Flamegraphs

Flamegraphs

A flamegraph draws every sampled stack as a column of frames, merges identical prefixes, and sets each frame's width to its share of samples. The x axis is sample share, not time order; the y axis is stack depth; colors carry no meaning unless a palette or a diff assigns one. The pipeline is always the same three stages: a profiler emits stacks, a stackcollapse-* script folds them to one line per unique stack with a count, and flamegraph.pl renders the SVG. The scripts live in https://github.com/brendangregg/FlameGraph and run from the checkout without an install step.

Contract

Inputs

• Profiler output: perf script text, a DTrace ustack() aggregation, Go pprof raw text, async-profiler collapsed output, or heaptrack_print -F folded stacks.
• The FlameGraph checkout on PATH: git clone https://github.com/brendangregg/FlameGraph then export PATH="$PATH:$PWD/FlameGraph".
• For a diff: two folded files from the same program under the same workload.
• Perl for the scripts. Rust inferno is a drop-in reimplementation when Perl is unavailable.

Procedure

1. Collect stacks with the profiler. For perf: perf record -F 999 -g -o perf.data ./prog and perf script -i perf.data > out.perf (linux-perf covers permissions and frame pointers). Done when: the raw stack text exists.

2. Fold the stacks with the collapser for the source format:

   Source	Collapser
   perf script	stackcollapse-perf.pl out.perf > out.folded
   DTrace ustack() aggregation	stackcollapse.pl out.stacks > out.folded
   Go go tool pprof -raw -output=prof.txt ./prog	stackcollapse-go.pl prof.txt > out.folded
   SystemTap	stackcollapse-stap.pl
   Java jstack dumps	stackcollapse-jstack.pl
   Lightweight Java Profiler	stackcollapse-ljp.awk
   async-profiler	none needed: asprof -d 30 -f out.collapsed <pid> writes folded output directly (the .collapsed extension selects the format)
   heaptrack	none needed: heaptrack_print -f run.zst -F out.folded

   A folded line reads main;parse;tokenize 1234. FlameGraph ships no Callgrind converter; for Valgrind Callgrind data use callgrind_annotate or KCachegrind through valgrind. Done when: the folded file has one stack per line with a trailing count.

3. Render: flamegraph.pl out.folded > fg.svg. Useful options: --title "text", --subtitle "text", --width <px>, --height <px per frame>, --minwidth <n> to hide frames narrower than n pixels (or n% of samples), --colors <palette> (hot default, mem, io, wakeup, chain, java, js, perl, and flat red, green, blue, aqua, yellow, purple, orange), --countname bytes when the counts are not samples, --inverted for an icicle chart, --reverse to merge from the leaf, --hash for a color per function name that is stable across runs, --cp for a consistent palette saved in palette.map. A one-liner: perf record -F 999 -g ./prog && perf script | stackcollapse-perf.pl | flamegraph.pl > fg.svg. Done when: fg.svg opens in a browser (xdg-open fg.svg).

4. Read the graph. Hover shows the frame name, sample count, and percentage; click zooms; the search box highlights matches. The actionable hotspot is the widest frame with no or only narrow children above it, because that is where the CPU is when sampled. Trace downward from it to learn who calls it and why.

   Shape	Meaning	Move
   Wide plateau at the top	Leaf function is the hotspot	Optimize its body
   Wide base with many narrow towers	One caller fans out to many callees	Reduce per-call overhead, batch, or cache
   Very tall narrow tower	Deep recursion or call chain	Check depth; consider iteration
   Row of tiny slivers sharing width	Cost spread across many small functions	Algorithmic change or inlining, not micro-tuning
   One frame spanning most of the width at every level	Single bottleneck	Fix it before anything else

   Done when: the hotspot is named with its percentage and its caller chain.

5. Diff two profiles: collapse both, then difffolded.pl before.folded after.folded | flamegraph.pl > diff.svg. Widths come from the after profile; red frames grew, blue frames shrank, pale shades are small changes, and a frame in only one profile is solid. -n normalizes the two sample totals before differencing, -s strips hex addresses so JIT or ASLR addresses do not split identical stacks. difffolded.pl -n after.folded before.folded | flamegraph.pl --negate > diff_inv.svg flips the sign so improvements read as red. Done when: the largest regression and the largest improvement are named.

Alternatives, when the SVG is not the right surface: perf script | speedscope - (npm install -g speedscope) for an interactive left-heavy and sandwich view; perf script -F +pid > profile.perf dropped onto profiler.firefox.com; cargo install inferno for perf script | inferno-collapse-perf | inferno-flamegraph > fg.svg; cargo install flamegraph for cargo flamegraph --bin mybin, which drives perf itself. Details and per-profiler recipes are in references/tools.md.

Failure and recovery

Output

The SVG path, the folded file paths, the hotspot named with its percentage and caller chain, and, for a diff, the largest regression and improvement with their frames.