[Beowulf] Academic sites: who pays for the electricity?

Glen Gardner Glen.Gardner at verizon.net
Thu Feb 17 17:12:27 PST 2005


It sounds to me as though you are seeking low power beowuf as a solution.
There have been a few people build such machines, and it is possible to 
build a fast, useful beowulf
cluster that uses very little electrical power and has sufficient muscle 
to do some serious work.

I have a small 14 node cluster which I built a year ago.  It  uses very 
little power , and runs so cool that no room air conditioning is needed.
In fact, my p4 machine makes more noise and heat than the beowulf 
cluster in my apartment.


The above link shows the original 12 node configuration.

At present, there are some motherboards available which give a very nice 
combination of cost, performance and low power use.
The trick is to "right-size" everything for your needs and available 
resources. The down side is  that the small low-power go-fast stuff is a 
little more pricey than the plain vanilla pc hardware a beowulf is 
usually buit from , but not insanely so.

Transmetta has some nice boards, and the minit-itx boards are not bad at 
all for the cost.  Also, there are some rather nice small form factor 
motherboards that use AMD's geode cpu. When I start comparing cost, 
power use, and performance, so far the most attractive motherboards seem 
to be the mini-itx boards with the nemiah core cpu.
However with some low power geode boards now running at up to 1500 MHz, 
that may change.  The Transmeta boards are probably the fastest of the 
low power boards, but the power use per MIPS is not as good as other 
boards if you believe the Transmeta printed specifications.



I have also made a few comments  below.

David Mathog wrote:

>At Wed, 16 Feb 2005 19:08:05 +0100 Vincent Diepeveen wrote:
>>Date: Wed, 16 Feb 2005 19:08:05 +0100
>>From: Vincent Diepeveen <diep at xs4all.nl>
>>Subject: Re: [Beowulf] Academic sites: who pays for the electricity?
>>To: "David Mathog" <mathog at mendel.bio.caltech.edu>,
>>	beowulf at beowulf.org
>>Message-ID: < at pop.xs4all.nl>
>>Content-Type: text/plain; charset="us-ascii"
>>At 08:16 16-2-2005 -0800, David Mathog wrote:
>>>In most universities services like electricity, water, and 
>>>A/C are paid for by the school.  To do so they take "overhead"
>>>out of every grant.  Partially as a consequence of this they
>>>typically have a very poor ability to meter usage on a room
>>>by room basis.
>>>Now somewhere between the 10 node Pentium II beowulf sitting on
>>>a lab bench and the 1000 node dual P4 Xeon beowulf in a machine
>>>room that takes up half the basement the cost of the electricity
>>>(both for power and A/C) goes from  a minor expense to a major
>>>one.  Really major. For instance, in that hypothetical large machine,
>>>at 10 cents per kilowatt hour (a round number), assuming 100 watts
>>>per CPU (another round number) that's:
For a dual p4 xeon machine at full throttle, it comes out to about 250 
watts per node (or a little less) including the network adapters and 

>>> 1000  (nodes) *
>>>    2  (cpus/node) *
>>>    .1 (kilowatts/cpu) *
>>>    .1 (dollars/kilowatt-hour) *
>>> 365   (days /year) *
>>>  24   (hours/day) =
>>> 175200 dollars/year
>>Complete academic nonsense calculation. If you use quite some electricity
>>the electricity gets up to factor 20-40 cheaper. Getting a factor 10
>>reduction in usage bill is pretty easy if you negotiate properly.
>Well, it isn't complete nonsense, unless you care to dispute the
>number of days in a year, hours in a day, or cpus in a dual node
>The only term you're complaining about is the price of
>electricity.  I'm not privy to the electrical rates that our
>school pays, they may well be an order of magnitude lower.  My
>home rates certainly aren't, but then, I don't buy as much
>power as the campus.  It's also not at all clear that the
>campus would sell power to the end users at the same rate
>which it pays the utility.
You are forgetting the cost of cooling the cluster. Big machines make a 
lot of heat, and need a lot of cooling.

>I don't really understand your point about keeping the units
>running versus restarting them.  Sure, it would be really bad
>to try to boot all 1000 nodes simultaneously, in all likelihood
>it wouldn't work.  That's why they are typically started at N
>second intervals, where N depends on your hardware.
>Surely there is some N large enough so that the peak current
>draw during the restart never exceeds the random fluctuations
>observed when all units are running normally.  Or is your
>point that the electricity company doesn't want the facility
>to draw _less_ current than it uses normally at
>steady state?
It is important to keep the cluster up and running, and only cycle the 
power when you must.
The inrush currents at turnon stress components and shorten the life of 
the nodes considerably.
Also, thermal cycling puts mechanical stresses on boards and components 
that can cause
components and connections to fail.

In a large cluster that is middle aged (@ 2 years old), you can 
reasonably expect to lose a couple of nodes
every time you power down and come back up. After a while, this can be 
Shutting down a big machine is not a trivial thing.

>On a somewhat related note, it would be nice if rack nodes
>had some graceful way to conserve electricity.  For instance,
>something along the lines of: if the CPU utilization goes
>below 5% for 10 seconds ratchet the clock down by a factor of 10.
>When CPU usage goes above 90% ratchet for 2 seconds move it back
>up again.  Notebooks can do this sort of thing, but it seems not
>to be a "feature" of most full size motherboards.  This should
>also lower the average temperature in the case, at the expense
>of increased thermal cycling.  Hard to say off hand if that's
>a plus or a minus as far as hardware longevity goes.  Certainly
>it would be a plus in terms of energy conservation.

A lot of modern cpu's have the ability to actually shut off unused 
internal circuitry.
VIA CPU's, AMD's geode, Transmeta, and some intel cpus have these features.

>David Mathog
>mathog at caltech.edu
>Manager, Sequence Analysis Facility, Biology Division, Caltech
>Beowulf mailing list, Beowulf at beowulf.org
>To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf

Glen E. Gardner, Jr.
Glen.Gardner at verizon.net


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.beowulf.org/pipermail/beowulf/attachments/20050217/be311034/attachment.html>

More information about the Beowulf mailing list