Abstract
This article describes an original approach for the optimized execution of computational tasks in Grid environments. Tasks are represented as workflows that define interactions between different services. Functional service descriptions written in OWL-S are extended with non-functional properties, allowing to specify the resource requirements of services depending on given inputs. Based on such annotations, a mathematical model is derived to estimate the execution costs of a workflow. Moreover, an optimization algorithm is presented that distributes the execution of a workflow in a Grid, supporting the dynamic deployment of software components on demand, in order to fulfill user requirements, such as a limit on the total workflow execution time. Workflows are executed in a fully decentralized way, avoiding inefficient triangular routing of messages.