A Survey of Phase Analysis: Techniques, Evaluation and Applications

Most programs exhibit significant repetition of behaviour, and detecting program phases is an increasingly important aspect of dynamic, adaptive program optimizations. Phase-based optimization can also reduce profiling overhead, save simulation time, and help in program understanding. Using a novel state space characterization, we give an an overview of state-of-the-art phase detection and prediction techniques. We use this systematic exposition to situate our own long-range phase analysis, and evaluate our approach using a novel phase prediction evaluation metric. Our results show that phases with relatively long periods can be predicted with very high accuracy. This new technique, as well as the unexplored regions in our high level characterization suggest that further improvements and variations on program phase analysis are both possible and potentially quite useful.