Ephemeral Instrumentation for Lightweight Program Profiling

Program profiling is a mechanism that is useful for performance evaluation and code optimization. Profiling techniques that provide detailed information with extremely low overhead are especially important for systems that continuously monitor or dynamically optimize running executables. In this paper, we describe an approach for program profiling called ephemeral instrumentation and show that it collects useful profiles with low overhead. This approach builds on ideas from both program instrumentation and statistical sampling; it produces binaries that are able to periodically record aspects of their executions in great detail. It works because program behavior is predictable and because we are able to convert ephemeral profiles into traditional formats. This paper describes an ephemeral instrumentation system for gathering branch biases and post-processing that data into a traditional edge profile. We evaluate the usefulness of such profiles by using them to drive a superblock scheduler. Our experimental results show that we can gather ephemeral profiles with extremely low overheads (1-5%) while acquiring profile data that rivals the usefulness of complete profiles gathered at much higher overheads.