[Beowulf] Re: CCL:Opteron or Nocona ?

[Maurice Hilarius]

>
>
>From: mark somers <m.somers at chem.leidenuniv.nl>
>Date: Mon, 9 May 2005 09:38:06 +0200
>To: Luigi Cavallo <cavallo at chemistry.unina.it>
>Cc: chemistry at ccl.net
>Subject: CCL:Opteron or Nocona ?
>
>Dear Luigi Cavallo,
>
>we have recently purchased and setup a new Beowulf cluster here in Leiden, 
>dedicated to run ADF and other DFT codes (Dacapo and Vasp) on it. We have 
>found the Nocona's, in combination with the Intel compilers, to be the best 
>combination especially for ADF and ADF-BAND.
>
>Before we decided to go for the Nocona's, we did try with two test machines, 
>one being a dual-cpu AMD Opteron 246 (2.0 GHz) and the other being a dual-cpu 
>Nocona (3.0 GHz). Tests showed that the AMD has, in general, a slightly 
>better cache hierarchy and a better memory scalability, but also that the 
>lack of compilers able to tune for the Opterons is severe. Of course we took 
>the differences in cpu clocks into account.
>
>  
>
LACK OF COMPILERS IS SEVERE?

You apparently did next to no research on the subject.
A simple Google search on the terms:  "Opteron compiler" provides a 
wealth of information.
GCC, Intel, Absoft, and especially Pathscale compilers are excellent for 
Opteron.
New Revision E and dual core Opterons have SSE3.

As for your comment about "taking the differences in CPU clocks into 
account" I am quite puzzled.
Even Intel is now pushing "performance ratings".
In current CPUs a comparison based on clock speed is nearly meaningless.

>As it turns out, after having contacted the SCM people in Amsterdam, ADF can 
>effectively use the SSE3 registers and cpu intruction set and that makes it 
>run fast on the Nocona's. This, together with the prices being roughly the 
>same, made us decide to go for the Nocona's. 
>
>BTW, maybe this is known to you already, but Intel offers you their OpenMP 
>capable compilers for free for academic use. 
>  
>
No, they do not!
They are only "free" for essentially hobbyist use.
If you use them in a university or research cluster that violates the 
license.

I think you had a nice relationship with the Intel reseller, and "drank 
the Kool-Aid" he gave you.

Sorry to be harsh, but the statements you make above are so one-sided 
and uninformed, I had to speak out.

For HPC the numbers are so overwhelmingly in favour of Opteron as to be 
virtually insurmountable.
Before I get flamed, yes, I agree there are some codes that are better 
suited to the Intel CPUs.
Just not the vast majority.

Further, now that dual core Opterons are shipping, the whole game just 
got reset.

Any advice you can give based on a purchasing decision made even 3 
months ago is now so obsolete as to be virtually irrelevant.

Here is a recent and fairly comprehensive comparison on the AMD versus 
the Intel single and dual core CPUs performance.
http://techreport.com/reviews/2005q2/athlon64-x2/index.x?pg=1

The power consumption and heat production is especially telling.
Add to that the fact that with dual cores the whole architecture just 
changed drastically and I think you will see that we now start over with 
"new rules".



With our best regards,

Maurice W. Hilarius
Hard Data Ltd
email: maurice at harddata.com
http://www.harddata.com/