At 03:49 PM 4/22/2009, you wrote:
On Wed, Apr 22, 2009 at 1:27 PM, Dan Nelson <dnel...@allantgroup.com> wrote:
> In the last episode (Apr 22), Rod Heyd said:
> > I've been running mysql on a T1000 (Ultrasparc T1) system for several
> > years now, and while I've been happy with the performance overall, the
> > poor floating point capability on these systems has been a
> disappointment.
> > Recently, I got my hands on a Sunfire T5420 system and I've been
> comparing
> > the floating point performance on this system with the T1000, and was
> > expecting to see a significant improvement in floating point by virtue of
> > the fact that the Ultrasparc T2 processor has 1 FPU per core as opposed
> to
> > 1 FPU per CPU. However, I'm only seeing a marginal improvement for
> > floating point calculations, for example, a "select benchmark(100000000,
> > 1.0 + 2.0)" takes roughly 50 seconds to run on both of these systems,
> > however I was expecting a lot better performance from the T2 processor.
>
> If you are running just one command, then you are only using one of the 8
> FPUs on the T2. Try comparing 8 parallel "select benchmark(100000000, 1.0
> +
> 2.0)" runs at once on each server.
>
Hi Dan,
Yes, actually, I already know that parallel performance will be much
improved, however, I was expecting more improvement on single threads as
well, since the specs say that it takes 40 clock cycles just to access the
FPU on the T1, but something like 6 clock cycles on the T2. So just from
that perspective it seems like there should be a significant improvement for
single threads, not just parallel performance. At least that's the way I
read the docs from Sun on this. At any rate, my expectations here are
clearly wrong, and I guess I'd just like a better understanding of why I'm
getting it wrong.
Rod,
Have you noticed any significant improvement with floating point
calculations with *any* software running on the Sunfire, not just MySQL?
For example, can you run some 3rd party benchmark to confirm that the FPU
will operate faster compared to your Ultrasparc? If you can't verify this,
then there is no point blaming MySQL. If it does show a significant
improvement, then maybe MySQL has to be recompiled to take advantage of the
faster FPU?
Mike
--
MySQL General Mailing List
For list archives: http://lists.mysql.com/mysql
To unsubscribe: http://lists.mysql.com/mysql?unsub=arch...@jab.org
