Building Online Performance Models of Grid Middleware with Fine-Grained Load-Balancing:

Abstract. As Grid computing increasingly enters the commercial domain, per-formance and Quality of Service (QoS) issues are becoming a major concern. To guarantee that QoS requirements are continuously satisfied, the Grid middleware must be capable of predicting the application performance on the fly when de-ciding how to distribute the workload among the available resources. One way to achieve this is by using online performance models that get generated and an-alyzed on the fly. In this paper, we present a novel case study with the Globus Toolkit in which we show how performance models can be generated dynam-ically and used to provide online performance prediction capabilities. We have augmented the Grid middleware with an online performance prediction compo-nent that can be called at any time during operation to predict the Grid perfor-mance for a given resource allocation and load-balancing strategy. We evaluate the quality of our performance prediction mechanism and present some experi-mental results that demonstrate its effectiveness and practicality. The framework we propose can be used to design intelligent QoS-aware resource allocation and admission control mechanisms.