DATA REORGANIZATION IN MEMORY: A DATA CENTRIC PERSPECTIVE
Greg Lloyd, R. B. Pearce, Maya B. Gokhale · University of North Texas Digital Library (University of North Texas) · 2013
Emerging 3D memory packaging technology offers new opportunities for computing near memory.We are investigating the potential benefits obtained by incorporating variations on simple, familiar data movement hardware within the memory subsystem that can transform source data structure(s) from the physical organization stored in memory into alternative organizations better suited to the varying access patterns that may occur during application execution.In our model, memory and logic are combined in a single 3D package as exemplified by the Hybrid Memory Cube.The proposed data structure reorganization logic assembles bytes of data within the 3D package so that only bytes that are needed by the application running on the main CPU leave the memory/logic package.Using a simulator built for this purpose, we present quantitative results of memory bandwidth savings on representative data centric benchmarks from image processing, graph analysis, and scientific simulation application domains.We find that using data reorganization engines in these benchmarks can reduce the number of bytes moving between CPU and memory to as low as 20% of the conventional memory access bandwidth, while still performing the computation in the main CPU.