Many large-scale scientific computing projects share a data archive and retrieval requirement, a call to perform customized modeling tasks, and a need to visualize data in a geospatially referenced manner. Today, researchers generally design and procure hardware and software to provide the above services for each individual project. The rocketing cost of duplicate development and implementation, and the difficulty to collaborate due to the lack of standardization for working process and data format often leave something to be desired. This paper presents a service-oriented architecture for integrating software and hardware resources in a unified manner in order to deliver highly reconfigurable, on-demand customized scientific computing tasks. A practical system has been developed to evaluate the proposed architecture.