A fuzzy ecosystem for evaluating the effect of numerical error on computational tools.
Computational analysis has become an essential component of science. The software tools used bear the weight of the countless models, discoveries, and interventions based upon them. However, computers were never intended to be perfect, and one doesn't need to look very far to find examples of where conceptually simple operations may fail due to floating point arithmetic (e.g. ...
$ python3 -c "print(sum([0.001] * 1000))"
1.0000000000000007
). While small issues like the one above are unlikely to cause catestrophic failures on their own, they can cascade over the course of an execution and it isn't uncommon to find a plane that requires a power cycle every two months or an unstable series that consistently converges to the wrong solution.
Fuzzy allows you to study the stability or trustworthiness of tools and pipelines. You start by instrumenting libraries in your pipeline which do the bulk of numerical heavy lifting, then run your tool multiple times, and finally analyze the variability in your results. This project aims to provide an environment in which the stability of programs can be evaluated, reducing the overhead on both tool developers and consumers.
You can get started with Fuzzy quite simply, just launch a Docker container as follows:
docker run -ti verificarlo/fuzzy
If you're on a shared system and would prefer to use Singularity, that's no problem, just convert the container using the appropriate method for your system (e.g.).
If you would like to build the environment locally on your system, look at the
Dockerfiles in docker/base/
to see how installation was performed. At the end of the
build chain, you'll find instrumented versions of libmath
, lapack
, python3
,
numpy
, and several other recompiled libraries.
An example for how to verify your installation for Python could be the following:
$ python3 -c "print([sum([.001]*1000) for _ in range(3)])"
[1.0, 0.9999999999999997, 1.0000000000000007, 1.0000000000000002]
If your code is written in, say, Python, then you can simply mount your code to
this environment, run it using the contained interpreter, and experince the :fuzz:.
If your code loads shared libraries, such as libmath
or lapack
, make sure that
they are using the instrumented versions of these libraries made available in
/usr/local/lib/
. You can check whether this is the case using ldd </path/to/yourbinary>
.
The next step is to make sure the environment is configured to introduce perturbations the way you expect. You can start with a configuration which performs perturbations akin to randomizing machine error with the following:
echo "libinterflop_mca.so -m mca --precision-binary32=24 --precision-binary64=53" > $VFC_BACKENDS_FROM_FILE
For more usage instructions on how to control how perturbations are introduced and where they occur within operations, please refer to the Verificarlo repository.
A simple sanity check to verify that your code is using the perturbed libraries
is to dramatically reduce the precision in your configuration by changing the
precision-binary32
and precision-binary64
values to something small (e.g. 1),
boot up a debugging session in your environment (e.g. a Python shell, GDB, etc.),
load a math library, and run a basic math function a few times (e.g. sin(1)
) —
you should get different results if your instrumentation is setup properly.
Important: Don't forget to set the precision back to your desired level
prior to performing your experiments! Fuzzy should print a log message to the
terminal when you run your commands, including the configuration, so you can verify
that the parameters were properly specified.
Since version v0.8.0, fuzzy-libmath comes with three modes (standard, quad, and
mpfr). Internally, the fuzzy-libmath computes the actual result of the math
function using the standard library of the OS and then perturbs the output by
adding MCA noise. Since the accuracy of the libm
used can vary from one
version to another (see this
article), one
can use higher precision to ensure accurate intermediate computations.
A new mode called fast
was introduced in version v0.9.1 that is equivalent
to a MCA noise with a fixed virtual precision (t=24 for binary32, t=53 for binary64)
that cannot be changed.
These modes specify the library used:
fast
: uses the standard libm provided by the OS with a virtual precision fixed to the last bit (fastest)standard
: uses the standard libm provided by the OSquad
: uses the libquadmath.so librarympfr
: uses the MPFR library with 113 bits of precision, equivalent to the binary128 precision (slowest)
Note that a higher precision implies a larger slowdown. The standard
mode is
the fastest and can be used if one does not require high accuracy. To switch
from one version to another, please use the set-fuzzy-libmath
tool already
installed in the docker image as follow:
usage: set-fuzzy-libmath [-h] --version {no-fuzzy,standard,quad,mpfr}
optional arguments:
-h, --help show this help message and exit
--version {no-fuzzy,standard,quad,mpfr}
Fuzzy libmath version to use:
- no-fuzzy: Disable fuzzy libmath
- fast: Use the standard libmath available on the system.
