Perf

Last update: 05. March 2023 (Created: 19. February 2023)

---

Perf

• Great resource for perf in general: https://www.brendangregg.com/perf.html
• Unofficial linux perf events page: https://web.eece.maine.edu/~vweaver/projects/perf_events/
• Perf Timechart manpage: https://manpages.ubuntu.com/manpages/impish/man1/perf-timechart.1.html

Visualizers

• Flamegraph:
• https://github.com/brendangregg/FlameGraph
• https://www.brendangregg.com/flamegraphs.html
• Speedscope visualizer:
• https://www.speedscope.app
• https://github.com/jlfwong/speedscope
• Hotspot:
• https://github.com/KDAB/hotspot
• Firefox Profiler:
• https://profiler.firefox.com

Install (Debian): apt install linux-tools-<kernel-version> linux-perf-<kernel-version>

Compile applications with: -fno-omit-frame-pointer

# Recording Examples

# All processes with simple callstack for 30 seconds.
perf record -ag -- sleep 30

# 99 samples per second.
perf record -F 99 -ag -- sleep 30

# Specific process with simple callstack for 60 seconds.
Perf record -g -p <pid> -- sleep 60

# Specific process with dwarf-callstack
perf record --call-graph=dwarf -p <pid> -- sleep 60
perf record -F 99 --call-graph=dwarf -p <pid> -- sleep 60

# For firefox visualizer
perf record -a -g -F 1000 > perf.data
perf script -F +pid | gzip > perf.gz

# Analysis

# Output as Tree
perf report -n --stdio
perf report --call-graph --stdio -G

# Output as flamegraph
# http://www.brendangregg.com/FlameGraphs/cpuflamegraphs.html
# Flamegraphs need `flamegraph.pl` and `stackcollapse-perf.pl` from https://github.com/brendangregg/FlameGraph

perf script | ./stackcollapse-perf.pl > out.perf-folded
./flamegraph.pl out.perf-folded > perf-kernel.svg

# perf timechart can also be used to see the timing of processes better

# Records
perf timechart record

# Makes SVG
# WILL BE VERY BIG! TRY INKSCAPE OR EVEN BETTER ADOBE ILLUSTRATOR TO VIEW!
perf timechart

# Compiler Flags:

V1:
-g -fno-inline -fno-omit-frame-pointer -fno-optimize-sibling-calls -marm

V2:
-g -fno-inline -fno-omit-frame-pointer -marm -mapcs-frame

V3:
-g -fno-inline -fno-omit-frame-pointer -fno-optimize-sibling-calls -marm -mapcs-frame