1U P4 Systems

Velocet math at velocet.ca
Tue May 22 19:41:57 PDT 2001

On Tue, May 22, 2001 at 09:50:56PM -0400, Bob Drzyzgula's all...
> > I regretfully opted for shelves of midtowers for my latest effort(s) --
> > one can achieve something like (within a factor of 2, surely) of the
> > packing of 2U cases -- I can fairly easily get 16 nodes in a 1-1.5 m^2
> > floorspace footprint (depending on how you like to count the gaps
> > between units) without going over two nodes high and could double that
> > with tall shelves.

the standard floor tile is the decidedly non metric 2'x2', which is
61x61cm, or .372m^2. So you're talking about some 2.7 to 4.0 floortiles
total for 16 machines (ie 6 to 4 machines per tile). This is insanely
low density for a real machine room. If anyone has any contraints for
floor space, even 12 machines per tile (your 'double' figure), which
is directly 12 machines per 19" rack (24" on outside frame). In 44 to
48U (1.75" per U) standard racks, you can usually fit 11 or 12 standard
huge klunky 4U cases. This matches your best figure. With 2U cases, 
its double, with 1U, its quadruple (48 machines).

The cost effectiveness of standard towercases cannot be denied, if you
are not operating in an environmentally-controlled server room raised
floor environment. Those cost a fair bit to install and maintain. If this
use or construction of space is factored into your cluster cost, its going
to add up. An old empty gymnasium full of towercases is hard to deny
is the best savings, but if you are only being alloted a small amount
of space in the machine room, you cant go that way.

THis is why I asked in the previous post about custom cabinest, abandoning
the high cost of standard rack mount relay racks and cabinets (which
are incredibly overpriced as are the 1U and 2U cases, IMHO). 

> This does a good job of fitting the requirements
> of an ad-hoc or perhaps a lab environment, but in a
> production data center environment these can be decidedly
> sub-optimal. It is relatively difficult to get good cooling

exactly :) [ just putting numbers to your comment  based on the
original poster's figures]

> system density gets high. (FWIW, what we are short on is
> chiller capacity; some of our racks are gulpling for
> chilled air.)

What do you find the required capacity of the chiller to watts of power
sucked down by the machines has to be? I've heard that its harder to
get better than 1.4:1 with incredibly good design, and closer to 
1.75 or 2:1 is more typical. (just think, thats 3 watts of power
per 'watt-computing'. Man, this must just make Feynman roll in his
grave at the disgustingly inefficient thermodynamics of the situation! 
What are we working at here, factors of 20 magnitudes of heat production
per joule spent processing bits? :)

> > It's not as pretty and it is a bit more work to put together neatly, but
> > I still assembled a 16 node cluster in about four hours total work
> > including assembling the heavy duty steel shelving.  The shelving itself
> > cost only $50 total (plus another $50 or so for cable ties and surge
> > protectors and miscellany to make it look nice) saving me about $2500.
> > That's serious money for six hours of work (allowing for the time to
> > drive to Home Depot and buy the shelving:-) -- 3 nodes worth of money.
> > I'll make the time back in the first DAY of fulltime operation.

16 nodes at $60 odd canadian (guestimate) per machine is ~$1000. We're
looking at a custom cabinet for around $2000 for 40 or 50 nodes. And
sides it will look *WAY* cooler than a row of standard pc cases.
I need to confirm the quote with the metalworker, could go up to
$3000 in fact, but thats still great economy.
> want to use the casters. These take about ten minutes to
> put together if you've done it a few times and know the
> tricks. You can easily fit as many as twenty minitowers on
> one of these, and wheel it around when you're done. For
> this they want $77. Talk about the advantages of volume
> production...

Gotta have everything in standard cases to start with though!
Ken Chase, math at velocet.ca  *  Velocet Communications Inc.  *  Toronto, CANADA 

More information about the Beowulf mailing list