Exploring Causes of Performance Overhead During Dynamic Binary Translation

Abstract

Dynamic Binary Translators (DBT) have applications ranging from program portability, instrumentation, optimizations, and improving software security. To achieve these goals and maintain control over the application's execution, DBTs translate and run the original source/guest programs in a sand-boxed environment. DBT systems apply several optimization techniques like code caching, trace creation, etc. to reduce the translation overhead and enhance program performance at run-time. However, even with these optimizations, DBTs typically impose a significant performance overhead, especially for short-running applications. This performance penalty has restricted the more wide-spread adoption of DBT technology, in spite of its obvious need. The goal of this work is to determine the different factors that contribute to the performance penalty imposed by dynamic binary translators. In this thesis, we describe the experiments that we designed to achieve our goal and report our results and observations. We use a popular and sophisticated DBT, DynamoRio, for our test platform, and employ the industry-standard SPEC CPU2006 benchmarks to capture run-time statistics. Our experiments find that DynamoRio executes a large number of additional instructions when compared to the native application execution. We further measure that this increase in the number of executed instructions is caused by the DBT frequently exiting the code cache to perform various management tasks at run-time, including code translation, indirect branch resolution and trace formation. We also find that the performance loss experienced by the DBT is directly proportional to the number of code cache exits. We will discuss the details on the experiments, results, observations, and analysis in this work.