More stable hardware ??

Robert G. Brown rgb at phy.duke.edu
Thu Jan 11 14:15:03 EST 2001


On Thu, 11 Jan 2001, Roche, Olivier {PRBT~Basel} wrote:

> Hi,
>
> I have scan a little bit the archive of the mailing list and I have
> seen many particular points on hardware, but I would like a more
> general view...
>
> I would like to build a 16 CPUs beowulf cluster. It seems that
> dual processors are better. So I would like to purchase 8 nodes.
> But could someone tells me which hardware, at the present time,
> have the most high rate between stability over performence ??

The "it seems" part is a bit worrisome.  I'm a longtime dual/beowulf
user and sometimes duals are better and sometimes they aren't.  For my
code they are marginally better in that one can get more CPUs per dollar
and my code is demonstrably CPU bound and so more CPU per dollar
translates effectively into more work done per dollar per unit time (for
me) which is my real goal.  However, for code that is memory I/O bound
(or network bound or bound by ANY CPU-shared resource) they are likely
to NOT provide more work done per dollar per unit time.

To put it another way, if one runs a cpu benchmark (like cpu-rate on the
brahma site) in cache, one gets one result (typically 100's of MFLOPS in
L1 or even L2).  If one runs the cpu benchmark out of memory, the speed
ends up being limited by the speed of the memory and not the CPU clock,
and PC133 will pretty much yield just about 100 MFLOPS as an upper
bound, independent of CPU speed or clock.  Running two copies of the
benchmark at the same time (when one copy alone is saturating the memory
base and is indeed rate limited by the memory bus) one is NOT going to
complete in the same time and get twice as much work done.  Neither is
it likely to take twice the time -- the blocking isn't perfect.
Performance is likely to be somewhere in between.

Other code is structured and organized so that "cache works" -- so that
memory accesses are relatively rare compared to intervals that
computations are done one data in cache.  In that case two jobs run on a
dual will complete in almost exactly the same amount of time that one
completes.

In the former case duals may well be a bad idea.  Buying a faster clock
CPU or more memory with the money that you might have spent on a dual
motherboard and extra CPU may be a better deal with less potential
instability.  In the latter case they are likely a good idea.  It
depends on your code.  Test and benchmark before making a decision.

>
> I have read that, maybe SuperDER is a good motherboard (or not ?).
> I would rather think about ABIT or ASUS and maybe 3com as a
> ethernet card, but I have no clue how this particular hardware
> fits with a beowulf cluster.

I've used the Abit BP6 with dual celerons without problems, but it no
longer exists and celeries are pretty much tapped out as a processor
family.

   rgb

>
> So if you can give me any indication about the "stable" hardware I
> can pourchase (I will administrate the cluster !!), It will be very helpful.
>
> Thanks,
>
> Olivier
>
> _______________________________________________
> Beowulf mailing list
> Beowulf at beowulf.org
> 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
http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list