cost vs value Re: [Beowulf] standards for GFLOPS / power consumption measurement?

[Jim Lux]

----- Original Message -----
From: "Vincent Diepeveen" <diep at xs4all.nl>
To: "Douglas Eadline - ClusterWorld Magazine" <deadline at clusterworld.com>;
"Ted Matsumura" <matsumura at gmail.com>
Cc: <beowulf at beowulf.org>
Sent: Tuesday, May 10, 2005 5:37 AM
Subject: Re: [Beowulf] standards for GFLOPS / power consumption measurement?


> How do you categorize second hand bought systems?
>
> I bought for 325 euro a third dual k7 mainboard + 2 processors.
>
> The rest i removed from old machines that get thrown away otherwise.
> Like 8GB harddisk. Amazingly biggest problem was getting a case to reduce
> sound production :)
>
> Network cards i got for free, very nice gesture from someone.
>
> So when speaking of gflops per dollar at linpack, this will destroy of
> course any record of $2500 currently, especially for applications needing
> bandwidth to other processors, if i see what i paid for this self
> constructed beowulf.

Essentially, then, you've invested your time (which you cost at zero) to
find low material costs.  This is no different than scrounging at surplus
places, swap-meets, etc.  What you've described is a fine example between
what you "paid" for something (the "cost") and what it's worth (the
"value").    Bear in mind, also, that the value (in a market sense) of a
used mobo is not the same as a new one, in a box, with a manufacturer
warranty, etc.  All those ancillary things (like the warranty) have some
non-zero value in the general case (although you might make a decision that
in your specific case, it doesn't have value... de gustibus non est
disputandum)

Most of the time, if you're doing comparisons, you need to establish ground
rules for the cost basis.  Typically, people use the "catalog price" or
"current street price" for these sorts of things.  I think that ClusterWorld
did this for their "value cluster".  They were given some of the hardware,
but they valued it at the going price for that hardware, which seems fair.

Another way to look at pricing (for $/GFLOP kinds of comparisons) is what
the required selling price of the completed system would be, if they were in
production.  This keeps you from using pricing based on "free labor" and
"surplus deals", because such things are "one time" events and can't be used
in a production system.

A friend of mine was telling me that he had built some little boxes for his
(very young) sons that had a battery, and some lights and switches. Maybe
$50 worth of scrounged parts, all told.  People would tell him: "Those are
great, you should sell them." and then, he'd have to explain that the
selling price would be more than a thousand dollars, and he'd say "Would you
pay $1500 for that?", and of course, they wouldn't.  The problem is that the
catalog price of the parts is probably $250. And, he'd have to pay someone
to assemble them, pay all the other costs of running a business (which is
what he does in his "real job", so he's well aware of it), market them, and
so forth.  In the manufacturing company where I used to work, we assumed
that the selling price (retail) would be 10 times the raw cost of the
components.

Also, if you compete on $ spent/GFLOP, it would open you up to all sorts of
"gaming" the system.  Uhh.. This cluster "fell off the back of a truck", so
it cost nothing (aside from the risk of time in prison), so I've got the
$0/GFLOP position staked out.  Beat that!!