[cap-talk] Capability accounting

[Ian G]

Jed at Webstart wrote:
> At 04:12 AM 6/23/2006, Ian G wrote:
> 
>>...
>>
>>Also, bear in mind that the typical charge would not be
>>on a "CPU processing" basis, that's a more or less old
>>notion that may harken back to the mainframe days when
>>CPUs were the bottleneck and that's all there was to
>>it.
> 
> 
> Ha!   Would it surprise you to know that we still account
> on a CPU processing equivalent basis at scientific computer
> centers today - e.g. NERSC:
> 
> http://www.nersc.gov/
> 
> where I work?  Of course there are equivalents for different
> processors, for short and long term storage, etc., but
> CPU time is still the dominate factor in figuring so-called
> "SRU"s (System Resource Units).
> 
> In so far as we do "GRID" computing (where people move
> around between such scientific computing centers and even
> have jobs automatically scheduled to appropriate centers
> with suitable resources - for a price) there is a "market"
> of some sort in place for such computing resources.


Funny!  Grid was a very big buzzword last year,
it seemed to have quietened down this year.

>>In proper marketing, you charge what you can, and you
>>align it to events rather than "pure consumption", because
>>that is more tractable in code.  But you try and simulate
>>as much as possible consumption because this gets you
>>closer to what the economists call "consumer surplus".
> 
> 
> Of course in the scientific computing market (such as it
> is), the "money" is almost entirely government "funny"
> money.  I don't know of any companies in the timesharing
> business any more, but scientific computing users do
> "shop around" for the best services among centers.

In ideal place for an internal currency.  Mind
you I'd be cautious of recommending it to a
government funded area :)

> 
> 
>>So the idea of:
>>
>>    "a new cap requiring a new coin, but you can
>>    keep using the cap forever."
>>
>>Is efficient, on paper at least.
> 
> 
> Sorry, you lost me above.


It's efficient because there is one event at
the beginning which has one single simple
charging action attached to it.  No meters
needed.  At the end there is a "reverse event"
of cleanup+reclamation.

(With a bit of handwaving) this avoids the
costs of the meter and the billing process.
What it does run slap bang into is the decision
threshold which Nick is talking about.  If you
ask the user "shall I pay for this Cap, it
costs 0.001 cents" you just blew it, because
you consumed 0.01 cents of the user's time.


>>Over time, the demand and the supply cross and efficiency
>>is achieved as you muck around with different charging
>>experiments.  How this is done is arcane, you probably
>>recall how Paypal was originally free - but over time
>>and dozens of migrations, little fees popped up here
>>and there.
> 
> 
> Seems to be true of a lot of the dot coms that survived.


Yes.  It's an evolution thing.


>>Managing resources *is* what money does - it is only
>>what money does - according to some views :)  The
>>reason we talk about money rather than accounting
>>is that in the old days, you could have an accounting
>>program, but you couldn't have a money system.
>>
>>It wasn't until David Chaum's ideas that this issue
>>changed in people's minds.
>>
>>Now you can have both.  And the money system is much
>>much better, albeit somewhat complicated to learn,
>>which somewhat inherent in money itself.
>>
>>( I personally haven't used an accounting system for
>>about 3-4 years, ever since my company issued its
>>own money.  Not that I have much to account for,
>>these days, but I can attest that the difference
>>between using an internal money and an accounting
>>system is staggering, if you can get it to work. )
> 
> 
> Fascinating.  I would certainly like to better understand
> what you mean by the above.  Would you suggest that
> changing to a "money system" would have a beneficial
> effect (e.g. more "efficient" resource use) on scientific
> computing?

Sure, absolutely.  On paper, in principle.


> By that "money system" term, would you
> consider it sufficient to have exchange rates between
> units like NERSC's "SRU"s and the equivalent at other
> centers, or would something more fundamental need to
> be done?  If the later, can you sketch the sort of transformation
> that you believe is needed?

There will be exchange rates, the trick is not
to manage them.

> Can you briefly suggest how old "big iron" things like
> processor time (which is still the primary scarce resource
> in such scientific computing systems) might be managed
> with a money system?  Naturally if there is a serious
> improvement possible (especially if there are extant models
> to view as examples) then I could see trying to introduce such
> a change to centers like ours.


Canonically, each "owner" of a big resource, such
as a data center would issue their own currency.
CrunchBux if you will.  They would then set a price
in CrunchBux for each job.  An hour on the big one
costs maybe 1000, an hour on the little ones 100
each.

That creates a "demand" for CrunchBux, assuming
that there is a demand for those machines.  Now,
the problem is we need a "supply".  That's the
business-wise tricky part.  Here's some ideas:

    * everyone who works in the center is paid
      CrunchBuxs.  Obviously they aren't going to
      buy groceries down at the Whole Foods (if
      that's what it is called) so what the sysadm
      has to do is trade his CrunchBux for a pocket
      full of green notes.
      So he posts on the trade list "hey, any offers
      for 1000 CrunchBux?" and some scientist puts in
      an offer for $3000.  A little bit of trading,
      and the trade is done.

    * CrunchBux are actually issued by the center, and
      the center is "owned" by the govt. and the govt.
      does all that stuff through the grants agency.
      Easy - pass all the CrunchBux across to the grants
      agency, and have it allocate the CrunchBux to
      the various grantees.

    * Go to Fujitsu, and say, hey, we want you to sell
      us your biggest iron possible, for a knockdown
      price.  And, to sweeten the deal, we'll give you
      CB10000000.  So Fujitsu does the deal, and then
      passes the crunchies over to the research dept.,
      which then gives grants to various science agencies,
      and thus directs the research it needs done in order
      to get the new chips built for the next gen big iron.

Now, I should say that there are a few devils in the
details.  As written, and it can be polished up, it
sounds enticing, like DSR.  But those devils are
really tricky.

iang