cheap clusters

[Robert G. Brown]

On Wed, 4 Apr 2001, Velocet wrote:

> On Wed, Apr 04, 2001 at 04:22:50PM -0400, Troy Baer's all...
> > On Wed, 4 Apr 2001, Robert G. Brown wrote:
> > > Speaking of DDR Athlons -- are there any motherboards that are actually
> > > available out there?  Most of what I was able to find in a modest
> >
> > There are a couple uniproc Athlon/DDR motherboards that are available
> > right now if you look hard.  The most widely available ones are the
> >
> > The major problem with DDR memory right now is finding DIMMs that are
> > stable at the 133MHz system bus speed.  Some batches of Nanya PC2100-rated
> > DDR RAM are very unstable (i.e. cause silent lockups) unless you downclock
>
> Im lost here. Why are people buying the most expensive single
> nodes for their computations? Arent we all building/running beowulfs
> because no single machine could deliver the power we need? So why
<deleted>
> Mebbe Im wrong. Mebbe DDR ram isnt still 3 times more expensive in your
> countries and the mainboards 2x as much as the average mainboard...
> mebbe this is all cheap. To get something like 1.5 to even 2.5x
> the performance for 2.5 to 3.0x the cost really doesnt seem to be
> a win, and these are extremely conservative numbers. (I need to get
> Atlas working with my copy of gaussian before I can compare directly.)

Goodness.  One of many problems is that MFLOPS is a fragile term.
Sometimes overall performance depends on memory speed (e.g. stream).
Sometimes it depends on CPU clock.  Sometimes in between.  So do other
subsystems that might be rate determiners.  Then there are the
nonlinearities.  As you say, there is a fragile (or rather complex)
dynamic.

I personally wasn't asking after DDR/Athlons because I want to waste
money (I was one of the list's biggest fans of the humble Celeron for
years:-).  My own calculations tend to be embarrassingly parallel and
cache size/memory speed insensitive, for the most part.  However, I'm
getting ready to parallelize a new calculation that will likely run
fastest (which is MUCH faster than it might otherwise be) in a big
memory, memory bound configuration.  It also isn't going to be
particularly homogeneous, as tasks go -- the computation separates and
different task components may well have different optimal architecture.
I was kicking around getting DDR for my head node/desktop and SDRAM for
the rest, depending on cost.  Once I found out what that (real, quoted)
cost was, by finding some actual motherboards and memory to consider
buying;-)

I agree that if the systems are even twice as expensive, one DDR system
is the most that I'm likely to buy because I also agree that in my
particular case compute power matters a lot and memory speed only
matters a medium -- not enough to offset a factor of two in cost for
sure.  1.3x the cost, I'd have to really think about it and would
certainly buy one for my desktop and to do the prototyping and thinking
with;-)

It is important to remember that I'm not "typical" -- I'm not sure
anybody is.  I'm certain there are folks out there for whom alphas, P4's
and DDR/Athlons as well as garden variety PIII's and KT7 Tbirds make
cost-benefit sense.

Oh, and one more point.  I still don't know what the memory limitations
of e.g. AMD's 760 chipset will be, but I do know that VIA's is 1.5 GB.
This alone is enough to justify DDR systems if they will hold 2 or more
GB, for at least some folks.  Local memory is likely to be LOTS faster
than memory read over or written to by any sort of network IPC.

> [ I realise there's a very fragile dynamic that can occur with
>   parallel computations - a whole slew of slower nodes that actually
>   adds up to more power (in Mhz/GFLOPs whatever) can be slower depending
>   on network chatter and switch partitioning and many other factors, or
>   the reverse is possible. Only detailed analysis of an actual cluster (or
>   a very good similation) running the computations in question can give you
>   the answers regarding what the best configuration here is. Obviously
>   people are asking about equipment they have no access to test something
>   this complex on, nor for extended periods in a reallife situation. Or perhaps
>   there is a definite indication that there is economy in faster and faster
>   single nodes for some types of computations and thats why everyone wants
>   faster nodes NO MATTER the cost, the advantage is that great. I cant believe
>   that thats true for all cases however. Can someone explain to me how this
>   factor plays into the cost analysis? ]

No, you're dead on the money -- cost benefit is the correct calculation
to make, not raw performance.  The important thing is getting your work
done in an acceptable amount of time for the least money.  However, "an
acceptable amount of time" means different things to different people,
and then there is all sorts of overhead to consider.  For some, reducing
the time to 0.8 might we worth 2x the cost, especially if the 0.2 time
saved costs far more than twice the cost or if the result lets you start
making 10x the money for those two extra months.  Cost benefit isn't
generally a simple linear function.

   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