The aim of this paper has been to formulate a trustworthy framework for analyzing models of system administration. There is good cause to view computers as dynamical systems, approximated by mechanistic rules developing in time, with idealized properties which can be summarized by a finite state lattice. The theory of games has been employed in order to select between alternative strategies in a contest for machine resources, moving the state of the system through the lattice, as if on a chequerboard. It has been shown that it is possible to see system administration as the effort to keep the system close to an ideal state, by introducing countermeasures in the face of competitive resource consumption. This is the formal basis which opens the way for objective analyses in the field.
It is important to understand that, even an answer obtained with the assistance of a mathematical formalism is not necessarily the last word on the subject. Mathematics is only a tools for relating assumptions to conclusions, in an impartial way. With a mathematical approach, it becomes easier to see through personal opinions and vested interests when assumptions and methods are clearly and rigorously appraised. However, one can only distinguish between those possibilities which are taken into account. That means that every relevant strategy, or alternative, has to be considered, or else one could miss the crucial combination which wins the game. This is the limitation of game theory. It is not generally possible to determine strategies without creative input; this means that human intelligence will be required for the foreseeable future. There can be no zero-maintenance computer system. With this caution, how can one know that the ideal state of a system can be reached? How can one know that the system will not run away in an unstable spiral to catastrophe?
Two things are clear from the limited analysis here. The first is that purely dumb automatic systems are inadequate to perform every task in system administration today. Intelligent incursions are required to solve complex problems, to extend or adjust the strategies of the automatic system. Interestingly, this is the approach by which evolution has solved the immunity problem: the automatic responses of lymphocytes only go so far; the emergence of intelligence in humans has enabled us to develop medical research and develop drugs and other treatments against damage and disease. It seems naive to believe that any simple mechanistic system would be able to do any better than this; we can expect to require the assistance of humans at least until alternative machine intelligences have been developed.
The second point is that the use of quotas is a highly inefficient way of counteracting the effects of selfish users. A quota strategy can never approach the same level of productivity as one which is based on competitive counterforce. The optimal strategies for garbage collection are rather found to lie in the realm of the immunity model[2]. However, it is a sobering thought that a persistent user, who is able to bluff the immune system into disregarding it, (like a cancer) will always win against the resource battle. The need for new technologies which can see through bluffs will be an ever present reality in the future. With the ability of encryption and compression systems to obscure file contents, this is a contest which will not be easily won by system administrators.
There is plenty of work to be done on the theory of system administration. This paper is merely a small push in the direction of progress.
I am grateful to Trond Reitan for a useful discussion about evolutionary stable strategies and to Hårek Haugerud, Lars Kristiansen and particularly Sigmund Straumsnes for their critical readings of the manuscript.