perfgrind

This is 'perfgrind', tools for collecting samples from Linux performance events subsystem and converting profiling data to callgrind format, allowing it to be read with KCachegrind.

Because of its own simplified format containing only the data necessary for creating the callgrind profile, the resulting file is commonly much smaller. One additional reason is that perfgrind explicitly ignores the kernel space during profiling.

Note: Perfgrind has a known limitation which is on the TODO list - it currently does not handle separate debug (neither on disk nor via debuginfod). Compiling with debug info and collecting data from non-stripped binaries will provide you with useful tracing data; especially when calling into system libraries you may see entries like func_7f2192e087070 in ld-2.31.so and similar.

License

This software is available to everyone under the license GPLv2.
It uses parts of code derived from:

• Linux kernel, license is GPLv2
• elfutils, license is GPLv3 or GPLv2 or LGPLv2

Usage

Overview

• collect samples using pgcollect into perfgrind format
• convert collected samples into callgrind format using pgconvert
• open resulting file in KCachegrind

pgcollect - collect samples

Usage: pgcollect filename.pgdata [-F freq] [-s] {-p pid | [--] cmd}

Options to specify output:

• filename.pgdata name of output file

Options to adjust profiling:

• -F freq profile at the given frequency freq
• -s profile using software events

Options to specify target:

• -p pid profile running process with PID=pid
• cmd command to profile, prefix with -- to stop command line parsing

pgconvert - convert collected samples to callgrind format

Usage: pgconvert [-m {flat|callgraph}] [-d {object|symbol|source}] [-i] filename.pgdata [filename.grind]
Note: If no output name is specified, then stdout will be used instead.
Examples:

• overview showing call stack
  pgconvert -d symbol filename.pgdata overview.grind
• full data with source annotation and instructions
  pgconvert -i filename.pgdata full.grind

Options to adjust generated callgrind data:

• -d specify detail level; default is "source"
• -i dump instructions, only possible with detail level "source"
• -m mode default mode is "callgraph" if detail level is not "object"

Note: To collect with hardware counters you may have to adjust the kernel parameter perf_event_paranoid as root.

pginfo - show event count and calculated entries

Usage: pginfo {flat|callgraph} filename.pgdata

• flat simple calculation, fast way to show number of events
• callgraph full calculation

Building

Dependency elfutils

either install from source or - preferably - via package manager, for example by issuing yum install elfutils-devel or apt install libdw-dev

Building the source

• optional step: create site.mak file and set FLAGS variable with paths to elfutils header and libraries (necessary if using a "local" version of elfutils)
  For example:
  FLAGS=-I/usr/local/elfutils/include -L/usr/local/elfutils/lib -O2 -march=native -Wl,-rpath /usr/local/elfutils/lib
• build it by issuing make
• optional: run tests with make check
• optional: install binaries to enable use by others with make install