[Beowulf] Earthquakes and raised floors...

[Justin Moore]

> Power comes in the bottom-rear of the racks, and network comes in the
> top-rear.  We have nearly no top-bottom homogeneity of ambient temperature
> in the racks, due to the up-down airflow and the large area in the front of
> the racks that allows much mixing.  I say "nearly no" because the top 2-3
> 1U machines *are* a few degrees C warmer, likely because of recirculated
> hot air *inside* the racks.
>
> Mind you, we have all 42U occupied in the racks (20U of machines in the
> bottom & top, and 2U space in the middle with a blanking panel & a 1U GigE
> switch).  So the power supply & the homogeneity of ambient temperature are
> important.

    Actually, homogeneity of ambient temperature doesn't save you 
anything on cooling, and in fact it might make things worse.  What we 
should be striving for is minimizing hot air recirculation, since that 
introduces inefficiencies into the cooling cycle.

    What you want to do is keep your hot exhaust flows hot and your cold 
inlet flows cold.  It's OK if part of your data center is toasty, just 
as long as that hot air is flowing directly back to the CRAC units.  In 
other words, if you address the cause of heterogeneous server inlet 
temperatures -- hot air recirculation -- as opposed to the symptoms, 
you'll be much better off.

    I had a paper in Usenix last year that examined scheduling from a 
cooling point of view, and it discusses the "homogenous ambient 
temperature" argument.

-jdm

Department of Computer Science, Duke University, Durham, NC 27708-0129
Email:	justin at cs.duke.edu
Web:	http://www.cs.duke.edu/~justin/