The emergence of data-intensive science as the fourth science paradigm has posed a data deluge challenge for enacting scientific work-flows. The scientific community is facing an imminent flood of data from the next generation of experiments and simulations, besides dealing with the heterogeneity and complexity of data, applications and execution environments. New scientific work-flows involve execution on distributed and heterogeneous computing resources across organisational and geographical boundaries, processing gigabytes of live data streams and petabytes of archived and simulation data, in various formats and from multiple sources. Managing the enactment of such work-flows not only requires larger storage space and faster machines, but the capability to support scalability and diversity of the users, applications, data, computing resources and the enactment technologies. We argue that the enactment process can be made efficient using optimisation techniques in an appropriate architecture. This architecture should support the creation of diversified applications and their enactment on diversified execution environments, with a standard interface, i.e. a work-flow language. The work-flow language should be both human readable and suitable for communication between the enactment environments. The data-streaming model central to this architecture provides a scalable approach to large-scale data exploitation. Data-flow between computational elements in the scientific work-flow is implemented as streams. To cope with the exploratory nature of scientific work-flows, the architecture should support fast work-flow prototyping, and the re-use of work-flows and work-flow components. Above all, the enactment process should be easily repeated and automated. In this thesis, we present a candidate data-intensive architecture that includes an intermediate work-flow language, named DISPEL. We create a new fine-grained measurement framework to capture performance-related data during enactments, and design a performance database to organise them systematically. We propose a new enactment strategy to demonstrate that optimisation of data-streaming work-flows can be automated by exploiting performance data gathered during previous enactments.