[Beowulf] power consumption for IBM J20 blade or Apple Xserve?

[Robert G. Brown]

On Tue, 8 Jun 2004, Mark Hahn wrote:

> Hi.
> we're looking into building a number of clusters, some of which 
> have serious power constraints.  rather than just biting the bullet
> and buying fewer dual opteron/xeon nodes, we'd like to consider 
> lower-powered processors.  as it happens, those clusters will focus
> on cache-friendly FLOPS; this makes the PPC970 look especially attractive
> 
> the IBM PPC970fx is rated at 24.5W dissipation at 2.0 GHz (theoretical 8
> Gflops), but does anyone have actual measurements from systems using it?
> I presume IBM's JS20 blade uses the fx chip; if so, does anyone have power
> measurements for it?  IBM's website seems to be sanitized of anything 
> past PHB levels of vapidity.
> 
> alternatively, does anyone have power measurements for Apple's Xserve?
> Apple.com also has plenty of pretty marketing info, but little non-spin...
> 
> our config would likley be 1 GB/CPU, single small disk, and for one cluster,
> a quadrics/myrinet/infiniband nic.
> 
> thanks, mark hahn.
> (yes, I'm aware there are lower-voltage/power opterons available, but they 
> compromise clock speed, cost noticably more, and are unavailable from most
> vendors.)

Hey, Mark.

Why don't you call a vendor or two, and see if they won't loan you a box
of each to play with.  Then you can put it on your very own kill-a-watt
and run your very own benchmarks for performance estimation.  Tell 'em
you'll post the results to the beowulf list and that the numbers will
probably get cited in Cluster World's "best of the beowulf list" monthly
column -- free advertising, if the numbers are good ones at any rate.
I'd guess that Jeff would be amenable;-).

I'd certainly be interested in hearing what they are, as there may well
be several clusters on campus here in the future with serious
power/cooling constraints. Besides, power+cooling costs roughly $1 per
watt per year in a cluster (and maybe even more in the future as energy
prices rise) and when one has hundreds of nodes drawing tens of KW of
power, fractional reductions can add up to real money in operating cost
savings over three years.  At the very least this cost savings needs to
be applied against the hardware cost in a CBA at purchase-decision time
as much as the energy/heat savings can factor into the engineering
decisions.

   rgb

-- 
Robert G. Brown	                       http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567  Fax: 919-660-2525     email:rgb at phy.duke.edu