Tips for parallel fuzzing
This document talks about synchronizing afl-fuzz jobs on a single machine or across a fleet of systems. See README.md for the general instruction manual.
Every copy of afl-fuzz will take up one CPU core. This means that on an n-core system, you can almost always run around n concurrent fuzzing jobs with virtually no performance hit (you can use the afl-gotcpu tool to make sure).
In fact, if you rely on just a single job on a multi-core system, you will be underutilizing the hardware. So, parallelization is always the right way to go.
When targeting multiple unrelated binaries or using the tool in “non-instrumented” (-n) mode, it is perfectly fine to just start up several fully separate instances of afl-fuzz. The picture gets more complicated when you want to have multiple fuzzers hammering a common target: if a hard-to-hit but interesting test case is synthesized by one fuzzer, the remaining instances will not be able to use that input to guide their work.
To help with this problem, afl-fuzz offers a simple way to synchronize test cases on the fly.
Note that afl++ has AFLfast’s power schedules implemented. It is therefore a good idea to use different power schedules if you run several instances in parallel. See power_schedules.md
Alternatively running other AFL spinoffs in parallel can be of value, e.g. Angora (https://github.com/AngoraFuzzer/Angora/)
2) Single-system parallelization
If you wish to parallelize a single job across multiple cores on a local system, simply create a new, empty output directory (“sync dir”) that will be shared by all the instances of afl-fuzz; and then come up with a naming scheme for every instance - say, “fuzzer01”, “fuzzer02”, etc.
Run the first one (“main node”, -M) like this:
./afl-fuzz -i testcase_dir -o sync_dir -M fuzzer01 [...other stuff...]
…and then, start up secondary (-S) instances like this:
./afl-fuzz -i testcase_dir -o sync_dir -S fuzzer02 [...other stuff...] ./afl-fuzz -i testcase_dir -o sync_dir -S fuzzer03 [...other stuff...]
Each fuzzer will keep its state in a separate subdirectory, like so:
Each instance will also periodically rescan the top-level sync directory for any test cases found by other fuzzers - and will incorporate them into its own fuzzing when they are deemed interesting enough. For performance reasons only -M main node syncs the queue with everyone, the -S secondary nodes will only sync from the main node.
The difference between the -M and -S modes is that the main instance will still perform deterministic checks; while the secondary instances will proceed straight to random tweaks.
Note that you must always have one -M main instance! Running multiple -M instances is wasteful!
You can also monitor the progress of your jobs from the command line with the provided afl-whatsup tool. When the instances are no longer finding new paths, it’s probably time to stop.
WARNING: Exercise caution when explicitly specifying the -f option. Each fuzzer must use a separate temporary file; otherwise, things will go south. One safe example may be:
./afl-fuzz [...] -S fuzzer10 -f file10.txt ./fuzzed/binary @@ ./afl-fuzz [...] -S fuzzer11 -f file11.txt ./fuzzed/binary @@ ./afl-fuzz [...] -S fuzzer12 -f file12.txt ./fuzzed/binary @@
This is not a concern if you use @@ without -f and let afl-fuzz come up with the file name.
3) Multiple -M mains
There is support for parallelizing the deterministic checks. This is only needed where
- many new paths are found fast over a long time and it looks unlikely that main node will ever catch up, and
- deterministic fuzzing is actively helping path discovery (you can see this
in the main node for the first for lines in the “fuzzing strategy yields”
section. If the ration
found/attempsis high, then it is effective. It most commonly isn’t.)
Only if both are true it is beneficial to have more than one main. You can leverage this by creating -M instances like so:
./afl-fuzz -i testcase_dir -o sync_dir -M mainA:1/3 [...] ./afl-fuzz -i testcase_dir -o sync_dir -M mainB:2/3 [...] ./afl-fuzz -i testcase_dir -o sync_dir -M mainC:3/3 [...]
… where the first value after ‘:’ is the sequential ID of a particular main instance (starting at 1), and the second value is the total number of fuzzers to distribute the deterministic fuzzing across. Note that if you boot up fewer fuzzers than indicated by the second number passed to -M, you may end up with poor coverage.
4) Syncing with non-afl fuzzers or independant instances
A -M main node can be told with the
-F other_fuzzer_queue_directory option
to sync results from other fuzzers, e.g. libfuzzer or honggfuzz.
Only the specified directory will by synced into afl, not subdirectories. The specified directory does not need to exist yet at the start of afl.
-F option can be passed to the main node several times.
