Path: utzoo!utgpu!watserv1!watmath!att!att!linac!pacific.mps.ohio-state.edu!zaphod.mps.ohio-state.edu!usc!snorkelwacker.mit.edu!mintaka!ai-lab!rice-chex!caroma
From: caroma@rice-chex.ai.mit.edu (Carl R. Manning)
Newsgroups: comp.object
Subject: Re: Stash Collection (Economic models)
Summary: Look in "The Ecology of Computation", Huberman, editor
Keywords: Economic GC
Message-ID: <12080@life.ai.mit.edu>
Date: 29 Nov 90 01:18:37 GMT
References: <2853@esquire.dpw.com> <2855@esquire.dpw.com> <1990Nov28.075224@Unify.com>
Sender: news@ai.mit.edu
Organization: MIT Artificial Intelligence Laboratory
Lines: 81

In article <1990Nov28.075224@Unify.com> grp@Unify.com (Greg Pasquariello) writes:
>I would be interesed in someone posting more information about the
>"economic model" mentioned yesterday, whereby objects pay rent, get
>evicted, etc.

A basic idea of the Agoric (market) model is that large systems which
allocate resources through markets (e.g. economies based largely on
capitalism) can often make better use of resources than those where
planners try to anticipate ahead of time what the resource needs will
be (e.g. command economies; programmers & system operators).  The
theory is that local market prices reflect the actual (not just
anticipated) locally available supplies of, and local needs for, a
resource, so resource decisions made according to these prices can
make better use of the resource.  So, as our computer systems become
integrated into larger Open Information Systems, perhaps we can best
make use of the resources by setting it up as a Market ecology, which
doesn't require anyone to understand the overall system in order for
it to evolve toward more efficient use of resources.

In order to set up markets for resources, we have to build systems
which charge for resource use: processing, storage, communications,
etc.  Part of the cost of using a resource is the overhead for
charging and collecting for use of the resource; part of the cost for
using market mechanisms is the transaction cost of negotiating prices
in the market system.  There's a tradeoff between (a) using the market
to allocate resources where they're needed most, but taking on
overhead and transaction costs, and (b) using a command-style planning
to allocate larger amounts of resources ahead of time and avoid the
overhead, but risk less efficient allocation since the decisions are
based on anticipated resource conditions rather than actual ones.

Therefore, in some situations rent will likely be charged for
resources at the `agency', `application', or `server' level rather
than at the individual object level, just as your company probably
doesn't pay rent (or property taxes) on each office individually.  In
other situations, the transaction/overhead cost of smaller allocations
is worth the benefit in flexibility, e.g. you might pay for an object
agent to migrate temporarily to a knowledgebase server to make some
queries, or to negotiate protocols for future interactions, just as
your company can pay for individual hotel rooms in another city when
it sends you to a conference or to negotiate a contract.

In such a market system, when local demand for a resource rises, the
price for that resource will also rise; those uses of the resource who
are renting at the market rate will find their costs rising, and if
the costs rise high enough, they will have incentive to cut back on
the use of that resource, e.g. put things off until later when the
price is lower, get rid of large tables of noncritical cached info,
move things to cheaper areas (disk or another machine).

The market provides *incentive* for doing these things.  The price
mechanisms provide the local information for making decisions about
what to do.  They do not provide actual mechanisms for actually doing
these things, just the competitive incentive to do them the best way
you can figure out how (or be outperformed by competitors).

So to get back to stashes... Markets mechanisms can help a system
decide when to stash and when to release the stashed info, but by
themselves won't help much in figuring out how to implement the stash
itself.

Enough rambling.  Those who are interested might take a look at 

   The Ecology of Computation
   B.A. Huberman, editor
   New York: Elsevier Science Publishers, 1988

In particular at the chapters

    "Markets and Computation:  Agoric Open Systems" by Miller & Drexler
    "Incentive Engineering for Computational Resource Management" by
      Drexler & Miller

If you find the book interesting, you might also find something interesting
in the more recent citations at the end of

    "Primed for Performance", Tad Hogg, June 1990 BYTE, p241

Cheers,

CarlManning@ai.mit.edu