Hi,

I wish to run Matlab on a Sun Solaris machine.
This application will be single threaded.
Problems will include matrix multiplication/ inversion and SVD's.


Which single processor Sun machine should I choose to equal the processing
performance of a P4 3.2Ghz Machine?

Thanks,

Mark

On Wed, 4 Feb 2004, Mark wrote:

> Which single processor Sun machine should I choose to equal the processing
> performance of a P4 3.2Ghz Machine?

None of them. Have a look at the Sun Blade 1500.

Now, which P4 3.2 GHz machine can make full use of 8 GB of RAM?

--
Rich Teer, SCNA, SCSA

President,
Rite Online Inc.

Voice: +1 (250) 979-1638
URL: http://www.rite-online.net

Rich Teer wrote:
>
> On Wed, 4 Feb 2004, Mark wrote:
>
> > Which single processor Sun machine should I choose to equal the processing
> > performance of a P4 3.2Ghz Machine?
>
> None of them. Have a look at the Sun Blade 1500.
>
> Now, which P4 3.2 GHz machine can make full use of 8 GB of RAM?
>

MATLAB 5.3.1 works as a true 64 bit application on Solaris, but MATLAB
6.5 does not. The next release of MATLAB will on some platforms, but
Solaris has not yet been declared to be one of them, so proceed with
some caution.

Having said that, Solaris seems most likely as the next choice of 64 bit
MATLAB platforms.

Choose a machine with a good amount of cache memory. The Sun Fire V100
I just bought is predictably slow with MATLAB when crunching large
matrices. My SGI Octane with its 250 MHz R10k does better, and it is
half the speed of the latest offerings from SGI. The reason the V100 is
slow on my MATLAB work is because it has only 256 k of cache, which is
rather pathetic, but then I knew that when I bought it.

Drop me a line if you want some more details.

--
Dr Tristram J. Scott
Energy Consultant

Mark wrote:

> Hi,
>
> I wish to run Matlab on a Sun Solaris machine.
> This application will be single threaded.
> Problems will include matrix multiplication/ inversion and SVD's.
>
>
> Which single processor Sun machine should I choose to equal the processing
> performance of a P4 3.2Ghz Machine?
>
> Thanks,
>

V65x with a single CPU installed?

P.

Tristram Scott <tristram.scott@ntlworld.com> writes:

> Choose a machine with a good amount of cache memory. The Sun Fire V100
> I just bought is predictably slow with MATLAB when crunching large
> matrices. My SGI Octane with its 250 MHz R10k does better, and it is
> half the speed of the latest offerings from SGI. The reason the V100 is
> slow on my MATLAB work is because it has only 256 k of cache, which is
> rather pathetic, but then I knew that when I bought it.

I'm not sure size of cache is quite enough to choose - what really
matters for matlab, surely, is overall floating-point throughput,
right? The new US-IIIi has less cache than US-III (which has 8MB), but
it moves the cache and the memory controller on-chip, so it's approx
neck and neck with the older chip on specFP at a lower price point. I
would suggest Sun Blade 2500 dual-cpu as a serious contender.

I hope to be benchmarking one of those later today as it happens

Chris
--
Chris Morgan
"Post posting of policy changes by the boss will result in
real rule revisions that are irreversible"

- anonymous correspondent

Rich Teer wrote:

> On Wed, 4 Feb 2004, Mark wrote:
>
>
>>Which single processor Sun machine should I choose to equal the processing
>>performance of a P4 3.2Ghz Machine?
>
>
> None of them. Have a look at the Sun Blade 1500.
>
> Now, which P4 3.2 GHz machine can make full use of 8 GB of RAM?

The Sun Blade 1500 is limited to 4GB according to Sun's product
website. So I'm not sure I get your point.

Unfortunately, for the original poster, you won't find a single
CPU machine from Sun's current product line that can match
the 3.2Ghz P4 except on the most rare of workloads.

This webpage has some Matlab benchmark results, and a 750Mhz
US-III (V880) is bested by a 1.5Ghz P-IV. Take a 1.28Ghz US-IIIi
and a 3.2Ghz P-IV and you'll likely see an even wider gap.

http://www-math.cudenver.edu/~jmande...benchmark.html

BTW, Wolfram is now shipping Mathematica for Linux on the AMD64
platform. Price/performance wise I don't think you can beat
that combo. Maybe Mathworks will come around now that
Itanium is effectively orphaned:

http://news.com.com/2100-1006_3-5150...?tag=nefd_lede


-cjs

Chris Morgan wrote:


> I'm not sure size of cache is quite enough to choose - what really
> matters for matlab, surely, is overall floating-point throughput,
> right? The new US-IIIi has less cache than US-III (which has 8MB), but
> it moves the cache and the memory controller on-chip, so it's approx
> neck and neck with the older chip on specFP at a lower price point. I
> would suggest Sun Blade 2500 dual-cpu as a serious contender.
>
> I hope to be benchmarking one of those later today as it happens
>
> Chris

Chris, please followup with your impressions and results. I'm
sure a lot of us are anxious to hear how the 2500 compares with the
1000/2000 workstations.

