Title: Footprint-based Locality Analysis Authors: Xiaoya Xiang, Bin Bao and Chen Ding (University of Rochester) Speaker: Xiaoya Xiang Abstract: Locality analysis traditionally focuses on data reuse patterns, in particular, the reuse distance of each access. At fine granularity such as cache blocks, the full analysis causes hundreds to thousands of times slowdown. Sampling, a common statistical tool, cannot reduce the cost by simply lowering the sampling rate, because it may still have to examine the whole trace to identify the next use of a sampled datum. An alternative locality model is footprint, the amount of data accessed in an execution window. Recent advances have made it possible to measure all-window footprints. In this talk, we show the relation between the new footprint statistics and the traditional locality statistics. By connecting the two, we reduce the cost of off-line modeling of cache sharing by over 200 times and enable on-line locality analysis through parallel sampling at a marginal cost. The results show empirically, with overwhelming evidence, a simple relation between all-window statistics and reuse-window statistics.