[Beowulf] machineroom design

Mark Hahn hahn at physics.mcmaster.ca
Wed Jan 10 07:04:03 PST 2007


I just had an episode in my machineroom where a small perturbation
in the heat load (turning off a handful of machines) caused a 30T
chiller to ice up and become mostly nonfunctional.  this was doubly
perplexing because there's plenty of load to keep it working.
the main factor was that we recently (well, month or to ago) 
remove some nasty plastic tarps which were trapping cold air inside
a cold aisle.  apparently this altered airflow enough to push a lot
of cold air towards the chiller.  these Liebert units have just one 
intake sensor, and it's nearly on a corner - out of the cold flow.

easy thing to fix, but reinforced to me once again that the canonical
hot/cold aisle approach is actually _not_ a good idea.  at the very 
least, you want something blocking the path from the cold aisle to 
the chiller intakes.

a 2-plenum design (say, cold underfloor and drop ceiling for return)
could certainly avoid this problem, assuming there's no large gap
between the ceiling and top-of-rack.

but in our case, it would have made a lot more sense to simply make 
a single row of compute racks, with their hot little bums mooning 
the row of chillers along one wall.  that's the only airflow that 
really matters, and the real problem with the current hot/cold setup
is the numerous possible bypass and counter-flows.

in short: don't build hot/cold machinerooms unless you control
both the cold outflow and hot intake locations quite carefully.
at the very least, plan to block off the end(s) of cold aisles,
since any flow out of them that doesn't go through machines is wasted,
and quite possibly problematic.

for raised floor, a simple row of machines facing away from chillers 
is a lot nicer behaved.  if you can't fit the row, consider folding 
it into sort of a W-shaped structure that still puts racks between 
cold air outflow and chiller intakes.  most servers these days generate
a pretty powerful jet of air out the back, and pointing them at least
partly towards the chiller intakes is certainly helpful.

regards, mark hahn.



More information about the Beowulf mailing list