Investigating evolutionary approaches for self-adaptation in largedistributed databases

摘要

As the size of typical industrial strength databases continues torise, particularly in the arena of the Internet and multimedia servers,the issue of managing data distribution over clusters or `farms' toovercome performance and scalability issues is becoming of paramountimportance. The general objective is to manage a self-adaptingdistributed database so as to reliably and consistently provide nearoptimal performance as perceived by client applications. Such amanagement system must ultimately be capable of operating over a rangeof time varying usage profiles and fault scenarios, incorporateconsiderations for multiple updates and maintenance operations, and becapable of being scaled in a practical fashion to ever larger sizednetworks and databases. This paper investigates evolutionary computationtechniques, comparing a genetic algorithm, simulated annealing, andhillclimbing on a test problem in this field. Major differentialalgorithm performance is found across two different fitness criteria.Preliminary conclusions are that a genetic algorithm approach seemssuperior to hillclimbing or annealing when the more realistic (from aquality of service viewpoint) objective function is in force. Further,the genetic algorithm approach displays regions of adequate robustnessto parameter variation, which is also critical from a maintained qualityof service viewpoint