[Beowulf] A Cluster of Motherboard.

[Florent Calvayrac]

Glen Gardner wrote:
> That is a nice cluster!
>
> Have you modeled it in a smaller, higher density form ?
>
> Glen
>
>   
Thanks !
To be perfectly honest this was a student project so we had a natural 
deadline
of the end of the semester to build it, since I found it particularly 
interesting
for the students to interact as they where designing the cluster
 with the  lab technician who was building it, G.Ripault, whom you see 
on the pictures,
and who suggested the rail mounting solution.  Thus, the students could work
on Beowulfery, Linux, HPC, CAD, mechanical design (bit of an overkill 
here), thermal design (they
were the ones to suggest we buy a discount Flotherm licence after some 
googling), budget,
and production problems with realistic feedbacks at each step (we even 
sat on the
box at the end...).

The final design was adopted as the calculations on Flotherm were not 
yet systematic
(we lost a month or two since we had forgotten to include properly 
turbulence and numerical
viscosity, as discussed in the report) so to be on the safe side we 
expanded the box
a little.  Amazingly enough the calculated temperature in the box (27 
degrees C for 17 degrees
ambiant air) was within 1 degree of the measured one when everything was 
taken into account
with the proper power specifications for the processors, memories and 
the board, other
characteristics coming from the library of components in Flotherm (a 
good selling
point for that program, who in absolute terms would not be competitive with
Star-CD or Fluent).

I think that the motherboards  could be  50% closer  without problems ; 
but we have
a dedicated room for our larger, more traditional clusters (68 PIII on 
TCP/IP  + 16 PIII on Myrinet
+ 10 AMD-64 with 16 Go RAM each processors now) because of the cold and the
noise, and had only budget for 8 processors on such an adventurous 
project, so...

In the same way, we could have used the same power supply for several 
nodes,
but ready-made ones with several outputs were expensive and hard to get 
with our
stupid rules for public procurements in France (kafkaian bureaucracy to buy
anything unusual), and we computed that the probability of frying a 
motherboard and a CPU if we
did not rewire correctly a single-output one was high enough not to try  
it from
a cost-benefit point of view (short of extreme bad faith with the 
motherboard
vendor...).

 

-- 
Florent Calvayrac  | http://www.jackywulf.com
Directeur du SC Informatique Ressources Num.  de l'Universite du Maine
Lab. de Physique de l'Etat Condense UMR-CNRS 6087 
Inst. de Rech. en Ingenierie Molec. et Matx Fonctionnels  FR CNRS 2575