Chapter 1. Introduction

When developing a program, one of the last steps often involves performance optimizations. As it makes no sense to optimize functions rarely used, because that would be a waste of time, one needs to know in which part of a program most of the time is spent.

For sequential code, collecting statistical data of the programs runtime characteristic like time numbers spent in functions and code lines usually is enough. This is called Profiling. The program is run under control of a profiling tool, which gives the summary of an execution run at the end. In contrast, for parallel code, performance problems typically are caused when one processor is waiting for data from another. As this waiting time usually cannot easily attributed, here it is better to generate timestamped event traces. KCachegrind cannot visualize this kind of data.

After analyzing the produced profile data, it should be easy to see the hot spots and bottlenecks of the code: for example, assumptions about call counts can be checked, and identified code regions can be optimized. Afterwards, the success of the optimization should be verified with another profile run.