Improving Data Locality Using Dynamic Page Migration Based on Memory Access Histograms

This Paper presents an approach which dynamically and transparently improves the data locality of memory references in Non-Uniform Memory Access (NUMA) characterized systems. The approach is based on run-time data redistribution via user-level page migration. It uses memory access histograms gathered by hardware monitors to make correct decisions related to the placement of shared data. First performance experiments on several applications show the potential for a significant gain in speedup. In addition, a graphical user interface has been developed showing the actual data movement thereby helping the user to understand the behavior of the application and to detect performance bottlenecks. This feature complements an already existing Data Layout Visualization tool for the observation of memory locality.