perf_tips.md

Tips for performance optimization

This file provides tips for troubleshooting slow or wasteful fuzzing jobs. See README.md for the general instruction manual.

1. Keep your test cases small

This is probably the single most important step to take! Large test cases do not merely take more time and memory to be parsed by the tested binary, but also make the fuzzing process dramatically less efficient in several other ways.

To illustrate, let's say that you're randomly flipping bits in a file, one bit at a time. Let's assume that if you flip bit #47, you will hit a security bug; flipping any other bit just results in an invalid document.

Now, if your starting test case is 100 bytes long, you will have a 71% chance of triggering the bug within the first 1,000 execs - not bad! But if the test case is 1 kB long, the probability that we will randomly hit the right pattern in the same timeframe goes down to 11%. And if it has 10 kB of non-essential cruft, the odds plunge to 1%.

On top of that, with larger inputs, the binary may be now running 5-10x times slower than before - so the overall drop in fuzzing efficiency may be easily as high as 500x or so.

In practice, this means that you shouldn't fuzz image parsers with your vacation photos. Generate a tiny 16x16 picture instead, and run it through jpegtran or pngcrunch for good measure. The same goes for most other types of documents.

There's plenty of small starting test cases in ../testcases/ - try them out or submit new ones!

If you want to start with a larger, third-party corpus, run afl-cmin with an aggressive timeout on that data set first.

2. Use a simpler target

Consider using a simpler target binary in your fuzzing work. For example, for image formats, bundled utilities such as djpeg, readpng, or gifhisto are considerably (10-20x) faster than the convert tool from ImageMagick - all while exercising roughly the same library-level image parsing code.

Even if you don't have a lightweight harness for a particular target, remember that you can always use another, related library to generate a corpus that will be then manually fed to a more resource-hungry program later on.

Also note that reading the fuzzing input via stdin is faster than reading from a file.

3. Use LLVM instrumentation

When fuzzing slow targets, you can gain 20-100% performance improvement by using the LLVM-based instrumentation mode described in the llvm_mode README. Note that this mode requires the use of clang and will not work with GCC.

The LLVM mode also offers a "persistent", in-process fuzzing mode that can work well for certain types of self-contained libraries, and for fast targets, can offer performance gains up to 5-10x; and a "deferred fork server" mode that can offer huge benefits for programs with high startup overhead. Both modes require you to edit the source code of the fuzzed program, but the changes often amount to just strategically placing a single line or two.

If there are important data comparisons performed (e.g. strcmp(ptr, MAGIC_HDR)) then using laf-intel (see llvm_mode/README.laf-intel.md) will help afl-fuzz a lot to get to the important parts in the code.

If you are only interested in specific parts of the code being fuzzed, you can instrument_files the files that are actually relevant. This improves the speed and accuracy of afl. See llvm_mode/README.instrument_file.md

Also use the InsTrim mode on larger binaries, this improves performance and coverage a lot.

4. Profile and optimize the binary

Check for any parameters or settings that obviously improve performance. For example, the djpeg utility that comes with IJG jpeg and libjpeg-turbo can be called with:

  -dct fast -nosmooth -onepass -dither none -scale 1/4

...and that will speed things up. There is a corresponding drop in the quality of decoded images, but it's probably not something you care about.

In some programs, it is possible to disable output altogether, or at least use an output format that is computationally inexpensive. For example, with image transcoding tools, converting to a BMP file will be a lot faster than to PNG.

With some laid-back parsers, enabling "strict" mode (i.e., bailing out after first error) may result in smaller files and improved run time without sacrificing coverage; for example, for sqlite, you may want to specify -bail.

If the program is still too slow, you can use strace -tt or an equivalent profiling tool to see if the targeted binary is doing anything silly. Sometimes, you can speed things up simply by specifying /dev/null as the config file, or disabling some compile-time features that aren't really needed for the job (try ./configure --help). One of the notoriously resource-consuming things would be calling other utilities via exec*(), popen(), system(), or equivalent calls; for example, tar can invoke external decompression tools when it decides that the input file is a compressed archive.

Some programs may also intentionally call sleep(), usleep(), or nanosleep(); vim is a good example of that. Other programs may attempt fsync() and so on. There are third-party libraries that make it easy to get rid of such code, e.g.:

https://launchpad.net/libeatmydata

In programs that are slow due to unavoidable initialization overhead, you may want to try the LLVM deferred forkserver mode (see llvm_mode/README.md), which can give you speed gains up to 10x, as mentioned above.

