Distance-Based Locality Analysis and Prediction

Profiling can accurately analyze program behavior for select data inputs. This article shows that profiling can also predict program locality for inputs other than profiled ones. Here locality is defined by the distance of data reuse. The article describes three distance-based techniques for whole-program locality analysis. The first is approximate measurement of reuse distance in near linear time. It can measure the reuse distance of all accesses to all data elements in fullsize benchmarks with guaranteed precision. The second is pattern recognition. Based on a few training runs, it classifies patterns as regular and irregular and, for regular ones, it predicts their (changing) behavior for other inputs. It uses regression and multi-model analysis to reduce the prediction error, the space overhead, and the size of the training runs. The last technique is online prediction, which uses distance-based sampling at the beginning of an execution to estimate the locality of the whole execution. When tested on 15 integer and floating-point programs from SPEC and other benchmark suites, these techniques predict with on average 94% accuracy for data inputs up to hundreds times larger than the training inputs. Distance-based locality analysis has been used in measuring and improving the cache performance of a wide range of programs.