Dual AMD systems in rackmount cases
Robert G. Brown
rgb at phy.duke.edu
Mon Jun 24 09:13:12 PDT 2002
On Mon, 24 Jun 2002, David Chalmers wrote:
> Hi All,
> Does anybody have any experience with Dual AMD systems (MP2000) in 2 RU
> rackmount cases? Is heat generation a problem? I am planning on
> building a system with Tyan 2466 motherboards and 400 W power
> supplies. Should I use 3RU cases?
Heat generation is a serious problem for dual AMD's. You can succeed
with 2U cases, but you will need a good case with several fans driving
flow through the case and high-grade heatsinks for the CPUs themselves.
As important as these two components is being able to provide ambient
air at around 65F (or cooler, of course:-) on the air intake side of the
cases and to get rid of the 75F air coming out of the air outflow side
of the cases -- the cooler the ambient air, the more forgiving the node
design in other ways.
Dual AMD's will just plain crash when they get hot, and they get hot
very, very easily as they draw about 150 W sustained under load. In our
new server room, our 2U cases were not originally getting enough, cold
enough, air and the temperature at the case rears was up in the 80's.
The nodes would crash after 20-30 minutes of operation, consistently.
Even with that situation repaired, we still get an occasional node crash
that might well be heat.
3U cases would likely run more stably IF you still can provide cold
ambient air and adequate stable air circulation patterns, but of course
that reduces your cpu density from 1/U to 1.5/U. Then it comes down to
whether you need the space/volume. If not, 3U would be better.
Based on my experiences thus far, I just wouldn't run 2466's in a 1U
case. In 2U, a full sized rack with (say) 16 2U cases and 32 CPUs plus
patch panel and/or switch plus a bit of spare room will burn about 2500
Watts -- something above 1 KW/meter^3. Doubling that to more than 2
KW/meter^3 seems unwise -- any sort of bobble in your cooling system
will rapidly cause your stack to reach furnace temperatures, and even
with nodes idling as they crash you might start to damage hardware
before enough crashes can occur. Worse, even idle load is nontrivial at
those densities -- the stack won't cool that much just because its
components have crashed.
> Thanks for any information.
> David Chalmers Lab: 9903 9110
> Victorian College of Pharmacy Fax: 9903 9582
> 381 Royal Pde, Parkville, Vic 3053 http://synapse.vcp.monash.edu.au
> Australia David.Chalmers at vcp.monash.edu.au
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
Robert G. Brown http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567 Fax: 919-660-2525 email:rgb at phy.duke.edu
More information about the Beowulf