Overhead Analysis

To ascertain the overhead values $\Delta_i$ and $\Delta_e$, we must measure the instrumentation code that calculates inclusive and exclusive performance. However, the $\Delta_i$ and $\Delta_e$ values may be at the same scale as the measurement precision. Thus, we must construct a test that guarantees statistical accuracy in our estimation of $\Delta_i$ and $\Delta_e$.

Our approach is to conduct a two-level experiment where the inclusive measurement code is executed $k$ number of times (in a tight loop) and the performance data being profiled is measured and stored. This procedure is then repeated $j$ times and the minimum performance value across the $j$ experiments is retained. We use that value divided by $k$ to compute $\Delta_i$. While this does not entirely insure against anomalous timing artifacts, its statistical safety can be improved by increasing $j$. The two-level experiment is then repeated, this time with the exclusive measurement code included, resulting in an approximation of $\Delta_i + \Delta_e$. We can substract our $\Delta_i$ approximation to find $\Delta_e$. It is important to note that we need to do this overhead evaluation at the beginning of the application execution for the specific set of profile performance data being measured.