[ibm-pc] Re: "INT64 Confirmed"

["FEEB" <feeb zavinac chem bod utoronto bod ca>]

On 29 Jun 2004 13:28:49 -0700, kujebak wrote:

>"FEEB" <feeb zavinac chem bod utoronto bod ca> wrote in message 
news:<srropurzhgbebagbpn bod i00sye1 bod pminews zavinac news1 bod chem bod 
utoronto bod ca>...
>
>> We did some benchmarking of our specific code on various HW.  The 
>> differences are quite staggering.  Code A would run the fastest on 
>> processor X, while slowest on processor Z.  Code C would be the exact 
>> opposite.  Considering execution times, the code that would be run, 
price 
>> and previous experience we ended up ordering a 128 CPU Opteron cluster 
as 
>> an addition to our Athlon, Alpha and XEON cluster farm.
>>   
>> Actually I have a dual Opteron 1U machine with serial ATA RAID on my 
desk 
>> right now.  Running FC2 64 bit kernel it outperforms Athlon 2800+ 
running 
>> some floating point code.  And its the cheapest Opteron at 1394.404 
MHz.  
>> 
>> If you are seriously interested in this area, take a look at the 
InfoWorld 
>> February 2, 2004 article about 2.6 Linux kernel.  There are lot of 
things 
>> besides just the HW.
>> 
>> 
>> Frank Bures, <feeb zavinac chem bod utoronto bod ca>
>
>The second page of http://www.overclockers.com/articles913/index.asp 
talks
>about the possible strategies behind Intel's own implementation of x86-64
>in their Xeon line (http://tinyurl.com/2lnhg).

Yes, I read that.

>Do you have a personal opinion about the hardware technology direction
>with respect to the two competing architectures (x86-64 vs. Itaniums)?

I do not have any fixed opinion nor I'm trying to have one :-)  My 
opinions in this regard are highly fluid :-).
We just use whatever is the best value for our purposes.  
Before investing this kind of money ($300,000), we do some benchmarking on 
the particular code samples and then order what's the best overall value.  
That's why we did not go for quad MoBo's as they are too expensive right 
now.
What makes Opterons so interesting is the way they manage memory access.  
Especially if you use dual and quad MoBo's in Beowulf setup.  If your code 
is written well for the architecture (you are aiming at lowering the 
inter-node data transfer), the subsequent increase in processing speed is 
more than considerable.
I'll let you know how the cluster works after it has become fully 
operational.  Currently working on A/C modifications in the room :-)

FYI:  We are using dual Opteron MoBo's with 4GB RAM per CPU (8GB on 
board).

Rather outdated and incomplete description of our resources is at
http://www.chem.utoronto.ca/~dgruner/cptg/frontpage.html


Frank Bures, <feeb zavinac chem bod utoronto bod ca>