Subscribe Free

Join 2670+ others. No spamming.
I promise!

We are currently under high development. Follow us at github.

Looking for Python Tutorials?
Check these awesome tutorials



brendangregg / FlameGraph


stack trace visualizer


Flame Graphs visualize profiled code

Main Website:

Example (click to zoom): Example

Other sites:

Flame graphs can be created in three steps:

  1. Capture stacks
  2. Fold stacks

1. Capture stacks

Stack samples can be captured using Linux perf_events, FreeBSD pmcstat (hwpmc), DTrace, SystemTap, and many other profilers. See the stackcollapse-* converters.

Linux perf_events

Using Linux perf_events (aka "perf") to capture 60 seconds of 99 Hertz stack samples, both user- and kernel-level stacks, all processes:

# perf record -F 99 -a -g -- sleep 60
# perf script > out.perf

Now only capturing PID 181:

# perf record -F 99 -p 181 -g -- sleep 60
# perf script > out.perf


Using DTrace to capture 60 seconds of kernel stacks at 997 Hertz:

# dtrace -x stackframes=100 -n 'profile-997 /arg0/ { @[stack()] = count(); } tick-60s { exit(0); }' -o out.kern_stacks

Using DTrace to capture 60 seconds of user-level stacks for PID 12345 at 97 Hertz:

# dtrace -x ustackframes=100 -n 'profile-97 /pid == 12345 && arg1/ { @[ustack()] = count(); } tick-60s { exit(0); }' -o out.user_stacks

60 seconds of user-level stacks, including time spent in-kernel, for PID 12345 at 97 Hertz:

# dtrace -x ustackframes=100 -n 'profile-97 /pid == 12345/ { @[ustack()] = count(); } tick-60s { exit(0); }' -o out.user_stacks

Switch ustack() for jstack() if the application has a ustack helper to include translated frames (eg, node.js frames; see: The rate for user-level stack collection is deliberately slower than kernel, which is especially important when using jstack() as it performs additional work to translate frames.

2. Fold stacks

Use the stackcollapse programs to fold stack samples into single lines. The programs provided are:

  • for DTrace stacks
  • for Linux perf_events "perf script" output
  • for FreeBSD pmcstat -G stacks
  • for SystemTap stacks
  • for XCode Instruments
  • for Intel VTune profiles
  • stackcollapse-ljp.awk: for Lightweight Java Profiler
  • for Java jstack(1) output
  • for gdb(1) stacks

Usage example:

For perf_events:
$ ./ out.perf > out.folded

For DTrace:
$ ./ out.kern_stacks > out.kern_folded

The output looks like this:

unix`_sys_sysenter_post_swapgs 1401
unix`_sys_sysenter_post_swapgs;genunix`close 5
unix`_sys_sysenter_post_swapgs;genunix`close;genunix`closeandsetf 85
unix`_sys_sysenter_post_swapgs;genunix`close;genunix`closeandsetf;c2audit`audit_closef 26
unix`_sys_sysenter_post_swapgs;genunix`close;genunix`closeandsetf;c2audit`audit_setf 5
unix`_sys_sysenter_post_swapgs;genunix`close;genunix`closeandsetf;genunix`audit_getstate 6
unix`_sys_sysenter_post_swapgs;genunix`close;genunix`closeandsetf;genunix`audit_unfalloc 2
unix`_sys_sysenter_post_swapgs;genunix`close;genunix`closeandsetf;genunix`closef 48


Use to render a SVG.

$ ./ out.kern_folded > kernel.svg

An advantage of having the folded input file (and why this is separate to is that you can use grep for functions of interest. Eg:

$ grep cpuid out.kern_folded | ./ > cpuid.svg

Provided Example

An example output from DTrace is included, both the captured stacks and the resulting Flame Graph. You can generate it yourself using:

$ ./ example-stacks.txt | ./ > example.svg

This was from a particular performance investigation: the Flame Graph identified that CPU time was spent in the lofs module, and quantified that time.


See the USAGE message (--help) for options:

USAGE: ./ [options] infile > outfile.svg

    --titletext             # change title text
    --width                 # width of image (default 1200)
    --height                # height of each frame (default 16)
    --minwidth              # omit smaller functions (default 0.1 pixels)
    --fonttype              # font type (default "Verdana")
    --fontsize              # font size (default 12)
    --countname             # count type label (default "samples")
    --nametype              # name type label (default "Function:")
    --colors                # "hot", "mem", "io" palette (default "hot")
    --hash                  # colors are keyed by function name hash
    --cp                    # use consistent palette (
    ./ --titletext="Flame Graph: malloc()" trace.txt > graph.svg

As suggested in the example, flame graphs can process traces of any event, such as malloc()s, provided stack traces are gathered.

Consistent Palette

If you use the --cp option, it will use the $colors selection and randomly generate the palette like normal. Any future flamegraphs created using the --cp option will use the same palette map. Any new symbols from future flamegraphs will have their colors randomly generated using the $colors selection.

If you don't like the palette, just delete the file.

This allows your to change your colorscheme between flamegraphs to make the differences REALLY stand out.


Say we have 2 captures, one with a problem, and one when it was working (whatever "it" is):

cat working.folded | ./ --cp > working.svg
# this generates a, as per the normal random generated look.

cat broken.folded | ./ --cp --colors mem > broken.svg
# this svg will use the same for the same events, but a very
# different colorscheme for any new events.

Take a look at the demo directory for an example: