocked algorithms the Sun BLAS
> > maintains its high performance and wins by a factor of about 7. The
> > great surprise is, that the naive  implementation suffers less than
> > uBLAS and wins by a factor of about 1.7. Currently I've no
> > explanation for that.
>
> We're switching from indexing to iterating access at N = 32. You could
> play with NUMERICS_ITERATOR_THRESHOLD to see the impact on your
> platform.

Thanks for the hint. GCC/UltraSparc seems to prefer indexing. The 
performance drop attributed to the L1->L2 transition seems to be caused 
by the indexing to iteration switch. Incidentally it happened at the same 
matrix size  where the L1->L2 transition must in fact be placed.

For the threshold at 1024, the new graph is available at 
http://www.zib.de/weiser/ublas_review-1024.gif .

As a surprising side effect, the change seems to affect (improve) the 
performance for small matrices (N<<32), too.

Yours,
Martin

-- 
Dr. Martin Weiser            Zuse Institute Berlin
weiser_at_[hidden]                Scientific Computing
http://www.zib.de/weiser     Numerical Analysis and Modelling