-cjs

What exactly does mathematica do?
I seem to remember the name from when I was in high school.
Is it the same company?

Vu.

"Colin Stuckless" <colinsRE@MOVEstuckless.com> wrote in message
news:cafUb.8684$bp1.460884@news20.bellglobal.com.. .
> Rich Teer wrote:
>
> BTW, Wolfram is now shipping Mathematica for Linux on the AMD64
> platform. Price/performance wise I don't think you can beat
> that combo. Maybe Mathworks will come around now that
> Itanium is effectively orphaned:
>
> http://news.com.com/2100-1006_3-5150...?tag=nefd_lede
>
>
> -cjs
>

Chris Morgan wrote:
>
> I'm not sure size of cache is quite enough to choose - what really
> matters for matlab, surely, is overall floating-point throughput,
> right? The new US-IIIi has less cache than US-III (which has 8MB), but
> it moves the cache and the memory controller on-chip, so it's approx
> neck and neck with the older chip on specFP at a lower price point. I
> would suggest Sun Blade 2500 dual-cpu as a serious contender.
>

No, size of cahce is not the only thing, but I have found with my own
use of MATLAB that it makes a significant difference when working with
large matrices. If they fit in cache memory, things fly along,
otherwise, maybe not. The specFP has never been much use for me as an
indicator of how my own code will perform. The MATLAB bench function is
somewhat better as an indication of MATLAB performance (as you would
hope), but it is a bit tame compared to some of the nasty things I like
to do to it.

As others have mentioned, you will find it hard to beat the P4 3.2Ghz
undr happy conditions. But if you start to annoy the machine, swapping
would be a good example, then I have always found the Windows machines
fall way behind. My laptop (P4 1.8 GHz) runs my own code at about twice
the rate of my SGI Octane (dual R10k, 250 MHz) up until it starts to be
low on RAM. At that point the Octane is in the order of maybe five
times as fast. I remember benchmarking a PC against an Origin 200 a
couple of years back. The machines had the same amount of RAM. I ran a
MATLAB job which took most of the RAM, and it was about two minutes on
either machine. Then I ran one twice as large, so it had to swap. The
Origin took four minutes. The PC took ten minutes.

One of the best things about MATLAB is that you don't need to change you
code to handle larger problem sizes. My application works on a 256MB
laptop if I keep the length of the simulations short. The exact same
application will consume 8GB of RAM on my Octane for large simulations.

If you want raw speed with MATLAB, and you know for sure you are not
going to step beyond the capacity of the machine, then perhaps aim
towards Linux on one of the AMD64 or Itanium which The MathWorks seem
fond of. But personally I would much rather Solaris or IRIX. Perhaps
slightly slower, but certainly more consistent.

--
Dr Tristram J. Scott | 44 Montague Road
Energy Consultant | Cambridge, CB4 1BX
| England
email tristram.scott@ntlworld.com | ph (+44 1223) 526 255

Colin Stuckless <colinsRE@MOVEstuckless.com> writes:

> Chris Morgan wrote:
>
>
> > I'm not sure size of cache is quite enough to choose - what really
> > matters for matlab, surely, is overall floating-point throughput,
> > right? The new US-IIIi has less cache than US-III (which has 8MB), but
> > it moves the cache and the memory controller on-chip, so it's approx
> > neck and neck with the older chip on specFP at a lower price point. I
> > would suggest Sun Blade 2500 dual-cpu as a serious contender.
> > I hope to be benchmarking one of those later today as it happens
> > Chris
>
> Chris, please followup with your impressions and results. I'm
> sure a lot of us are anxious to hear how the 2500 compares with the
> 1000/2000 workstations.

I will do, but unfortunately this machine is so "late" from my own
personal viewpoint that I wont be able to spend company time on giving
it the damn good thrashing my blade 1000 dual 750 has had. My usual
first test is a complete rebuild with optimisation of all our software
using forte 6.2 c++. That I can manage (and post comparisons against
said older machine). I might be able to squeeze in some other
benchmarks if people have some favourites. They can email me at
m i h a l i s at s p e a k e a s y d o t n e t
with suggestions in that case. I have not yet even established
precisely what spec I have - it's sitting in its box yet. I have more
important things to do (from employers point of view) else there is
almost nothing I would rather do right now than spend the rest of the
day setting it up and testing.

Chris
--
Chris Morgan
"Post posting of policy changes by the boss will result in
real rule revisions that are irreversible"

- anonymous correspondent