DEC 2 1974
7 SLEDERBERG>ALLOCATION@POLICY.SUMEX38 FRI 29=NOVe74 1216PM

Further thoughts om economics of SUMEX,

Since we will he giving out tokems with one handy and collecting with
the others, there are ibvious limitations tn analogizing the liberty
of consumer praterance om SUMEX and on a tree market, It will then
helo in oullding am ecomamie model to clarify Just what we are trying
to optimize, I will make a fisrt attempt at this, amd then suggest
that we May be able to dispense with a tokem economy (Cand its
monetrivial costs ot administration, megotiation etc.) im favor of a
centrally administered patterming of priorities,

Certainiy we have no objective PROFIT fumetion to maximizes though {n
the long run SUMEX shoyld be operated AS IF it maximized the {ncame
that it could extract from users at their own valuation of marginal
Servicesse {e@,s what the traffic could bear == what the service is
worth to trem in aggrenate,

Howaver, Health Kas heen orofessiomalized, and HEALTH=RESEARCH
nationalized long since, amd we can get into serious trouote by
inaporooriate mixes of central planning and free narket arrangements
The MAXYT@PROFIT motion isam abstraction that can give us some
(limited) guidance to planning,

Some postulated erinciples of operation imeclude?

The SUMEX*=AIM community comprises a limited set of imvestigators who
afe tO be encouraged and sypnorted tm the pursuit of 1) explicit
research prograns and 2) pelated but less welledefined explorations,

We have a constrained budget for carital tnvestment that is the
princie¢al limitation to the overall volume of service that can be
delivered,

3, Each user will he Judaqed te have some service=value functiom which
falls asymptotically to zero (that {8s plotting the utility of the
Mext imcrement of Service aqgaimst total volume consumed), At zero
but mot necessarily to first order soeial utility may or may not
correspond to how a user would Spend his own dollar budget im a
comvertyiole currency,

4, Given the constraints of a fixed=size machine, and the mamagertal
ones Of a finite community of users we should, roughtys optimize the
integral of the product of services , utilitys 1.e,7 allocate the
mext increment of available service to the user Judged to consume it
at Ruishest utilfitv. In practice this means we must assure each user
an opportunity to get his cemtral auantum of work dome? and we must
take account of the sidewcosts of deliverina services are umeven high
loadings etc,

5S, “e@ have to ne able to justify any cvert inefficiency, idie time
etc, We alS2 have to ne accountable for various aspects of FAIRNESS,
esoecially in re the charter of a 40"d0=20 sitcing,
p SLEMDERSERG>ALLOCATION#POLICY,SUMEX]8 FRI 29"NOVe974 LrloPm

6, There aee three evident measures of services how=much (cpu); when

(tinesofeday or demand availability): amd how=fast (throughout rate)

athat will cear differentiy otn different users at different times,
These consideratuions suggest the follawing aparoach to allocation,

We will mot have tokens at all (except perhaps im re connect time,
esoecially for renote users a/e the relevant costs,) Instead a 3 or
Getier orjiority system,

PRIORITY 1, USE Guring SCHEDULED time of day, Eacr USER will be
allotted am hour per day of commect time (ar some multiple thereof)
which he can

‘reserve’ for a9 weex im advamnce by voluntary posting, This does not
interfere with overlaoping scheoules by others, mor with voluntray
Side=aqreements to avoln overlap, The point $8 to have some
franework in which users can PLAN to have the most efficient access
possible SUMEX management can also play a persuasive role im such
postings,

Users im ordomity=] status wtl!l comeete for the first 60% of
machine cycles (divided 39230 AIMSSUMEX) regardless of other quotas,

PRIORITY 2. (SYS lives here chronically), Users compete for 80% of
the RESIDUE of the machines (1.e, 24% olus spillover from priority tL,
Users will work im tris level umti! they have used up their daily
Quota of CPU time,In anditionm work Involving routine EXEC and text
and file=handling smould be ungraded from level 3 into this queue,
PRIURITY $ gets the rest, either at mare or after some further
adjustments,

This general scheme of course admits of many further tuning
Steos, Under level 3, for axample, we should consider unloading Jobs
that do mt get ersough CPL attention to warrant keeping them in the
aQueve. In the same vaim we should have some provision for autolegout
of jmactive connect lines that merely burm up communications costs,
and for Randling irmaetive forks... but we have to analyze what
Penalty these dreaas (amd their solutions) impose,

