Due to advances of modeling techniques, massive models are easily generated these days. Such massive models can consist of hundreds of millions of primitives and thus use more than tens of gigabytes. This high memory requirement is likely to cause serious performance issues on visualization and rendering because of the heavy loads of data accesses. Moreover, it is not trivial to make the data access pattern of global illumination be coherent which aggravates the performance degradation. In this thesis, three directions to address the issues are proposed. To reduce the expensive data transmission from external drives, we propose two kinds of compressed representations of massive models as in-core and out-of-core representations. The proposed representations show significant performance benefits of various applications requiring random access. We further accelerate the performance by fully utilizing heterogeneous computing resources, CPU and GPU. We propose a novel framework which drastically reduces data transmission overhead between the heterogeneous resources. By using the framework, an interactive performance of rendering massive models with global illumination effects is achieved. Finally, we propose an optimization method which maximizes both rendering throughput and responsiveness. The rendering framework with the optimization robustly works with running machines of different performances. A user study is performed to show the benefit of the optimization.