Due to the increasing heterogeneity, dynamicity and scale of Grid environments, one of the emerging trends in high performance computing is to design Grid services to easily and efficiently use the vast amount of computing, storage, and visualization facilities that are available at geographically dispersed sites. In this paper, we propose an extensible, generalized design model for future Grid services, called Computational Federation Service (CFS), which is based on virtualization of Grid resources in a single system image. The approach builds on and extends our previous research on Computational Federations (CF). By utilizing a Distributed Virtual Machine (DVM) platform, the CF architecture virtualizes node heterogeneity and communication protocols, automates load balancing and includes features of persistence and multitasking, while supporting a location and registration service and the dynamic attach and detach operations of resources. The CFS can exist as part of the Grid middleware, built on top of the collective Grid, resource, and connectivity layers. Our architecture is application-profile aware and can flexibly and dynamically configure and interconnect any type of Grid resources, e.g., computational and storage nodes, instrumentation, and can also extend to different domains, change in time, and support complex application workflow compositions and collaborative applications.