The presert oroposal is of course substantially what Rainer has
discussed and nartliy implemented, except for the overlay of
Priority=1, J] am mot averse to some gradualism in shifting between
the levels» but a user should have enough credictahility about how he
fs being manmdied py the system that he cam plam his work, rather than
just sit totally passively hoping for the best possibile, In
allocations ot Jlevelel svheduless SUMEX=management cam of course slay
a more or less active role in structuring the traffic if the
cirstances reauire (lie a treatfic com relates to signal lights and
ateo signs): amd I can eventually foresee some gamewlike algorithms
to nelo organize these schedules, The assigmment of %=ages among
differemt jievels is of course a further mamagement option,
$y <LEDERBERGOALLOCATION=POLICY.SUMEX}38 FRI 29"NOVe74 1t16PM

The proplem with tre token economy alternative is just how te
allocate the chinas to start with and the time we will spend
mecotiatina qrievanceas whemever the currency is reeflated and the
pricing system altered, If we did gp that route, we might want to
think of a comtinuovs auction to set current prices «= which is a
kimd of parody of the difficulties,

But there is Some room for chips im the present schemet 1, The
user's choice im sosting level=1 connect timenadvancemreservation, 2,
Spending CPU quota to stay im this priority 3, Overall comnectetime
limitations,

The effuciemt exploitation of the resource is certainiy going
to sequire Sone form of quasi#batch=bhackground level of operation to
Spend the monworine=time cycles, I need to know mere about connect
time costs to judge the related issue of DETach/REDIRect aetivity,

Behind all this discussion is a model of user activity that I
will oe@ tryina te make more explicvit and perhaps to simulate, (12
have wanted for some time to start some work on aoplying AI to treaty
megotiations, ana the mechanizedf induction of schjmes like this one
might be a teasonahle challenge.)

FURTHER COMMENT in re Priority | (11/29/74)

The basic logic of this arrangement jigs to find the ootimum
comoromisSe between the level of structuring that will enable users to
Plan the most efficient use of their own time, and the flexibility
that enables ad hoe response to the exigencies of their task,
Management mas avatlable the ootiens of varying the relative
allocation of priority one cycles down or up from 60%» and also of
taking a "ore active role im arranging for the staggering of suck
Scheduleo use,

It may he asSkecdse cuite seasonably, whether there will mot be an
automatic regulation of usage im the light of the diurnal eyele of
Pesoonsiveness of the system == {,0,, whethers users will mot simply
adjust their owm sehedules to what they observe im loading, This
may well Rappem; but I foresee that there will stil! be
disappointments arising from umpredicted interference, We must also
take account that many of our users will have speeial comstraints
M3e.d. imteraction with kmowledge=comsultanmts,patients,
demonstrations, etc,s that will be greatly hindered by lack of some
Scheduling structure, On the other hand, we wish to avoid an
excessively competitive framework == a rush to the starting line =«
that will distract from the actual substance of working om the system,

Anotnmer possible option will be te allow a highly timited
opportunity for trading im some other ‘assets! in favor of a URGENT
priority level that cam override the current schesules,

Only the major project directors meed to inform themselves of
the details
y SLENDER BERGPALLOCATION@POLICY,SUMEX38 FRT 29eNfive74 1316PM

of these scheduling arrangements which are rather complex, and may be
Subject to change from time to time during our shakedowm period, They
should be anle to communicate to their collaborators simple
ewidelimes abcut when to work and what orfority to request,

Another advantage of this system {ts the possibility ¢t offers of
altruistic coomerations vizer the voluntray selfeassighment of a
Peduced priority level for tasks that do not have am urgent seed for
orenot completion, Likewise, manmagewil! be ahle to emabie a wider
Pamge of orojects, for the most efficient utilization of the machine,
if some of these cam be automatically kept from interfering with
high=eriority users at busy times,

Some eariier comments an economic model

ONE OF THE OANGERS OF SETTING UP TOKENS IS THAT IT WILL PROMPT
QUESTIONS, AkKY FUNNY MONEY JNSTEAD OF "REAL" DOLLARS?

WE SHOULD RE PREPARED WITH THE FOLLOWING ARGUMENT: TO JUSTIFY TOKENS
VS, REAL MONEY, NE SHULILD STRESS THAT THIS IS AN EXPERIMENTAL SCHEME,
UNLIKE REAL MONEY, *E CAN START OVER AGAIN AND M&kKE MANY OTHER KINDS
OF ADJUSTMENTS IN THE COURSE OF PERFECTING THE PRICING ALGORITHMS
ETCae

MANY & COMPANY HAS GONE BROKE OVER ERRORS IN ITS PRICING
ALGORITHMS, JOSH