Profiling

Profiling is the process of determining how a program is using the resources that it is consuming.

Resource that can be Profiled

Profiling produces a clear view of the call graph -- the hierarchy of function/procedure/method calls that takes place during the execution of the program.

Resources consumption that can be analyzed during profiling include:

• Time (both clock time (total real time, user time, and the amount of time the kernel spent on behalf of the program)
• Memory
• Temporary storage
• Energy (this is a relatively new area of profiling)

Profiling Granularity and Techniques

Most profiling systems determine resource usage on a per-function basis. Data may be gathered through two different techniques:

1. Sampling - interrupting the program frequently (such as 10 000 times/second) and determining which function is currently executing (by comparing the program counter to the debuginfo of the program).
2. Instrumentation - adding code to the binary or using debug controls (such as breakpoints) to determine when and how often specific actions take place, such as entry/exit to/from a function/procedure/method.

Profiling Tools

There are many profiling tools available. Open source options include:

• gprof
• perf
• oprofile
• SystemTap

These tools provide different combinations of profiling capabilities, and may provide additional functions.

Example: Profiling with 'gprof'

The gprof tool provides basic profiling capability using a combination of sampling (for times) and instrumentation (for call graph and counts). To use it:

1. Build the software to be profiled using the -pg (profile generation) and -g (debug) options to the gcc compiler. This may require that you modify the makefile or other build instructions, but it can often be done using the CFLAGS or CCOPTS variables passed to configure or make -- for example, ./configure CFLAGS="-g -pg -O2" (Important: make sure that you include all of the flags that are usually used in the build. The easiest way to discover what these are is usually to perform a regular build and observe the values used).
2. Execute the program. Ensure that you give it a typical to long execution time; if it is an interactive program, run through most of the commonly-used features, and if it is non-interactive, invoke it with common options and give it a good amount of data to process.
3. Check that a file named gmon.out was produced when the program ran. If not, recheck the previous steps.
4. Run the gprof program to generate a report: gprof filenameOfExecutableBeingProfiled (Important: use the same executable as you used to generate the gmon.out file).

The output from gprof is a text report. It can be converted to a graphical representation, which is often more useful, using the gprof2dot script to convert it to the GraphViz "dot" format, then using the dot utility to output it in the desired graphics format.

Examples:

# produce a PDF
gprof nameOfExecutable | gprof2dot | dot -Tps | ps2pdf - gprof.pdf 

# produce a PNG file
gprof nameOfExecutable | gprof2dot | dot -Tps | convert - gprof.png

# produce an SVG file
gprof nameOfExecutable | gprof2dot | dot -Tsvg > gprof.svg

# display on the screen
gprof nameOfExecutable | gprof2dot | dot -Tx11

For more details and help on interpreting the report, see the gprof documentation.

Example: Profiling with perf

Perf does not use instrumentation, so no changes are required to the build process.

To record performance data, run perf with the record argument: perf record nameOfExecutable To display performance data, run perf with the report argument: perf report|less for non-interactive output, or perf report for interactive display

Function vs. Method vs. Procedure
In procedural languages, called code blocks are often called functions. When programming in an object-oriented language, called code blocks may be called methods. Older or more general documentation may refer to called code blocks as procedures, routines, or subroutines. The distinction between functions, methods, procedures, routines, and subroutines is effectively one of terminology and "packaging" only - at the machine code level, the distinction effectively disappears.

Optional Lab

The SPO600 Profiling Lab was used in previous semesters in the SPO600 course. It is not a required lab in the current version of the course.