$devvkit learn --librarie perf-(linux)-guide

perf (Linux) Guide

[linux][profiling][cpu][tracing]

Performance & Profiling

Install

sudo apt install linux-tools-common linux-tools-$(uname -r)
# Or from kernel source: tools/perf/
# macOS: use Instruments.app or xctrace instead

perf is the built-in Linux profiler that accesses hardware performance counters (PMCs) via the perf_event_open syscall. It can profile CPU cycles, cache misses, branch mispredictions, page faults, and context switches with minimal overhead.

The most common workflow: `perf record` samples the call stack at 1kHz (adjust with `-F`), then `perf report` shows the hottest code paths as a flame-graph-friendly call tree. Use `perf top` for real-time CPU hotspot viewing.

perf trace is a strace alternative using tracepoints. perf stat aggregates hardware counters. For Java/Python/Node, use `--call-graph dwarf` for frame-pointer-unwound stacks. GUI: FlameGraph scripts by Brendan Gregg (`perf script | stackcollapse-perf.pl | flamegraph.pl`).

Stats

CPU stats— Aggregate event counts.

perf stat ./myapp
perf stat -e cycles,instructions,cache-misses,cache-references,branches,branch-misses sleep 5
perf stat -e context-switches,cpu-migrations,page-faults -p $(pgrep myapp)

perf stat with multiplexing— Many counters at once.

perf stat -d ./myapp          # Default detail (-d = -e instructions,cycles,cache-misses)
perf stat -dd ./myapp         # More detail
perf stat -ddd ./myapp        # Maximum detail
perf stat -M IPC ./myapp      # Instructions-per-cycle metric

Sampling

Sample CPU— Record call stacks.

perf record -F 99 -a -g -- sleep 10    # 99 Hz, all CPUs, call-graph
perf record -F 99 -p $(pgrep nginx) -g --sleep 30
perf record -F 999 -e cache-misses -a -g -- sleep 5

perf top— Real-time CPU view.

perf top                    # Live, like htop for CPU
perf top -p $(pgrep mysqld)  # Single process
perf top -e cache-misses     # Show cache miss hotspots
perf top -F 999 -g           # With call chains

Reporting

Report hottest— Interactive TUI report.

perf report                  # TUI: navigate with arrows, Enter to zoom
perf report --stdio          # text dump
perf report --sort=dso,symbol # by shared library
perf annotate --stdio        # annotated source + assembly

Tracing

perf trace— Syscall tracing.

perf trace -s ./myapp                # Syscall summary
perf trace -e openat,read,write -p 1234
perf trace --duration 100              # Show syscalls > 100ms

Analysis Tricks

Flame graph— Generate SVG flame graph.

perf script > out.perf
# Download FlameGraph:
# git clone https://github.com/brendangregg/FlameGraph
./stackcollapse-perf.pl < out.perf | ./flamegraph.pl > flame.svg
# Open flame.svg in browser — search for hotspots

Compare before/after— Regression detection.

perf stat -r 10 -e cycles,instructions ./myapp-v1 > v1.log
perf stat -r 10 -e cycles,instructions ./myapp-v2 > v2.log
paste v1.log v2.log | awk '{print $1, $2, ($2-$1)/$1*100 "% change"}'

Profiling with debug symbols— Get symbol names.

# Compile with -fno-omit-frame-pointer
# Install debug symbols:
sudo apt install myapp-dbg
# Or set perf map for JIT (Node, Python):
export PERF_RECORD_JIT=1
perf record -F 99 -g node --perf-basic-prof app.js