The unprecedented amounts of data generated from large scientific simulations impose a grand challenge in data analytics, and I/O simply becomes a major performance bottleneck. To address this challenge, we present an application-aware I/O optimization technique in support of interactive large-scale scientific visualization. We partition a scientific data into blocks, and carefully place data blocks in a memory hierarchy according to a characterization of data access patterns of user visualization operations. We conduct an empirical study to explore the parameter space to derive optimal solutions. We use real-world large-scale simulation datasets to demonstrate the effectiveness of our approach.