Setvectors for Memory Phase Classification

Phase classification has frequently been discussed the recent years as a method to guide scheduling, compiler optimizations and program simulations. In this paper, we propose a new classification method called setvectors. We show that the new method outperforms classification accuracy of state of the art methods by approximately 6 to 25 percent, while it has comparable computational complexity to the fastest state of the art methods. As a second contribution, we introduce a new method called PoEC (Percentage of Equal Clustering) to objectively compare phase classification techniques.