(Message inbox:16)
Replied: Sat, 11 Feb 89 16:02:58 +0000
Replied: ai-people@uk.ac.ucl.cs
Replied: geneticalg@uk.ac.ucl.cs
Received: from Cs.Ucl.AC.UK by pyr1.cs.Ucl.AC.UK   via List-Channel  id aa06710;
          10 Feb 89 17:20 GMT
To: ai-people@uk.ac.ucl.cs
cc: geneticalg@uk.ac.ucl.cs
Subject: Distributed Resource Management
Date: Fri, 10 Feb 89 17:11:34 +0000
From: Jon Crowcroft <J.Crowcroft@uk.ac.ucl.cs>


I have been playing with a notion for some time for a
distributed resource management system, based round some very simple
tools, but with the possibility for some serious learning system to be
built.

We have:
a number of workstations.
a number of fileservers.
a number of users.
a number of non-trivial processes.
a number of non-trivial data files.

We want to optimise:
global utilisation
user response

We can form a linear combination of the (weighted) load indices for a 
process-server/fileserver, and decide for each new process/file, its
optimal initial placement - *in the absence of change*

[Aside - rstat + ruses + rexec are enough to construct this info in an
afternoons programming - i have one i prepared earlier].

In the presence of change, we would like a system to *evolve* the
weights and possibly a non-linear function of some of the load indices
(such as process size/type etc etc).

A genetic alg would suit [such as the one constructed with the use of
some old clothes hangers and sticky back plastic that always
seems to lurk at the bottom of the cutlery drawer], with each user 
literally being a member of
the poulation - their resource manager would record the
weights+function (possibly coded as the first few terms in a taylor
series expansion). 

A global manager would pick these weight/function 
records up, do the usual GA thing,
(payoff/reward function is global utilisation + user response weighted
(possibly to be fed into another GA later)),
and redistrbute them to the users for the next generation of commands.

[Brings to mind the iterated prisoners dilemma of D. Hoffstadter or
whathaveyou]

An experiment could be run with the help of undergrads on the
teaching sun's, a paper written, and a few pints downed in 
celebration of the joys of C 'n Unix

any contributions (on the back of a postcard please)

jon
