Evaluation of Dynamic Binary Instrumentation Approaches: Dynamic Binary Translation vs. Dynamic Probe Injection

From web browsing to bank transactions, to data analysis and robot automation, just about any task necessitates or benefits from the use of software. Ensuring a piece of software to be effective requires profiling the program’s behavior to evaluate its performance, debugging the program to fix incorrect behaviors, and examining the program to detect security flaws. These tasks are made possible by instrumentation---the method of inserting code into a program to collect data about its behavior. Dynamic binary instrumentation (DBI) enables programmers to understand and reason about program behavior by inserting code into a binary during run time to collect relevant data, and is more flexible than static or source-code instrumentation, but incurs run-time overhead. This thesis attempts to extend the preexisting characterization of the tradeoffs between dynamic binary translation (DBT) and dynamic probe injection (DPI), two popular DBI approaches, using Pin and LiteInst as sample frameworks. It also describes extensions to the LiteInst framework that enable it to instrument function exits correctly. This evaluation involved using the two frameworks to instrument a set of SPEC CPU 2006 benchmarks for counting function entries alone, counting both function entries and exits, or dynamically generating a call graph. On these instrumentation tasks, Pin performed close to native binary time while LiteInst performed significantly slower. Exceptions to this observation, and analysis of the probe types used by LiteInst, suggest that LiteInst incurs significant overhead when executing a large number of probes