A Framework for Performance Analysis Tools

perform statistical sampling by tak-ing periodic snapshots of a program’s state. Statistical sampling is nonintrusive—unlike binary instrumentation, statistical sampling doesn’t add any lines of code to the appli-cation being tested—but the computing community generally regards this technique as imprecise. It imposes low overhead on a program’s runtime performance and can be used for time-critical experiments, but the measurements derived from statistical sam-pling can contain imprecisions. For exam-ple, without special hardware support, a sampled instruction pointer may not be related to the instruction address that caused a particular sampling event.1,2 Dynamic binary instrumentation, on the other hand, lets program instructions be changed on the fly, which can lead to more precise analysis. But because dynamic instrumentation modifies a pro-gram’s code—a program’s cache and pag-ing behavior, for example—execution time can slow significantly, even up to four times the normal execution speed. You typically perform dynamic instrumenta-tion at runtime, and you only test those parts of a program that will actually exe-cute. Doing so minimizes the overhead of the instrumentation process itself. Analysis tools based on dynamic instru-mentation—see the “Related work ” sidebar for information about other tools and sys-tems—require no special executable prepa-ration. HP Caliper, a framework for build-ing dynamic instrumentation tools, lets you change program instructions on the fly with instrumentation probes. Caliper integrates performance measurement unit (PMU) sampling with dynamic instrumentation. This article describes Caliper’s architecture, its public interfaces, and its dynamic instru-mentation algorithm. While Caliper will eventually run on sev-eral different platforms, this article specifi-cally describes Caliper’s implementation on HP-UX powered by the IA-64 Itanium processor. The IA-64’s instruction set archi-tecture (ISA) offers a set of architectural fea-tures designed to create synergy between compiler and processor.2 Caliper’s archi-tecture takes advantage of these features. HP Caliper offers a common framework for building performance analysis tools that can integrate hardware-supported sampling with dynamic instrumentation