Profiling means measuring the performance characteristics of a running application, for the purpose of identifying compute-intensive areas—often called "hot spots"—that may be worth improving. In general, profiling can suffer from an "observer effect" in that the act of collecting performance data about an application while it executes may significantly degrade performance. When the goal of profiling is to locate relative performance hot spots in an application, this is not a problem. When the goal is to characterize actual performance characteristics of an application as it operates in the field, this observer effect is undesirable.

There are two main approaches to profiling an application: instrumentation and statistical sampling. Instrumentation involves adding instructions to an application to collect information (function call duration, number of invocations, etc.). As this method inherently alters the execution of a program, it is most often associated with degrading the performance of the application as a whole while it is being profiled.

Statistical sampling involves querying the state of unmodified executable at regular intervals. This results in a less comprehensive set of measurements as achievable through instrumentation, but usually does not degrade the performance of an application at all.