Improving the Performance of 3D-Stacked DRAM Based Main Memory System

Modern data-intensive applications running concurrently on multiple processing cores require high memory bandwidth based on the size of data-set and locality found within the application. Traditional organization of DRAM modules cannot cope with this increasing memory bandwidth requirement. 3D- stacked DRAM modules provide unique advantages like huge bandwidth, memory-level parallelism, and logic area. In this thesis, we analyze the application’s run-time memory access behavior and propose novel memory management schemes that improve the system performance and energy efficiency by taking advantage of 3D-stacked DRAM architecture. First, we present CAMPS, a conflict-aware memory- side prefetching scheme proposed for Hybrid Memory Cube based main memory system. Secondly, we propose FAPS-3D, a feedback-directed adaptive page management scheme for 3D-stacked DRAM, that analyzes application’s high and low locality phases and recommends open- or close-page policy for the DRAM banks. Next, we present a memory-side prefetching scheme incorporating dynamic page mode in 3D-stacked DRAM, which categorizes the open- and close-page phases of the running application and suggests the prefetching policy that is optimized for the corresponding phase. Finally, we propose a dynamic page policy using perceptron learning, where perceptron learning is used to get deeper insight into the application’s long term memory access behavior and a perceptron is trained to predict the page open or close decision for the future accesses. Our evaluation shows that these proposed schemes improve the performance and energy efficiency of the 3D-stacked DRAM based main memory systems with a trivial area overhead.