Institute for Data Analysis and Visualization

In this paper we present a consistent, distributed, shared memory system for GPU texture memory. This model enables the virtualization of texture memory and the transparent, scalable sharing of texture data across multiple GPUs. Textures are stored as pages, and as textures are read or written, our system satisfies requests for pages on demand while maintaining memory consistency. Our system implements a directory-based distributed shared memory abstraction and is hidden from the programmer in order to ease programming in a multi-GPU environment. Our primary contributions are the identification of the core mechanisms that enable the abstraction and the future support that will enable them to be efficient.