[Beowulf] Re: Setting up a new Beowulf cluster

Mark Hahn hahn at mcmaster.ca
Wed Feb 13 14:16:55 PST 2008

>> Your cluster will need a home, and there are good homes and not
>> so good homes depending on its scale.

and design.  a noisy cluster is one that's trying hard to stay cool.
a non-noisy cluster is not trying hard - perhaps because it has 
low-power processors, or because its intake air is cold,
or perhaps it's turned off ;)

there are some cluster nodes that are just unnecessarily noisy - 
they run their fans at 100% all the time, or have too many fans,
or poor design.  in general, you want fewer big fans, preferably 
one large impeller-type fan rather than a dozen tiny muffin fans.
a taller chassis can make it easier to move the requisite air,
but remember that cooling requires moving the air in one side
and out the other.  intra-case circulation is no real help.

the one-side-to-other-side thing is what gives rise to front-to-back 
server chassis and hot/cold-aisle machineroom layouts.  but the 
best machineroom circulation is simple and well-controlled.
don't introduce h/c aisles if you can just have a single row 
with a single hot side and a single cold side.  if you can
arrange for both hot and cold plenums, do, since the goal is to 
contol the airflow, not let it do what it wants.  for instance, 
building a physical partition - drapes or even a wall, between
hot and cold regions is a very good thing.  leaving space above
racks that permits hot air to travel to the rack-front is very bad.

> too much noise.  Did I mention NOISE?  There is no inexpensive way to
> quiet down a rack because to first order sound insulation == heat
> insulation.  It can be quiet, or it can be cool, but it is really hard
> to do both.

if you take it as given that nodes are noisy, I suppose.

but it's easy to imagine a large cluster where nodes are actually passively
cooled or have quite slow fans.  but the machineroom would need exceptionally
good partitioning of hot and cold sides, and (if nodes are fanless) would
need some hefty air-moving hardware somewhere.  in this imagined layout,
you'd really have a hot room and a cold room, with the fronts of the racks in
cold, and the backs in hot.

putting a dozen 15k rpm fans in each node is the noisiest possible way
to move the air.  a few very large fans could be quite quiet, and 
not involve much in the way of either sound or heat insulation.

there are some instances of using heatpipes to get heat out of the node
and into chilled water or refrigerant.  it's a little unclear to me why
this is not done more, especially among the blade-loving community.

> A normal framed wall may be enough to
> subdue most of the noise for a smallish machine room, so that the rooms
> on either side can be used for normal tasks.  (The low frequencies may

the wall between my office and our machineroom is duplex, and it 
works well.  the machineroom is pretty noisy, but most of the noise
leakage is from the door.  across the hall is a conference room,
which has the same kind of duplex wall shared with the machineroom, 
and it's quite nice.  we have those 1-sq-ft ceiling tiles on one wall to
deaden the room a little, but that's mainly to improve the teleconf 

I guess most of my comments were for fairly large clusters.  for small
clusters, I'd probably first try to avoid generating much noise.  it still
makes sense to aim for a simple, well-controlled circulation pattern,
but if you can afford the space to use desktop machines, for instance, 
the result might be quieter than 1U screamers.

then again, a quiet cluster is less impressive 
when it comes time to impress the visitors ;)

More information about the Beowulf mailing list