7 OProfile—system-wide profiler #

OProfile consists of a kernel driver and a daemon for collecting data. It uses the hardware performance counters provided on many processors. OProfile is capable of profiling all code including the kernel, kernel modules, kernel interrupt handlers, system shared libraries, and other applications.

Modern processors support profiling through the hardware by performance counters. Depending on the processor, there can be many counters and each of these can be programmed with an event to count. Each counter has a value which determines how often a sample is taken. The lower the value, the more often it is used.

During the post-processing step, all information is collected and instruction addresses are mapped to a function name.

To use OProfile, install the oprofile package. OProfile works on AMD64/Intel 64, IBM Z, and POWER processors.

It is useful to install the *-debuginfo package for the respective application you want to profile. If you want to profile the kernel, you need the debuginfo package as well.

OProfile contains several utilities to handle the profiling process and its profiled data. The following list is a short summary of programs used in this chapter:

With OProfile, you can profile both the kernel and applications. When profiling the kernel, tell OProfile where to find the vmlinuz* file. Use the --vmlinux option and point it to vmlinuz* (usually in /boot). If you need to profile kernel modules, OProfile does this by default. However, make sure you read https://oprofile.sourceforge.net/doc/kernel-profiling.html.

Applications usually do not need to profile the kernel, therefore you should use the --no-vmlinux option to reduce the amount of information.

7.4.2 Getting event configurations #

  

The general procedure for event configuration is as follows:

1. Use first the events CPU-CLK_UNHALTED and INST_RETIRED to find optimization opportunities.

2. Use specific events to find bottlenecks. To list them, use the command perf list.

If you need to profile certain events, first check the available events supported by your processor with the ophelp command (example output generated from Intel Core i5 CPU):

# ophelp
oprofile: available events for CPU type "Intel Architectural Perfmon"

See Intel 64 and IA-32 Architectures Software Developer's Manual
Volume 3B (Document 253669) Chapter 18 for architectural perfmon events
This is a limited set of fallback events because oprofile does not know your CPU
CPU_CLK_UNHALTED: (counter: all))
        Clock cycles when not halted (min count: 6000)
INST_RETIRED: (counter: all))
        number of instructions retired (min count: 6000)
LLC_MISSES: (counter: all))
        Last level cache demand requests from this core that missed the LLC (min count: 6000)
        Unit masks (default 0x41)
        ----------
        0x41: No unit mask
LLC_REFS: (counter: all))
        Last level cache demand requests from this core (min count: 6000)
        Unit masks (default 0x4f)
        ----------
        0x4f: No unit mask
BR_MISS_PRED_RETIRED: (counter: all))
        number of mispredicted branches retired (precise) (min count: 500)

Specify the performance counter events with the option --event. Multiple options are possible. This option needs an event name (from ophelp) and a sample rate, for example:

# operf --events CPU_CLK_UNHALTED:100000

Warning: Setting sampling rates with CPU_CLK_UNHALTED

Setting low sampling rates can seriously impair the system performance while high sample rates can disrupt the system to such a high degree that the data is useless. It is recommended to tune the performance metric for being monitored with and without OProfile and to experimentally determine the minimum sample rate that disrupts the performance the least.

Before generating a report, make sure the operf has stopped. Unless you have provided an output directory with --session-dir, operf has written its data to CUR_DIR/oprofile_data/samples/current, and the reporting tools opreport and opannotate will look there by default.

Calling opreport without any options gives a complete summary. With an executable as an argument, retrieve profile data only from this executable. If you analyze applications written in C++, use the --demangle smart option.

The opannotate generates output with annotations from source code. Run it with the following options:

# opannotate --source \
   --base-dirs=BASEDIR \
   --search-dirs=SEARCHDIR \
   --output-dir=annotated/ \
   /lib/libfoo.so

The option --base-dir contains a comma separated list of paths which is stripped from debug source files. These paths were searched prior to looking in --search-dirs. The --search-dirs option is also a comma separated list of directories to search for source files.

Note: Inaccuracies in annotated source

Because of compiler optimization, code can disappear and appear in a different place. Use the information in https://oprofile.sourceforge.net/doc/debug-info.html to fully understand its implications.

This chapter only provides a short overview. Refer to the following links for more information: