Ensuring that a job executes at the desired timeframe in a Grid environment is possible only when reservation policies are employed. Such policies often require users to predict job needs in terms of duration and CPU capability. However, the efficient use of resources can be jeopardised by the low accuracy of user predictions.
This paper proposes: (i) an automatic support for detecting the job characteristics and timing requirements, and (ii) a novel admission control policy relaying on such a knowledge and some indications on credential and reputation that requests could bring. The developed reservation component has been integrated into Globus Toolkit in a way that let users perform just allocation, when reservation is unnecessary or impossible.