Not possible to modify the virtual precision!
Fastest instrumentation but possibly not accurate
- standard: Use the standard libmath available on the system.
Fast instrumentation but possibly not accurate
- quad: Use the libquadmath for intermediate results.
Slower than the standard version but more accurate.
- mpfr: Use the mpfr library for intermediate results.
Slowest version but gives the correct rounding.
Fuzzy provides a set of recompiled shared objects and tools that facilitate adding Monte Carlo Arithmetic to tools. If you've got a Docker container which relies on some of these libraries, you can easily add Fuzzy with a Multi-stage Docker build.
For example:
FROM verificarlo/fuzzy:latest as fuzzy
# Your target image
FROM user/image:version
# Copy libmath fuzzy environment from fuzzy image, for example
RUN mkdir -p /opt/mca-libmath/{fast,standard,quad,mpfr}
COPY --from=${2} /opt/mca-libmath/set-fuzzy-libmath.py /usr/local/bin/set-fuzzy-libmath
COPY --from=${2} /opt/mca-libmath/fast/libmath.so /opt/mca-libmath/fast/libmath.so
COPY --from=${2} /opt/mca-libmath/standard/libmath.so /opt/mca-libmath/standard/libmath.so
COPY --from=${2} /opt/mca-libmath/quad/libmath.so /opt/mca-libmath/quad/libmath.so
COPY --from=${2} /opt/mca-libmath/mpfr/libmath.so /opt/mca-libmath/mpfr/libmath.so
COPY --from=${2} /usr/local/lib/libinterflop* /usr/local/lib/
# If you will also want to recompile more libraries with verificarlo, add these lines
COPY --from=fuzzy /usr/local/bin/verificarlo* /usr/local/bin/
COPY --from=fuzzy /usr/local/include/* /usr/local/include/
# Preloading the instrumented shared library
ARG FUZZY_LIBMATH_VERSION=standard
RUN set-fuzzy-libmath --version=${FUZZY_LIBMATH_VERSION}
ENV VFC_BACKENDS 'libinterflop_mca.so --precision-binary32=24 --precision-binary64=53 --mode=mca'
In the context of Fuzzy experiments, it is important to remember that by default each execution will be evaluated with a unique random state, meaning that when you run it again you may get slightly (or very) different results. If your goal is to characterize the variability in your results, or obtain a robust estimate of the "true" mean answer, you will need to then run your tool multiple times and compare each execution.
It's also important to remember that the execution time will be increased when using Fuzzy, as compared to running tools in a determinitic environment. Depending on the tool, the instrumentation, and the MCA mode, this additional overhead may range from negligible to the order of a 30x slowdown.
Included in the references below are some references which can be referred to when deciding how many perturbations to run, how to consider groups of results, and demonstrating the differences in overhead that may exist between different instrumentations.
A detailed explanation of Monte Carlo Arithmetic can be found in the references below or the Verificarlo repository. In short, here is some terminology to get you started.
In this form of stochastic arithmetic you have three modes for perturbing your
floating-point operations x = a op b
where op
is in {+,-,*,/}
:
- Random Rounding (RR):
x = inexact(a op b)
- Precision Bounding (PB):
x = inexact(a) op inexact(b)
- Full MCA (MCA):
x = inexact(inexact(a) op inexact(b))
In this implementation, the inexact
operation is the addition of a 0
-centered
uniform random variable at the target bit of precision.
