Maintaining quality of service with dynamic fault tolerance in fat-trees

Abstract

A very important ingredient in the computing landscape is Utility Computing Data Centres (UCDCs), large-scale computing systems that offer computational services to concurrently running jobs through virtual servers. As UCDC systems increase in size and the mean time between failure decreases, it is becoming an increasingly important challenge to expediently tolerate failures (dynamically), while distributing the effects of the failure amongst the virtual servers according to their service level agreements. We propose and evaluate a strategy for offering predictable service in fat-trees experiencing faults, by reprioritising packets. The strategy is able to distribute the effect of network faults in order to satisfy a number of quality-of-service demands. Which demands to favour depends on the computer system and the characteristics of the jobs it is running, and in the presence of a moderate number of faults it is to some degree possible to meet the demands. © 2008 Springer Berlin Heidelberg.