Large-scale grids and other federations of distributed resources that aggregate and share resources over wide-area networks present major new challenges because they couple the behavior of resources and networks. These infrastructures support a new breed of applications which interact dynamically with their resource environment, making it critical to understand dynamic application and resource behavior to design for performance, stability, and reliability. Coupled use means that accurate study of dynamic applications, middleware, resource, and network behavior depends on coordinated, accurate, and simultaneous simulation of all four of these elements. Thus, the long-term challenge is to support scalable, high-fidelity, online simulation of applications, middleware, resources, and networks to support enable scientific and systematic study of grid applications and environments. That challenge is the focus of this dissertation. We define the problems in performing large-scale, high-fidelity, online simulation. We consider a number of approaches, and then present our approach in detail. Our approach includes a set of techniques which enable the use of real application and middleware software, and modeling of essentially arbitrary network and resource properties. These techniques include resource virtualization via application interception, computation resource simulation based on soft real-time scheduling, and packet-level online network simulation. Our studies and experiments show that these techniques can support simulation experiments with complex software packages as well as resource and network structures. While most of the techniques in our approach are inherently scalable, one major challenge is online network simulation – which we implement as a parallel distributed discrete-event simulation, well-known to be challenging to scale. A range of techniques for scaling our online network are studied. Exploiting advanced graph partitioners, we explore a range of edge and node weighting schemes based on a variety of static network and dynamic application information. While simple approaches do not achieve acceptable load balance, our studies show that detailed network structure and behavior can be combined with the graph partitioners to achieve both good load balance and parallel efficiency. For example, our improvements increase efficiency and scalability by over 100 times, achieving a parallel efficiency of over 40\\% on 90-node clusters for a range of experiments. Our online simulation techniques are embedded in a working simulation tool, the MicroGrid, which enables accurate and comprehensive study of the dynamic interaction of applications, middleware, resource, and networks. We present experimental results with applications which validate the implementation of the MicroGrid, showing that it not only runs real grid applications and middleware, but also accurately models underlying resource and network behavior. Our scalability experiments show that our load balance algorithms are effective, and the best of them, hierarchical profile-driven load balance, scales well, enabling simulation networks of 20,000 routers with 90 cluster nodes. This is the largest detailed network simulation ever performed, and corresponds in size to a large ISP’s network. Realistic packet level network simulation with tens of thousands of routers enables accurate study of grid and network dynamics at unprecedented scale, and we believe great opportunities for new insights.