|
As i'm not a complete layman in fluid dynamics
software i'll make a few assumption to advice a choice:
assumptions:
a) that gigabit is enough for you and that you
don't need bandwidth to other nodes
b) that you are interested in having a huge amount
of RAM at each compute nodes
c) that price is important as you probably want to
put there a lot of nodes
d) that your code usually runs single cpu and uses
all RAM
In that case perhaps an idea is getting a new dual
opteron mother board and put inside
a single core opteron 2.6Ghz or 2.8Ghz if they are
there. That's unbeatable in terms of
latency to RAM.
Those new boards sometimes support up to 16 DIMMs.
That might be very interesting for you.
If you put in 2x166Mhz (PC333) 2GB dimms * 16 ecc
registered CL2 = 32GB.
That will be faster for you than PC3200 (as it
might get clocked back to pc2700 anyway) cl 2.5
for latency.
Probably this is very interesting for you. At least
1 manufacturer i saw (tyan) has a dual opteron board that
supports 16 DIMMs. Perhaps there is others
manufacturers which offer 16 dimms. I can't recall
manufacturers offering more DIMMs than
that.
Perhaps certain quad boards do?
Note that there is 1 solution not so cheap that
offers 32 DIMMs. That's creating a 8 core machine.
Iwill and Tyann should both have a solution there.
Put in cheap single core 8xx chips (which are there
especially at ebay cheap as everyone wants to
get rid of them and get dual cores for codes that
need the cpu speed instead of huge RAM).
And fill it with 32 DIMMs PC2700. That would give 1
compute node a 64GB ram.
I'm not sure whether pc2700 cl2 or pc3200 cl2.5 is
faster in latency at those quads.
Could be pc3200 cl2.5 wins it there. It's a lot
more expensive though.
Good luck,
Vincent
|
_______________________________________________ Beowulf mailing list, Beowulf@beowulf.org To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
