[Beowulf] Airflow

[Jim Lux]

At 11:51 AM 6/28/2006, Andrew Piskorski wrote:
>On Sun, Jun 25, 2006 at 03:52:27PM -0700, Robert Fogt wrote:
> > Hello beowulf mailing list,
> >
> > How much harm does removing the computer case do?  I know computer 
> cases are
> > designed for the air flow, and without them I was wondering if there 
> will be
>
>In theory, computer cases are designed for good air flow.  How many of
>them actually are, well, I don't know.  I've certainly never seen an
>airflow specification for a consumer tower case.


The airflow spec would be more that imposed by the mobo mfr, or the 
processor mfr's temperature limits.  And, most mfrs provide a "reference 
design" which generally meets the requirements. Back when I was looking at 
densely packaging a bunch of Intel mobos in the Flex-ATX formfactor, I 
found a whole series of apnotes on Intel's website on just this subject. 
Any large consumer manufacturer (read: Dell, HP, etc.) will almost 
certainly have done analysis and testing for airflow on their consumer 
machines.  The LAST thing they want is reduced reliability from 
overheating.  Given their volumes, they have to be fairly careful.




> > heating problems. My air conditioner will be enough for the amount of heat
> > generated, but will I need circulation fans in the room also?
>
>The short answer is, I don't know.  But, why NOT use some fans?
>Installing a 20 inch box fan to blow directly across each mounted
>group of N motherboards seems like cheap insurance.
>
>According to my Kill-A-Watt, my dirt-cheap 20 inch, 120 V AC box fan
>uses from 79 to 136 Watts depending on its speed.  I assume a better
>quality fan would be more efficient for the amount of air it moves.

Not necessarily.  Fan efficiencies are more a matter of packaging and 
ducting. Most people aren't too worried about efficiency in a CFM against 
pressure drop per watt sense.  They're more concerned about things like 
noise, vibration, airflow patterns, vanes, etc.   The single thing you 
could do to improve fan efficiency is to make sure that the fan is operated 
in an orifice a very small amount larger than the diameter of the blades 
(i.e. clearance in the 1-2mm area)


> > I created a test cluster of 8 nodes, removing all the cases and mounting
> > everything onto a wire rack. It works fine with no heat problems. But I am
> > concerned that once I build the entire 100 node cluster there will be
> > problems. The book I am reading on HPC does not go into that detail.
> >
> > I was thinking an air purifier that is always on will slowly circulate the
> > air when the air conditioner is off.
>
>I would not place any bets on "slowly" being good enough.  Why would
>you want to risk it?  I'd at LEAST walk around with a temperature
>probe looking for hot spots.

With 100 nodes, you're looking at 10kW.  If the AC goes off, the 
temperature will rise remarkably fast, unless you are in a VERY big room. 
RGBs online book at Duke goes into this in some detail, but in short 
strokes: if your room is, say, 3x3x3 meters = 27 cubic meters.  There's 
about 45 kilos of air in the room.   Cp for air is about 1 kJ/(kg 
K).  Simple math will get you to a temperature rise of about 1 degree every 
5 *seconds*, or 12 degrees/minute. If you started at, say, 20C, you'd be to 
a node killing 50C in, oh, about 2.5 minutes.  This is so fast that the 
thermal mass of the stuff in the room is irrelevant.




James Lux, P.E.
Spacecraft Radio Frequency Subsystems Group
Flight Communications Systems Section
Jet Propulsion Laboratory, Mail Stop 161-213
4800 Oak Grove Drive
Pasadena CA 91109
tel: (818)354-2075
fax: (818)393-6875