[Beowulf] Re: The Walmart Compute Node?

Mark Hahn hahn at mcmaster.ca
Fri Nov 9 20:23:46 PST 2007


>> Most of the clusters I end up building or working on (academic,
>> government and corporate sites) are intended to support periodic
>> spikes in computing demands.
>
> Exactly.  It's not just academic.  This mode of parallel operation comes
> up any time the user needs the result to proceed with a project, and
> either their time has value or time is critical.  So cutting a job
> down from 2 hours to 6 minutes even occasionally can be justified
> economically even if the cluster sits idle the rest of the time.  BUT,

we live in different worlds.  we support spikes in demand by interleaving
them in a large enough pool of resources.  sure, there's no guarantee 
that spikes won't conflict.  do you really have users who are willing 
to pay for a whole cluster and only use it for a 5% spike of the time?

> while it is sitting idle you'd naturally like it to burn as little power

uh, you have idle clusters!?!  to me, that means you don't have a large
enough or perhaps diverse-enough user community.  or else your
time-to-solution is orders of magnitude higher than I see.  in a 
very rough sense, it looks like it costs me about 1% of a cluster's
purchase cost to power it for a year.  or to put it another way:
perfect power savings would save a handful percent of cost over the 
cluster's lifetime.

> power on the processor to something negligible.  Clearly it would be
> better if the AMD and Intel processors kept their capabilities and added
> equivalent low power modes.

the latest gen are pretty impressive that way, at least according to 
reports on desktop versions.  no idea whether server-oriented models 
(and their bioses) support so well.

>> Would it be interesting to open a discussion how to reduce sound of
>> cheapo clusters as well?

eh?  strip out the crap fans and stick in one larger decent one.
yeah, cheapo cluster will likely to have ugly hacks like active 
cpu heatsink-fans, but quiet versions can be had even there.

> isn't "cheapo", the problem is peak power and getting rid of it.  Sound
> insulation is very closely related to heat insulation.  The net result
> is that it's relatively easy to cool a cluster, or to muffle its sound,
> but fiendishly difficult to do both at the same time.  The cheapest

I'm not sure I agree.  good cluster design in my book is simple 
airflow - ideally, I'd like cold air ducted straight into the front
of racks containing passive 1U-ish blades tightly sealed to prevent
leakage.  back of racks flow straight into chiller/fancoil intake.
ideally, I'd also dump the heat into a big in-ground thermal sink
rather than running a big bank of condensers outside.

it's all about the fans - get rid of as many as possible, and trade
little ones for fewer big ones.  compressors on our chillers are not 
silent, but trivial compared to the 12 fans per node.



More information about the Beowulf mailing list