CIM: A Reliable Metric for Evaluating Program Phase Classifications

We propose the use of the Confidence Interval of estimated Mean (CIM), a metric based on statistical sampling theory, to evaluate the quality of a given phase classification and for comparing different phase classification schemes. Previous research on phase classification used the Weighted Average of Coefficient of Variation (CoVwa) to estimate phase classification quality. We found that the phase quality indicated by CoVwa could be inconsistent across different phase classifications. We explain the reasons behind this inconsistency and demonstrate the inconsistency using data from several SPEC CPU2000 benchmark programs. We show that the Confidence Interval of estimated Mean (CIM) correctly estimates the quality of phase classification with a meaningful statistical interpretation.