Analyzing Dynamic Binary Instrumentation Overhead

Robust and powerful software instrumentation tools are essential for dynamic program analysis tasks such as profiling, performance evaluation, and bug detection. Dynamic binary instrumentation (DBI) is a general purpose technique that eases the development of program analysis tools by facilitating automatic low-level instrumentation. DBI-based program analysis can introduce high overhead and it is crucial for tool writers to minimize the cost. Analyzing the performance of instrumentation tools is challenging because most systems use a just-in-time compiler (JIT) to dynamically generate code. In this paper, we describe our method for analyzing the performance of instrumentation tools. The instrumented code is itself instrumented with basic block counters. We implement the profiler in Pin and use it to analyze the behavior of simple and complex instrumentation tools. The analysis yields several unexpected results about the dynamic behavior of instrumented programs. By examining these results, we often find effective solutions to improve performance