DUAL CPU board vs 2 Single CPU boards: bang for buck?

Robert G. Brown rgb at phy.duke.edu
Thu Mar 7 09:03:55 PST 2002

On Thu, 7 Mar 2002, Jim Fraser wrote:

> Robert,
> You bias the calculation a bit with the selection of your hardware...if you
> use inexpensive COTS..we are talking something like 50 per case not
> 350...also a decent mother board with NIC can be had for 150 and you   need

I don't bias it at all.  Find me a 2U rackmount case for $75 and I'll
buy it on the spot.

To put this in greater perspective:

  Cost of renovating a 40 m^2 server room in the physics department in
Duke to provide 75,000 watts of electricity and 75,000 watts of
continuous AC capacity: $150K, or nearly $4000/m^2.

In a 43U standard rack (allowing a few U at the top for punchblocks) I
can fit perhaps 18 2U cases in roughly 1 m^2 of floor space, although in
practical terms on cannot put anything like 40 racks in this room
because one needs room to move around and because the overhead AC/return
is too low in parts of the room to allow a 43U rack underneath it
anyway.  We expect to get perhaps 20 racks into the room eventually --
some space is still devoted to shelfmount towers since in our old room
that's what we used and we plan to use (up) the systems until forced to
retire them.  Note that if we DID put 18 2U cases per rack and had 20
racks, we'd have 360 2U cases and 720 processors, drawing roughly 100W
per process and hence would be "at" room capacity in other dimensions,
so this is a pretty safe upper bound.  A better cost estimate for the
space is therefore $150K/20 or $8000/m^2 in terms of practically usable
space.  It cost us something like $200 per U in a RACK just for the
floor space -- the actual cost of the case is considerably less than

This, of course, argues for duals in >>1U<< cases as being the most
cost-beneficial, but a) we can't afford to run that many CPUs in the
room anyway because of the power dimension; and b) in a 2U case, a dual
runs "hot" but there is a tiny bit of thermal ballast associated with
the case size.  In a 1U case a dual runs excruciatingly and dangerously
hot.  Not dangerously hot in that the cases aren't necessarily
well-engineered -- dangerously hot in that a FAILURE of that engineering
(say a case fan) can cause a chain-reaction meltdown in a very short
period of time because of that LACK of any thermal ballast, especially
in a 1U node in the middle of a whole rack of 1U nodes that trap its
surplus heat.  I just don't think it is quite as robust.

1U singles vs 2U duals is a fairer comparison at the same power density,
but a 1U case and a 2U case cost just about the same.

Obviously, any sort of shelfmount/tower packaging, which DOES INDEED
reduce the relative cost advantage of duals (without quibbling about
"best of pricewatch" prices for hardware vs "best price you're likely to
get from a vendor you want to do business with" prices for hardware)
won't permit us to come CLOSE the spec capacity of this rather expensive
space, but the space does require us to use rackmount packaging and not
towers and cheap shelving (and I am a LONG TIME FAN of towers and cheap
shelving when they will suffice for you, don't get me wrong:-).

So you see, I was if anything generous by omitting the cost of the space

Besides (and I repeat) -- the REAL point of my overall reply is:

  Your Mileage May Vary

It is just (and forgive me, this isn't intended to be a flame) silly to
make a statement like "dual packaging never makes sense" in a universe
of applications and system installations filled with nonlinear
constraints in cost-space (like the cost of the space in which the nodes
must be located).  All your reanalysis of the cost/benefit below COULD
show is that perhaps there are configurations for which is it six of one
or half a dozen of the other (or for which single packaging even
slightly wins because of the cost differential for CPUs -- although from
MY local "trusted vendor" their current price is $310 for XP 2000's,
$250 for 1900's, with fan, which is not tremendously different from the
MP cost).

In other words, sure, sometimes single packaging makes sense.  It did in
my LAST cluster purchase last year, when they still hadn't decided to
renovate the server room for us.  So my last cluster is lovely towers on
Home Depot shelving.

Sometimes it doesn't.  As in now, they did, and floor space is suddenly
dear. And I'm the same guy running the same code, and sometimes even
hand-build my systems out of component parts.  Think about how many
other permutations of individual needs there are out there.  Some folks
have never even heard of pricewatch and would only THINK of buying
turnkey rackmount clusters or prebuilt systems, and in both of these
cases I think you'd find that street price favors the dual, per CPU, by
hundreds of dollars, as LABOR costs for ASSEMBLING a dual are ALSO
roughly $50-100 less compared to two singles.

My point.

> to look at the difference between XP and MP proc costs currently a
> difference of 90 bucks per cpu on pricewatch.  I think if you re-evaluate
> your calcs based on that you get something like:
>                      Single            DUAL
> MB (w/NIC)            150                220
> CPU (1900)            180                239 X 2
> case                   50                 75 (50+ 25 some super PS and
> cooling)
> memory(512DDR)        125               125 *2
>                    _________            __________
>                      505                1023
>                       X2
>                      1010      vs       1023
> As far as I see it is awash price wise.  Most dual CPU AMD setups I have
> seen in rack mounts exceed $2000 easy as they require exotic cooling
> measures.  On the other hand, there will be some cost penalty associated
> with switch costs having more single CPU's.
> I agree with your response to the application and effective throughput and I
> think you have to run your program before you actually know how it will
> perform on these things...looking at benchmarks can be very misleading.
> I have not seen any benchmarks on the dual AMD setups comparing them to
> singles, have you?

I've DONE a rather large suite of benchmark test on dual AMDs, and
actually posted them to the list back when I was testing them on a
loaned (thanks ASL!) dual Thunder with two 1200 MHz Tbirds.  As one
might expect, cpu bound processes run at full speed, and memory bound
processes collide to a greater or lesser degree depending in fair detail
on what you look at.  Dual streams definitely show some effects of
memory bandwidth saturation.

> That was quite response you posted!

A legend in my own time, I am, although I've been relatively quiet in
recent months;-)


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