Athlon SDR/DDR stats for *specific* gaussian98 jobs

Robert G. Brown rgb at phy.duke.edu
Thu May 3 05:23:56 PDT 2001


On Wed, 2 May 2001, Velocet wrote:

> > Not yet, but maybe soon.  The fast Tbirds do require a big "certified"
> > power supply, but I'm guessing they draw a lot less than they "require"
> > except maybe in bursts.  I'm betting they draw around 100-150W running,
> > a number that recently got some support on the list.
>
> Bursts of CPU usage? Arent all our clusters all hammering our CPUs as much as
> possible? And if ATLAS is really doing its job, arent we hammering all parts
> of the CPU as much as possible? :)

Not variations in the CPU/memory load, which as you note is nearly
constant (and not really all THAT different between idle and loaded -- a
lot of juice is expended just keeping it and the memory running in idle
mode), variations in the peripheral load -- using the disk(s), the
network and so forth.  In your configuration (very stripped, if I
recall) you don't think you'll see much variation.

I'm going to see if I can get Duke to spring for some tools to measure
power draw properly.  There are all sorts of peak vs rms issues that
were discussed on the list a year or so ago, and this matters for a
project we have underway (we're getting a "beowulf room" built from
scratch in the department and are dealing with the feeding and cooling
of some as yet indetermined number of nodes, where I've been using 100
Watts/node average power consumption as a rough guestimator for node
consumption).

> What kind of deals can you get in bulk? From AMD themselves? do you
> need to be a big university and have a big press release event to
> get these deals from them? How many do you need to get a batch deal?
> How deep is the discount?

Ah, I don't shop that way.  I love to save money and all, but:

  a) It's other people's money (mostly -- I do run my home 'wulf out of
pocket, sort of).
  b) I have considerable personal experience of the penny wise, pound
foolish variety.  I'll cheerfully spend some of my OPM budgets on
convenience, vendor relationships, and extranea that cost me a few nodes
overall but may get more work done with less hassle.
  c) I factor in the cost of my own time at a pretty hefty rate for
setting everything up and maintaining it.  I therefore get things set up
so they will be very low maintenance, as my time ends up being worth
(trading off) quite a lot of hardware.  To me, anyway.

In my case the best solution to optimize these parameters seems to be do
business with a reliable local vendor, missing the absolute best deals
available anywhere by an easy 10% or so but getting the warm fuzzies of
a place where I'm on a first name basis with the staff (who of course
adore all the lovely money I spend there and are willing to earn it)
that will fix things for me without invoking the daemon-gods of
depot-repair hell (or playing the mail game -- they mail you a piece,
the next day or so you try it, if it works you mail back the bad part,
otherwise you mail back the good part and they mail you a piece, ad
nauseam).  I just carry my vendor the node or the part if I've
definitely isolated it (ten minute drive) and the next day or next hour
they hand back a working node, with a full shop and lots of immediately
available parts to swap to apply to the repair.  Only same-day on-site
service contracts from e.g. Dell can beat it, but Dells cost another 10%
or more per node and its harder to shop and microconfigure.

After all, nodes will be with us always, and new budgets for buying them
are arranged every few years.  Whatever I get will be obsoleted in three
months so anything I get will be a "mistake" from one point of view (if
only I'd held out for the 1.9 GHz SuperCruncher with relativistic
predictive dram -- delivers the memory before you ask for it -- THEN I'd
have been Happy).  Better to take the long view, buy in mid-sized chunks
(giant purchases of 256-node 'wulfs can require real expertise to get
right and cost you a LOT of money if you get them wrong and are often
best arranged via a turnkey provider with some consultative experience).
That way, if one mid-size chunk is less than perfectly optimal, so be it
-- you learn from it and arrange the next mid-sized chunk to be better
spent.  You also get to buy that 1.9 GHz SuperCruncher with the next
round of money spent.

This last point is worth examining.  The way Moore's Law works it is
amusing but true that if you take a fixed three year budget of 3A and
spend it all at once, you get (3A)*(3 years) = 9 work units done over
three years.  If you spend it A per year, you get (A)*(3 years) +
2*(A)*(2 years) + 4*(A)*(1 year) = 11 work units done over the same time
(the numbers reflecting the approximate annual doubling in speed from
ML).  That is, you break even in work done between years 2 and 3 and
thereafter accumulate work units at A+2A+4A = 7A per year vs 3A.  Also
note that you break even in the RATE at which work is done at the
BEGINNING of the second year -- by spending your money incrementally
(likely matching the ramp-up in work load, unless your users are "ready"
to jump in and simply crank up to full speed immediately) you get almost
as much work done in the third year alone as one would in three spending
everything all at once.

Ain't exponential growth wonderful?

So my one piece of parting advice is to worry less about getting "the"
absolute best (most cost effective) hardware as it exists right now --
your cost-benefit optimization calculation may not survive literally
from week to week anyway.  Last week the bleeding edge Tbird dropped by
almost 25% of its price (so my concern about NIC prices in the cluster
I'm getting turned out to be specious -- I'm getting upgraded systems at
constant cost with the originally quoted 3c905's in place).  P4 prices
are plummeting.  Clock speeds keep edging up within CPU families, and
then there are the 64 bit children of those families waiting to be
delivered to the world.  Nothing you get now will appear to be a wise
purchase six months from now, but if one DOESN'T buy in at some point
one never gets started and Moore's law doesn't quit.

[...a clever person perhaps have noting that if one did nothing but bask
in the rays of the Caribbean sun for two years and mosey back north to
buy 3A*4-speed systems at the start of the third year, one would get 12
work units done in the third year alone and thus beat out even my A per
year purchase schedule and get a nice tan besides.  Or worse, waiting
one MORE year gets 3A*8 = 24 work units done in the fourth year alone.
In the limit, NOBODY should EVER buy computers to do numerical
calculations now, as the longer they wait the less time it will take to
complete them once they start and if we all just waited long enough a
single desktop unit would get more work done than all the beowulf units
currently in existance put together...  hmmm, something wrong with this
logic, head hurts, must seek solution -- oh hell, might as well go get
tickets to Jamaica...:-)

Forgive my morning ramblings...

     rgb

-- 
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







More information about the Beowulf mailing list