5) Multi-system parallelization
The basic operating principle for multi-system parallelization is similar to the mechanism explained in section 2. The key difference is that you need to write a simple script that performs two actions:
Uses SSH with authorized_keys to connect to every machine and retrieve a tar archive of the /path/to/sync_dir/<main_node(s)> directory local to the machine. It is best to use a naming scheme that includes host name and it’s being a main node (e.g. main1, main2) in the fuzzer ID, so that you can do something like:
for host in `cat HOSTLIST`; do ssh user@$host "tar -czf - sync/$host_main*/" > $host.tgz done
Distributes and unpacks these files on all the remaining machines, e.g.:
for srchost in `cat HOSTLIST`; do for dsthost in `cat HOSTLIST`; do test "$srchost" = "$dsthost" && continue ssh user@$srchost 'tar -kxzf -' < $dsthost.tgz done done
There is an example of such a script in utils/distributed_fuzzing/.
There are other (older) more featured, experimental tools:
However these do not support syncing just main nodes (yet).
When developing custom test case sync code, there are several optimizations to keep in mind:
The synchronization does not have to happen very often; running the task every 60 minutes or even less often at later fuzzing stages is fine
There is no need to synchronize crashes/ or hangs/; you only need to copy over queue/* (and ideally, also fuzzer_stats).
It is not necessary (and not advisable!) to overwrite existing files; the -k option in tar is a good way to avoid that.
There is no need to fetch directories for fuzzers that are not running locally on a particular machine, and were simply copied over onto that system during earlier runs.
For large fleets, you will want to consolidate tarballs for each host, as this will let you use n SSH connections for sync, rather than n*(n-1).
You may also want to implement staged synchronization. For example, you could have 10 groups of systems, with group 1 pushing test cases only to group 2; group 2 pushing them only to group 3; and so on, with group eventually 10 feeding back to group 1.
This arrangement would allow test interesting cases to propagate across the fleet without having to copy every fuzzer queue to every single host.
You do not want a “main” instance of afl-fuzz on every system; you should run them all with -S, and just designate a single process somewhere within the fleet to run with -M.
Syncing is only necessary for the main nodes on a system. It is possible to run main-less with only secondaries. However then you need to find out which secondary took over the temporary role to be the main node. Look for the
is_main_nodefile in the fuzzer directories, eg.
It is not advisable to skip the synchronization script and run the fuzzers directly on a network filesystem; unexpected latency and unkillable processes in I/O wait state can mess things up.
6) Remote monitoring and data collection
You can use screen, nohup, tmux, or something equivalent to run remote instances of afl-fuzz. If you redirect the program’s output to a file, it will automatically switch from a fancy UI to more limited status reports. There is also basic machine-readable information which is always written to the fuzzer_stats file in the output directory. Locally, that information can be interpreted with afl-whatsup.
In principle, you can use the status screen of the main (-M) instance to monitor the overall fuzzing progress and decide when to stop. In this mode, the most important signal is just that no new paths are being found for a longer while. If you do not have a main instance, just pick any single secondary instance to watch and go by that.
You can also rely on that instance’s output directory to collect the synthesized corpus that covers all the noteworthy paths discovered anywhere within the fleet. Secondary (-S) instances do not require any special monitoring, other than just making sure that they are up.
Keep in mind that crashing inputs are not automatically propagated to the main instance, so you may still want to monitor for crashes fleet-wide from within your synchronization or health checking scripts (see afl-whatsup).
7) Asymmetric setups
It is perhaps worth noting that all of the following is permitted:
Running afl-fuzz with conjunction with other guided tools that can extend coverage (e.g., via concolic execution). Third-party tools simply need to follow the protocol described above for pulling new test cases from out_dir/<fuzzer_id>/queue/* and writing their own finds to sequentially numbered id:nnnnnn files in out_dir/<ext_tool_id>/queue/*.
Running some of the synchronized fuzzers with different (but related) target binaries. For example, simultaneously stress-testing several different JPEG parsers (say, IJG jpeg and libjpeg-turbo) while sharing the discovered test cases can have synergistic effects and improve the overall coverage.
(In this case, running one -M instance per target is necessary.)
Having some of the fuzzers invoke the binary in different ways. For example, ‘djpeg’ supports several DCT modes, configurable with a command-line flag, while ‘dwebp’ supports incremental and one-shot decoding. In some scenarios, going after multiple distinct modes and then pooling test cases will improve coverage.
Much less convincingly, running the synchronized fuzzers with different starting test cases (e.g., progressive and standard JPEG) or dictionaries. The synchronization mechanism ensures that the test sets will get fairly homogeneous over time, but it introduces some initial variability.