As the size of typical industrial strength information systems continues to rise, particularly in the arena of Internet based management information systems and multimedia servers, the issue of managing data distribution over clusters or ‘farms’ to overcome performance and scalability issues is becoming of paramount importance. Further, where access is global, this can cause points of geographically localized load contention to ‘follow the sun’ during the day. Traditional site mirroring is not overly effective in addressing this contention and so a more dynamic approach is being investigated to tackle load balancing. The general objective is to manage a self-adapting, distributed database so as to reliably and consistently provide near optimal performance as perceived by client applications. Such a management system must be ultimately capable of operating over a range of time varying usage profiles and fault scenarios, incorporate considerations for communications network delays, multiple updates and maintenance operations. It must also be shown to be capable of being scaled in a practical fashion to ever larger sized networks and databases. Two key components of such an automated system are an optimiser capable of efficiently finding new configuration options, and a suitable model of the system capable of accurately reflecting the performance (or any other required quality of service metric) of the real world system. As conditions change in the real world system, these are fed into the model. The optimiser is then run to find new configurations which are tested in the model prior to implementation in the real world. The model therefore forms an evaluation function which the optimiser utilises to direct its search. Whilst it has already been shown that Genetic Algorithms can provide good solutions to this problem, there are a number of issues associated with this approach. In particular, for industrial strength applications, it must be shown that the GA employed can provide reliable and consistent performance. This paper investigates evolutionary computation techniques, comparing results from genetic algorithms, simulated annealing and hillclimbing. Major differential algorithm performance is found across different fitness criteria. Preliminary conclusions are that a genetic algorithm approach seems superior to hillclimbing or simulated annealing when more realistic (from a quality of service viewpoint) objective functions are used. Further, the genetic algorithm approach displays regions of adequate robustness to parameter variation, which is also critical from a maintained quality of service viewpoint. ©1999 American Institute of Physics.