Re: [eigen] benchmarks for large matrices?

[David Roundy <daveroundy@xxxxxxxxx>]

On Wed, Feb 18, 2009 at 5:08 AM, Benoit Jacob <jacob.benoit.1@xxxxxxxxx> wrote:
>
> 2009/2/18 Gael Guennebaud <gael.guennebaud@xxxxxxxxx>:
> > I have to say, those poor results of ATLAS puzzle me because all
> > benchmarks I can see show that ATLAS is close to MKL...
>
> From what I read, ATLAS is able to "tune" itself for the host machine?
> I.e. tune the optimal value of internal parameters such as cache size,
> etc? Could it be that the package that you used was just tuned for a
> different machine than yours, hence not optimal on yours?
>
> I just mention that as a possible explanation for discrepancy in
> benchmarks, but even so, I believe that our benchmarking is perfectly
> fair, because we aim Eigen to be useful for widely deployed software,
> where there can be no question of tuning for a specific machine (we
> still allow to control some internal parameters such as expected cache
> size). By contrast, if what I understood of ATLAS's tuning is correct,
> it is really aimed at university computing servers, not at wide
> deployment.

I would be interested to see how ATLAS shifts the benchmarks if you
compile it on your computer.  If you're using an ATLAS tuned for a
machine with a larger cache, it'd be no surprise that you'd get poor
numbers...

That said, I don't generally recompile ATLAS on my machines, so your
benchmark may reflect my experience.  On the other hand, it may be
that my machines happen to be similar to those that my ATLAS is
configured for, in which case I may be getting better performance.
Which is why it'd be nice to know the ATLAS best-case performance
(i.e. tuned for your machine).
--
David Roundy