Debugging the code
Sometimes, the code does not give you the result that you are expecting. This can be due to a variety of reasons, from misunderstandings or changes in the input parameters, system specific quirks, or bugs. You might also want to debug your code as you implement new features in WarpX during development.
This section gives a step-by-step guidance on how to systematically check what might be going wrong.
Debugging Workflow
Try the following steps to debug a simulation:
Check the output text file, usually called
output.txt
: are there warnings or errors present?On an HPC system, look for the job output and error files, usually called
WarpX.e...
andWarpX.o...
. Read long messages from the top and follow potential guidance.If your simulation already created output data files: Check if they look reasonable before the problem occurred; are the initial conditions of the simulation as you expected? Do you spot numerical artifacts or instabilities that could point to missing resolution or unexpected/incompatible numerical parameters?
Did the job output files indicate a crash? Check the
Backtrace.<mpirank>
files for the location of the code that triggered the crash. Backtraces are read from bottom (high-level) to top (most specific line that crashed).Was this a segmentation fault in C++, but the run was controlled from Python (PICMI)? To get the last called Python line for the backtrace, run again and add the Python
faulthandler
, e.g., withpython3 -X faulthandler PICMI_your_script_here.py
.
Try to make the reproducible scenario as small as possible by modifying the inputs file. Reduce number of cells, particles and MPI processes to something as small and as quick to execute as possible. The next steps in debugging will increase runtime, so you will benefit from a fast reproducer.
Consider adding runtime debug options that can narrow down typical causes in numerical implementations.
In case of a crash, Backtraces can be more detailed if you re-compile with debug flags: for example, try compiling with
-DCMAKE_BUILD_TYPE=RelWithDebInfo
(some slowdown) or even-DCMAKE_BUILD_TYPE=Debug
(this will make the simulation way slower) and rerun.If debug builds are too costly, try instead compiling with
-DAMReX_ASSERTIONS=ON
to activate more checks and rerun.If the problem looks like a memory violation, this could be from an invalid field or particle index access. Try compiling with
-DAMReX_BOUND_CHECK=ON
(this will make the simulation very slow), and rerun.If the problem looks like a random memory might be used, try initializing memory with signaling Not-a-Number (NaN) values through the runtime option
fab.init_snan = 1
. Further useful runtime options areamrex.fpe_trap_invalid
,amrex.fpe_trap_zero
andamrex.fpe_trap_overflow
(see details in the AMReX link below).On Nvidia GPUs, if you suspect the problem might be a race condition due to a missing host / device synchronization, set the environment variable
export CUDA_LAUNCH_BLOCKING=1
and rerun.Consider simplifying your input options and re-adding more options after having found a working baseline.
Fore more information, see also the AMReX Debugging Manual.
Last but not least: the community of WarpX developers and users can help if you get stuck. Collect your above findings, describe where and what you are running and how you installed the code, describe the issue you are seeing with details and input files used and what you already tried. Can you reproduce the problem with a smaller setup (less parallelism and/or less resolution)? Report these details in a WarpX GitHub issue.
Debuggers
See the AMReX debugger section on additional runtime parameters to
disable backtraces
rethrow exceptions
avoid AMReX-level signal handling
You will need to set those runtime options to work directly with debuggers.
Typical Error Messages
By default, the code is run in Release mode (see compilation options). That means, code errors will likely show up as symptoms of earlier errors in the code instead of directly showing the underlying line that caused the error.
For instance, we have these checks in release mode
Particles shape does not fit within tile (CPU) or guard cells (GPU) used for charge deposition
Particles shape does not fit within tile (CPU) or guard cells (GPU) used for current deposition
which prevent that particles with positions that violate the local definitions of guard cells cause confusing errors in charge/current deposition.
In such a case, as described above, rebuild and rerun in Debug mode before searching further for the bug.
Usually, the bug is from NaN
or infinite
numbers assigned to particles or fields earlier in the code or from ill-defined guard sizes.
Building in debug mode will likely move the first thrown error to an earlier location in the code, which is then closer to the underlying cause.
Then, continue following the workflow above, adding more compilation guards and runtime flags that can trap array bound violations and invalid floating point values.