JDFTx  1.3.1
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Friends Macros Groups Pages
Test compiled executable

After building, in the build directory, run

make test

to check that the code runs correctly.

If all the tests fail immediately, then the executables are not able to start, likely due to a runtime library path issue. Run ./jdftx to check the specific error messages; if it is a library path issue, use 'ldd ./jdftx' to identify all the missing libraries.

If the tests take a while, but then some or all of them fail, most likely one of the linked libraries is behaving inconsistently. Try linking to alternate BLAS, LAPACK and FFT libraries, if possible. If you linked to MKL and found tests that fail, try using ForceFFTW, and if it still doesn't work, try using ThreadedBlas=no. (See Customization for more details.)

If all the tests pass in the serial mode above, then you can test it with MPI using:

make testclean
export JDFTX_LAUNCH="mpirun -n %d"
make test

This will invoke each test with a different optimal number of MPI processes (in range 1 - 4) on the same machine. (See Getting started for more details on running the code with MPI.) If any of these tests fail, check the MPI installation or try compiling with an alternate MPI, if possible.

Next, if you compiled with CUDA and have a GPU available, try:

make testclean
export JDFTX_LAUNCH=""   #disable MPI (not needed if you have 4 GPUs)
export JDFTX_SUFFIX="_gpu"
make test

If you encounter issues with the GPU test, try different CUDA compilation flags or a newer CUDA version, if appropriate.

If the tests work, then make sure the performance is reasonable. Here are some approximate timings (all tests in the latest JDFTx combined) on selected hardware to serve as a guide:

Hardware Total time
4-core laptop (i7-3632QM) 800 s
16-core workstation (2x E5-2620) 400 s
Tesla K80: one GPU (without MPI) 250 s

If the tests run substantially slower on comparable hardware and you compiled with MPI (even if you didn't issue mpirun explicitly), a very likely cause is CPU binding of threads imposed by MPI: try passing "--bind-to none" to mpirun (for OpenMPI). Also check whether linking to alternate BLAS and FFT libraries improves performance. Note that MPI will not likely speed up the test calculations much (or at all) on a single computer because JDFTx already parallelizes aggressively using pthreads to use all cores of the machine; just make sure that it does not slow down by more than 20 - 30%.

The run time on Windows/Cygwin will be substantially slower, and there is no easy fix (in the JDFTx compilation anyway).