[Beowulf] FAWN

Ellis H. Wilson III ellis at runnersroll.com
Thu Jan 12 11:40:15 PST 2012


On 01/12/2012 02:08 PM, Lux, Jim (337C) wrote:
> -----Original Message-----
> From: beowulf-bounces at beowulf.org [mailto:beowulf-bounces at beowulf.org] On Behalf Of Ellis H. Wilson III
> Sent: Thursday, January 12, 2012 10:26 AM
> To: beowulf at beowulf.org
> Subject: Re: [Beowulf] FAWN
>
> On 01/12/2012 01:22 PM, Lux, Jim (337C) wrote:
>> But it's a whole 'nother twist on the idea of clustering of low
>> performance nodes (by some metric.. they've got good
>> nanojoule/operation
>> metrics) .
>>
>
> Not just good, from a sorting perspective, /best/:
> http://sortbenchmark.org/
> -------------
>
> I was thinking that their low powered nodes are poor in an absolute performance standpoint (i.e. MIPS), but actually quite good on a computation work per joule basis.
>
> Yes, for sorting, they are kicking rear.
>
>
> This is interesting, but when you start talking power consumption, one needs to be careful about where you draw boundaries and what's "in the system".  Do you count conversion efficiency in the power supply?   At one level, you say, no, just worry about DC power consumption, but even there.. is it at the board edge, or at the chip?  Something drawing 100Amps at 0.5V is a very different beast than something drawing 10Amps at 5V, and you can't locally optimize too far because your choices inside box A start to affect the design and performance of Box B and Box C.
>
>
> The contest rules point to a variety of power measurement systems, but based on what I see there, I think there's some scope for "gaming" the system. It sort of seems it's "wall plug power", but then, they do allow DC power systems.
>
> For instance, one could tune the power supply for the expected load conditions..  You could run those fans at warp speed before the test run starts to cool down as much as possible, and then slow them down (saving power) during the run, maybe even letting the processor get pretty hot.
>
> Sort of like running a top fuel dragster. Only has to go fast for 3 or 4 seconds, so why bother putting in a water pump.

All fair points, and I can't contest the suggestion that they likely 
tune their algorithm and physical units very highly to perform well for 
this sorting environment.  Dave actually keeps a pretty balanced 
perspective when discussing this, as shown in his reaction to Google 
talking down wimpy nodes.  Wired has a nice article on it, with inside 
it a link to Googles pub that discusses the other half of the coin:

http://www.wired.com/wiredenterprise/2012/01/wimpy_nodes/

Some more reading material for the weekend ;).

ellis



More information about the Beowulf mailing list