Specific Question about Single vs. Dual Processor System

Robert G. Brown rgb at phy.duke.edu
Sun Apr 6 19:57:24 EDT 2003


On Sun, 6 Apr 2003, Chih King wrote:

> Hello.  I am a member of the University of Iowa Student Supercomputing
> Project (UISSP), and we are planning for the purchase of our first cluster.
> Currently we are divided between a sixteen node single-processor Pentium 4
> system and a seven node dual-processor Xeon system.  Here are the brief
> specification of both machines:
> 
> 16 Pentium 4 single-processor system (total cost $7,407):
> 
> Intel Pentium 4 2.4GHz 533FSB 512KB
> ASUS P4S8X 533FSB SATA GB 8X MB w/ 10/100/1000 LAN
> 512MB PC2700 DDR333
> Maxtor 20GB Ultra100 Hard Drive
> ATI Rage Mobility VGA Card 8MB AGP
> CG 6039L 350W USB Midtower Case
> Linksys LNE100TX 10/100 Network Adaptor (2nd LAN channel)
> 
> 7 Xeon dual-processor system (total cost $8,400):
> 
> INTEL XEON 2.4GHZ 533FSB PROCESSOR x2
> TYAN S2723GNN E7501 GLAN MOTHERBOARD
> PC2100 256MB ECC/REG DDR x2
> Maxtor 20GB Ultra100 Hard Drive
> Chenbro Beige Server Case
> NMB 460W Xeon Power Supply
> MITSUMI 54X CD-ROM Drive
> 
> As you can see, the single-processor system is about $1,000 cheaper than the
> dual-processor system.  We have a total of $9,500 in our budget (to pay for
> the system, the switch, and everything else).  Taking into consideration
> both performance and economical issues which system would you choose and
> why?  Some more details: since Gigabit LAN is built in both motherboards we
> will probably establish one Gigabit channel, and if necessary have a second
> 100Mbps LAN channel as well.  Therefore we will probably have to spend an
> additional $500-600 on switches.  Currently we are not sure about specific
> application that we will be running on the cluster, but we would like to run
> a broad range of calculations/simulations (ie. biological, economical,
> mathematical, etc.)  We would really appreciate any response in this matter.
> Thank you very much!

Hmmm, I think I just responded with ONE plan -- looks like you already
have better quotes than I expected EXCEPT that you look like you're
getting less memory than I think you should get on the duals.  I'd
recommend at least 512 MB per processor, maybe 1 GB per processor if you
can afford it.

You also haven't said anything about a server -- if the cluster is going
to do any serious work, you'll likely want a "server node" in either
configuration with a lot more than 20GB in relatively unreliable IDE
drives.  You are also getting more nodes (either way) than will
comfortably fit on a cheap KVM, which is ok but not as convenient on a
starter/demo cluster when you'll have relatively many occasions to
connect directly to nodes to mess with them.  So replace the KVM with
just the monitor, keyboard, mouse themselves and a cart to put them on.

Now, about your question.  The UP systems have faster memory and more
memory and more processors total.  If you REALLY have no more money,
having 16 systems is better than having 7 if you have to deal with node
failures.  You do have to get a bigger switch, you do give up a bit of
speed when processors have to talk at least some of the time.

I think I'd get the UP configuration, with one node pulled and beefed up
in a bigger case into a server node.  The other 15 and the switch will
fit very neatly onto a single heavy duty steel shelf unit, and can be
cabled up to look lovely.  This should be very serviceable.  For coarse
grained or embarrassingly parallel code you've optimized CPU and have
more memory for applications; for parallel code with a fair bit of IPC's
you no longer can talk to at least ONE processor locally, but neither do
you have to share a single gigE connection among two processors.  It
will look more impressive.  It will run slightly hotter and cost
slightly more to operate, if you are paying the power bill (about
$2K/year, at a guess, so I hope you are NOT paying the power bill:-).

   rgb

> 
> Sincerely,
> 
> Chih King
> UISSP
> 
> 
> _______________________________________________
> 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