Last but not least, if you are using ASAN and the performance is unacceptable, consider turning it off for now, and manually examining the generated corpus with an ASAN-enabled binary later on.

5. Instrument just what you need

Instrument just the libraries you actually want to stress-test right now, one at a time. Let the program use system-wide, non-instrumented libraries for any functionality you don't actually want to fuzz. For example, in most cases, it doesn't make to instrument libgmp just because you're testing a crypto app that relies on it for bignum math.

Beware of programs that come with oddball third-party libraries bundled with their source code (Spidermonkey is a good example of this). Check ./configure options to use non-instrumented system-wide copies instead.

6. Parallelize your fuzzers

The fuzzer is designed to need ~1 core per job. This means that on a, say, 4-core system, you can easily run four parallel fuzzing jobs with relatively little performance hit. For tips on how to do that, see parallel_fuzzing.md.

The afl-gotcpu utility can help you understand if you still have idle CPU capacity on your system. (It won't tell you about memory bandwidth, cache misses, or similar factors, but they are less likely to be a concern.)

7. Keep memory use and timeouts in check

If you have increased the -m or -t limits more than truly necessary, consider dialing them back down.

For programs that are nominally very fast, but get sluggish for some inputs, you can also try setting -t values that are more punishing than what afl-fuzz dares to use on its own. On fast and idle machines, going down to -t 5 may be a viable plan.

The -m parameter is worth looking at, too. Some programs can end up spending a fair amount of time allocating and initializing megabytes of memory when presented with pathological inputs. Low -m values can make them give up sooner and not waste CPU time.

8. Check OS configuration

There are several OS-level factors that may affect fuzzing speed:

• If you have no risk of power loss then run your fuzzing on a tmpfs partition. This increases the performance noticably. Alternatively you can use AFL_TMPDIR to point to a tmpfs location to just write the input file to a tmpfs.
• High system load. Use idle machines where possible. Kill any non-essential CPU hogs (idle browser windows, media players, complex screensavers, etc).
• Network filesystems, either used for fuzzer input / output, or accessed by the fuzzed binary to read configuration files (pay special attention to the home directory - many programs search it for dot-files).
• On-demand CPU scaling. The Linux ondemand governor performs its analysis on a particular schedule and is known to underestimate the needs of short-lived processes spawned by afl-fuzz (or any other fuzzer). On Linux, this can be fixed with:

    cd /sys/devices/system/cpu
    echo performance | tee cpu*/cpufreq/scaling_governor

On other systems, the impact of CPU scaling will be different; when fuzzing,
use OS-specific tools to find out if all cores are running at full speed.

• Transparent huge pages. Some allocators, such as jemalloc, can incur a heavy fuzzing penalty when transparent huge pages (THP) are enabled in the kernel. You can disable this via:

    echo never > /sys/kernel/mm/transparent_hugepage/enabled

• Suboptimal scheduling strategies. The significance of this will vary from one target to another, but on Linux, you may want to make sure that the following options are set:

    echo 1 >/proc/sys/kernel/sched_child_runs_first
    echo 1 >/proc/sys/kernel/sched_autogroup_enabled

Setting a different scheduling policy for the fuzzer process - say
`SCHED_RR` - can usually speed things up, too, but needs to be done with
care.

• Use the afl-system-config script to set all proc/sys settings above in one go.
• Disable all the spectre, meltdown etc. security countermeasures in the kernel if your machine is properly separated:

ibpb=off ibrs=off kpti=off l1tf=off mds=off mitigations=off
no_stf_barrier noibpb noibrs nopcid nopti nospec_store_bypass_disable
nospectre_v1 nospectre_v2 pcid=off pti=off spec_store_bypass_disable=off
spectre_v2=off stf_barrier=off

In most Linux distributions you can put this into a `/etc/default/grub`
variable.

9. If all other options fail, use -d

For programs that are genuinely slow, in cases where you really can't escape using huge input files, or when you simply want to get quick and dirty results early on, you can always resort to the -d mode.

The mode causes afl-fuzz to skip all the deterministic fuzzing steps, which makes output a lot less neat and can ultimately make the testing a bit less in-depth, but it will give you an experience more familiar from other fuzzing tools.