A framework for testing a library using afl-fuzz

Monolith offers facilities for testing an OCaml library (for instance, a
data structure implementation) by comparing it against a reference
. It can be used to perform either random testing or fuzz
testing. Fuzz testing relies on the external tool

The user must describe what types and operations the library provides. Under
the best circumstances, this requires 2-3 lines of code per type or
. The user must also provide a reference implementation and a
candidate implementation of the library.

Then, like a monkey typing on a keyboard, Monolith attempts to exercise the
library in every possible way, in the hope of discovering a sequence of
operations that leads to an unexpected behavior (that is, a situation where
the library either raises an unexpected exception or returns an incorrect
result). If such a scenario is discovered, it is printed in the form of an
OCaml program, so as to help the user reproduce the problem.

Monolith assumes that the candidate implementation behaves in a deterministic
way. (Without this assumption, one cannot hope to reliably produce a
problematic scenario.) It does however allow nondeterministic specifications,
that is, situations where the candidate implementation is allowed to behave in
several possible ways.


To install the latest released version, type:

  opam update
  opam install monolith

To install the latest development version, type:

  git clone
  cd monolith
  opam pin add monolith .


The documentation of the latest released

is available online.

The documentation is built locally by make doc and can be viewed via make view.


The paper
Strong Automated Testing of OCaml Libraries
by François Pottier describes the use and the design of Monolith
in somewhat greater depth than the documentation.

Demos and Workflow

A number of demos are found under demos/. The demos under demos/working do
not have any known bugs, so make test should run forever without finding any
problem. The demos under demos/faulty intentionally contain bugs, so make test should very quickly find a number of problems.

Some demos require external libraries. Running make dependencies once at the
top level should install all of the libraries needed by the demos.

These demos share a common workflow, which is implemented in
This file is installed at the same time as the library. The
directory where it resides can be found via the command
ocamlfind query monolith.

First, the external tool afl-fuzz must be installed. Here is
a suggested installation script.

Then, an appropriate opam switch, such as 4.09.1+afl, must be created. The
name of this switch is controlled by the variable SWITCH. This variable can
be set by passing something like SWITCH=4.09.1+afl on the command line in
every make invocation, or (better) in a parent Makefile which includes

Creating an opam switch is done only once, as follows:

  make setup               # this takes a few minutes

If you are using MacOS, you will probably need to disable the system's
built-in crash reporter by typing make unload. (Otherwise, afl-fuzz will

If you are using an antivirus, it is advisable to disable it. As a case in
point, Kaspersky Endpoint Security on MacOS imposes a 5x speed penalty when
its protection is turned on.

After these steps, you are ready to run a demo. In each subdirectory of
demos, testing is carried out as follows:

  make test

This launches afl-fuzz, which keeps running forever, or until it is
interrupted by Ctrl-C. The number of problematic scenarios (also known as
crashes) found by afl-fuzz is displayed in red.

Once one or more crashes have been found and afl-fuzz has been interrupted,
it is recommended to first minimize the problem scenarios by using make min.

Then, you can ask for a report using either make show (which produces OCaml
code for each crash) or make summary (which shows a one-line summary of each
crash, by decreasing order of length).

A typical workflow is to first use make summary, so as to reveal the length
of the shortest crash, then type make show and search its output (by typing
/) for the shortest crash.

To perform purely random testing, without using afl-fuzz, type:

  make random

This uses a single processor core and runs in an infinite loop until a
problem is detected. Random testing is significantly faster than fuzz
testing, so it may be a good idea to first look for obvious bugs using
random testing, then look for nastier bugs using fuzz testing.

When running in random mode, after Monolith has a found a scenario, it reduces
the amount of fuel and continues searching for a shorter scenario. It is
therefore possible to start with a relatively large amount of fuel.

Parallel Fuzzing

The command make test launches only one afl-fuzz process. Once you are
confident that it works, you can instead use make multicore or make tmux,
which launch several processes in parallel. They differ in their user
interface: whereas make multicore displays only periodic progress reports in
the terminal, make tmux splits the terminal window using tmux and shows
the activity of each afl-fuzz process in its own pane.

These commands launch one master process plus a number of slave processes. The
number of slave processes can be adjusted via the command line: for instance,
make SLAVES=1 tmux uses just one master process and one slave process.

While parallel fuzzing is ongoing, typing make whatsup displays a progress
report. The most interesting line is Crashes found, which reports the number
of violations found so far. The Cumulative speed line is also of interest,
but be aware that it seems to take a few minutes for this information to
become stable.


Performance is measured in executions per second. It is visible in the
interactive interface, and is also reported by make whatsup.

Of course, the performance that you can expect depends on the cost of the
operations that you are testing. As a single data point, in a version of the
map demo, on a Linux machine equipped with two eight-core Intel Xeon CPUs
(E5-2620 v4 @ 2.10GHz), I am seeing 12k execs/second when using a single
afl-fuzz process, and about 5k execs/second/core when using all 32 cores
in parallel (therefore, about 160k execs/second).

Fuzzing under MacOS appears to be about 10 times slower, although I do not
know why; I have disabled both the MacOS crash reporter and the antivirus.

Real-World Applications

Monolith has been used to test
a library that offers 4 abstract types and well over 100 operations on them.

26 May 2021
>= "20200410"
>= "2.0"
>= "4.03"
Reverse Dependencies
>= "0.2.0"
>= "0.0.5"
>= "20201012"