[TriLUG] 32+ GB RAM PC advice

[Carlos J. Cela]

Kevin,

Do you know if that means that as far as the software application is 
concerned the memory "bridging" between processors is transparent? or do 
I need to make something inside  the application when writing the 
simulation programs?

Thanks,
Carlos-

Kevin Flanagan wrote:

>I recently saw an AMD presentation, they way that they access memory has
>to do with the need for the 3rd and 4th processors.  RAM is local to the
>processors, then there's a "bridge" (my word, not theirs) between the
>pairs of processors.   When you get to >16GB of RAM this bridge is
>required, or if you have >2 Processors.  Their claim is that this bridge
>is so fast that it's not a delay.  When using only 2 processors and
><16GB of RAM it sure does appear that it would be faster.  I don't know
>enough about the details to say for sure, but to me it sure looks like a
>dual proc 16GB system would have really high speed access to RAM, and
>with 64bits it should really fly.
>
>
>
>Kevin
>
>On Sun, 2004-09-12 at 12:03, Carlos J. Cela wrote:
>
>  
>
>>Ed,
>>
>>Thank you for your detailed answer. I have been using Fedora Core for a 
>>while, so I guess Opteron/Fedora Core will be the combination of choice 
>>here...I have noticed that you can configure 32GB RAM only if you have 4 
>>processors installed in the machine....do you know if in a 32GB RAM/Quad 
>>Opteron any of the 4 processors can access the full range of memory? I 
>>mean, if my application is single-threaded, can I use the 32GB od RAM or 
>>just to 1/4th of it? is all of the memory shared? if so, why the '4 
>>processors installed' requirement?
>>
>>Thanks,
>>Carlos-
>>
>>
>>Ed Hill wrote:
>>
>>    
>>
>>>Hi Carlos,
>>>
>>>For large memory systems (>4GB RAM) the Xeons are awful.  Instead, you
>>>should consider Opterons, Itaniums, or other hardware.  We have a number
>>>large-memory systems within our network
>>>
>>> http://acesgrid.org/technical_blueprint.html
>>>
>>>including:
>>>
>>> + Dell dual-Itaniums w/ 16GB RAM
>>> + dual-Opterons: (mostly Tyan 2885 MB w/ 16GB RAM)
>>> + an Altix 350 w/ 16 CPUs and 32GB RAM
>>>
>>>and they perform *far* better than the Xeons on predominantly memory-
>>>bandwidth-limited scientific applications.
>>>
>>>Dual (up to 16GB RAM) and quad (up to 32GB RAM) Opteron systems can be
>>>purchased and/or assembled for a few thousand dollars and they scale
>>>*remarkably* well.  I'm very fond of our Opterons.  We see almost linear
>>>(that is, per-CPU) speed-ups with many applications including our main
>>>ocean/atmospheric modeling software.  And Linux (we use both Fedora Core
>>>2 for x86_64 and RHEL v3 for x86_64) runs quite nicely on the Opterons.
>>>
>>>For more than 32GB RAM, you'll have to look into SGI Altix or other
>>>high-end hardware.  For truly huge simulations it is often best to get
>>>time at a center such as:
>>>
>>> http://www.ncsc.org/
>>> http://www.psc.edu/
>>>
>>>rather than trying to buy your own.  As an academic, you can usually get
>>>grants of computing time on their (often huge) systems quite easily.
>>>
>>>Good luck with your simulations!
>>>
>>>Ed
>>>
>>> 
>>>
>>>      
>>>
>
>  
>