As you become familiar with Fuzzy, you may run into some of the following common sources of error in your (or other) software:
-
Fuzzy not running on your system, and giving
Illegal instruction
errors (likely exit-code132
)? This is probably because our environments were built in a different architecture than yours! We tried to turn off as many optimizations as we could, increasing the portability of these images, but sometimes libraries don't listen to instructions very well... Try rebuilding the images on your local machine. to fix the problem. To our knowledge,Scipy
is the first package in the install chain which has some dependencies that may ignore the optimization-disabling compiler flags, so you could try starting your rebuild from theNumpy
image. -
When using MCA or PB modes of perturbation, operations which rely on integer values which happen to be stored in floating point containers may crash. For instance, imagine you're creating an array and store the desired length as
3.0
instead of3
; introduced perturbations may shift your length away from the exact-value of3.0
, and then your tool could (justifiably) crash when trying to allocate a2.99..
element array. You can fix these types of bugs by simply casting your length variable to an integer. -
In cases where piecewise approximations are used to solve complex functions (e.g. the sin function in libc.), it is possible that perturbations will trigger distinct branching and result in possibly large discontinuity between results. Fuzzy appraches instrumenting these libraries via wrappers which ultimately perturb the function inputs and outputs, rather than the internal arithmetic operations which may take place. An example of this can be found for
libmath
.
For instructions on how to contribute, please refer to the Contribution Guide.
If you prefer a visual illustration of everything we've written above, feel free to check-out the slides we recently presented at Scipy2021!
The Fuzzy ecosystem has emerged from — and been used in — several scientific publications. Below is a list of papers which present the techniques used, the tools which have been developed accordingly, and demonstrate how decision-making and applications can be built atop them:
Parker, Douglas Stott, Brad Pierce, and Paul R. Eggert. "Monte Carlo arithmetic: how to gamble with floating point and win." Computing in Science & Engineering 2.4 (2000): 58-68.
Frechtling, Michael, and Philip HW Leong. "Mcalib: Measuring sensitivity to rounding error with monte carlo programming." ACM Transactions on Programming Languages and Systems (TOPLAS) 37.2 (2015): 1-25.
C. Denis, P. De Oliveira Castro and E. Petit, "Verificarlo: Checking Floating Point Accuracy through Monte Carlo Arithmetic," in 2016 IEEE 23nd Symposium on Computer Arithmetic (ARITH), Silicon Valley, CA, USA, 2016 pp. 55-62.
Chatelain, Yohan, et al. "VeriTracer: Context-enriched tracer for floating-point arithmetic analysis." 2018 IEEE 25th Symposium on Computer Arithmetic (ARITH). IEEE, 2018.
Kiar, Gregory, et al. "Comparing perturbation models for evaluating stability of neuroimaging pipelines." The International Journal of High Performance Computing Applications 34.5 (2020): 491-501.
Sohier, Devan, et al. "Confidence Intervals for Stochastic Arithmetic." ACM Transactions on Mathematical Software (TOMS) 47.2 (2021): 1-33.
Kiar, Gregory, et al. "Data Augmentation Through Monte Carlo Arithmetic Leads to More Generalizable Classification in Connectomics." bioRxiv (2020).
Chatelain, Yohan, et al. "PyTracer: Automatically profiling numerical instabilities in Python." IEEE Transactions on Computers (2022).
Des Ligneris, Morgane, et al. "Reproducibility of tumor segmentation outcomes with a deep learning model." 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI). IEEE, 2023.
Pepe, Inés Gonzalez, et al. "Numerical Uncertainty of Convolutional Neural Networks Inference for Structural Brain MRI Analysis." International Workshop on Uncertainty for Safe Utilization of Machine Learning in Medical Imaging. Cham: Springer Nature Switzerland, 2023.
Chatelain, Yohan, et al. "A numerical variability approach to results stability tests and its application to neuroimaging." arXiv preprint arXiv:2307.01373 (2023).
Vila, Gaël, et al. "The Impact of Hardware Variability on Applications Packaged with Docker and Guix: a Case Study in Neuroimaging." (2024).
Gonzalez Pepe, Inés, et al. "Numerical stability of DeepGOPlus inference." Plos one 19.1 (2024): e0296725.
The Fuzzy copyright belongs to all contributors of this repository, and it is licensed for public use under the same terms as the LLVM project, which is a modified version of the Apache 2.0 license.