Performance tuning with instruction-level cost derived from call-stack sampling

Abstract

Except for program-counter histogramming, most modern profiling tools summarize at the level of entire functions or basic blocks, with or without additional information such as calling context or call graphs. This paper explicates the value of information about the cost of specific instructions, relative to summaries that do not include it. A good source of this information is time-random sampling of the call stack. To get the diagnostic benefit of instruction costs it is not necessary to measure them with high precision or efficiency. In fact, manual sampling suffices quite well, when it can be used. Other benefits of call stack sampling are that it can be used with unmodified software and libraries, and it is easily confined to the time intervals of interest. As with other profiling techniques, it can be employed repeatedly to remove all significant performance problems in single-thread programs.