Profiling utilities

Profilers are tools that analyze the dynamic behavior of your program: how fast, how often and, sometimes, when the parts of its code are executed when the program is run. First, the profilers gather data at run time. Then, they interpret this data, for one of two purposes:

• to display the information to programmers so that they can imporove their code or

• to give the information to the compilation system.

There are four types of profiling:

• time profiling

• line profiling

• flow profiling

• block profiling

Complementary utilities are available for profiling C and C++ source programs:

• prof is a time profiler. It reports the amount of time and the percentage of time that was spent executing the parts of a program. See ``Using prof'' for a discussion of prof.

• lprof is a line-by-line frequency profiler. It reports whether and how many times each line of source code was executed. See ``Using lprof'' for a discussion of lprof.

• fprof is a flow and time profiler. It reports when each function was called and when the function exited. See ``Using fprof'' for a discussion of fprof.

• fur is a tool that manipulates object files. It can rearrange functions, rearrange the code within functions, insert code into functions. fur can be used to do block and branch profiling. It reports when each function was called and when the function exited. See ``Using fur to perform block profiling'' for a discussion of fur.

There is a good deal of overlap in functionality between time profiling and flow profiling.

fprof has all of the functionality of prof and more, but this extra functionality comes with a cost. This cost is manifested in 3 ways: fprof experiments require a good deal of disk space, often require root privilege to run and generate more overhead in the experiment. If none of these issues matters to you, use fprof. If all you want is the kind of information that prof provides, use prof.

There seems to be overlap between line profiling and block profiling as well. However, line profiling is used for display to programmers while block profiling is intended for consumption by automated tools. (Block profile logs can be printed, but they are difficult to read).

Line profiling is complementary to time and flow profiling in this sense. You can use prof or fprof to identify the most time-consuming (or, for other reasons, critical) parts of a program. You can then use lprof to obtain line-specific information only about those parts. In that way, you can avoid generating uninformative output while targeting the lines of code whose performance needs to be improved. You may then be able to rewrite those lines to execute more efficiently. See ``Improving program performance by programmer analysis'' for more information.

The tuning performed using flow and block profiles is also complementary. Flow profiling determines the best order to lay out functions in the program. Block profiling determines the best order to lay out blocks within each function. (Actually, block profiling can also choose an order for the functions but does not do as well as flow profiling.)

The automated tools are very powerful, but only if you already have efficient programs. If you are just beginning to look into why your program is running slowly, the automated tools will have little effect. The automated tools have an effect similar to optimization: they make code run more efficiently, but neither can overcome inefficient program design.