As computational science simulations produce ever increasing volumes of data, executing part or even the entire visualization pipeline in the supercomputer side becomes more a requirement than an option. Given the uniqueness of the high performance K computer architecture, the HIVE visualization framework was developed, focusing on meeting visualization and data analysis demands of scientists and engineers. In this paper, we present an analysis on the input/output (I/O) performance of post-hoc visualization. The contribution of this research work is characterized by an analysis of a set of empirical study cases considering huge simulation datasets using HIVE on the K computer. Results from the experimental effort, using a dataset produced by a real-world global climate simulation, provide a differentiated knowledge on the impact of dataset partitioning parameters in the I/O performance of large-scale visualization systems, and highlight challenges and opportunities for performance optimizations.