advice on cluster purchase

Robert G. Brown rgb at phy.duke.edu
Sun Apr 6 19:40:42 EDT 2003


On Sat, 5 Apr 2003, jbassett wrote:

> Hi, I am an undergraduate involved with a totally student run parallel 
> computing experience. We have approximately 10,000 of university money with 
> which to produce the best possible machine. I would  be interested to hear 
> from you all what configuration you would choose if someone just said "here's 
> the money, build the best system you can." The system will do both cpu 
> dominated and network intensive activities, so it would be tailored for 
> neither. Do SMP nodes tend to be superior in a cost/performance framework? I 
> have worked with other peoples systems and they are always dual cpu nodes, my 
> impression being that it is for the purpose of minimizing overall size- as I 
> tend to start a process on each cpu. Any advice would be appreciated. 

I think you can barely afford the following:

3 dual Xeon or dual Athlon systems.  Budget them for $1800-2000, get at
least 512 MB of memory per, small/wimpy IDE hard disk, gigabit ethernet
card.  Tower cases are cheaper and 4 nodes don't need a rackmount.  No
CD drives.  A floppy is ok, a cheap video card is ok although likely to
be onboard on the motherboard along with a possibly useful 100BT
interface.

1 dual processor P4 or Athlon with a gig card, a SCSI interface, and 3-4
SCSI disks set up in a RAID, in a server (supertower) case.  If data
preservation is very importanty to you and you can afford it, add a tape
or CD-RW to back it up.  If this "head node" is to connect to an
external network, buy it an extra 100BT interface.  Get it some
bric-a-brac, as well -- a CD RW, a nice sound card (if one isn't
onboard), some decent speakers -- this is where one will "work".  A bit
of extra memory (relative to the nodes) wouldn't hurt as well.

1 small gigabit ethernet switch.  Netgear has a cheap one.  So do other
vendors.  I leave it to your shopping process to determine the number of
ports -- at least 4, of course, but you might want to try for 8, or 16,
if you think your cluster might grow later.  You may want a cheap 100BT
switch as well (or extra ports for the 100BT interfaces) if you'd like
to preserve the gig network for IPC computations only.

1 four port KVM switch.  Don't go cheap -- good cables, maybe a Belkin
switch.  This should cost you $200+ (including cables) not $100-.  The
cheap serial/switch ones suck, and cheap cables will distort video.

1 monitor as large and nice as you wish.  If you can afford it, I'd go
for e.g a NEC 17" flatpanel that does 1280x1024.  Oh, and a nice mouse
and keyboard too.

1 heavy duty shelf unit.  See pictures on http://www.phy.duke.edu/brahma
for a nice one I got at Home Depot for $60 or so -- you only need a half
of one for four nodes, but your cluster might grow...

Miscellaneous cables, UPS/surge protectors, some nifty LEDs and glowing
lights to make people think it is a really powerful computer;-)

A name, and a nice logo.  Never underestimate the importance of
marketing...:-)

I make it (3x$1800=$5400) + (1x$3000) + $600 + $400 = $9400, plus
several hundred for the miscellaneous -- cables, shelf, KVM, UPS, and
anything I might have forgotten.  At least you have something to
structure a price search around while shopping.

Note that you won't get bleeding edge systems at these prices.  I'd
guess 2.4 GHz P4 Xeons or 2000+ Athlons with 1 GB of DDR, maybe a bit
better.

   rgb

> Joseph Bassett
> 
> 
> _______________________________________________
> 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



_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list