Closed ikirker closed 6 years ago
Okay, so, installed new version of arpack-ng
.
module purge
module load gcc-libs
module load compilers/intel/2017
module load mpi/intel/2017
module load arpack-ng
wget https://github.com/dftbplus/dftbplus/archive/17.1.tar.gz
tar -xzf 17.1.tar.gz
cd dftbplus-17.1
cp sys/make.x86_64-linux-intel ./make.arch
./utils/get_opt_externals dftd3
./utils/get_opt_externals slakos
make LIB_BLAS=-mkl MKL_LIBDIR= OTHERLIBS=-qopenmp LNOPT=
make test
Hmmm. One of the tests fails as incomplete, with some FPU warnings.
In /dftbplus-17.1/_build/prog/dftb+/_autotest/spinorbit/As4S4/stderror.log
:
forrtl: warning (525): IEEE_DIVIDE_BY_ZERO is signaling
forrtl: warning (526): IEEE_INVALID is signaling
forrtl: warning (527): IEEE_OVERFLOW is signaling
forrtl: warning (528): IEEE_UNDERFLOW is signaling
16.34user 1.98system 0:02.53elapsed 724%CPU (0avgtext+0avgdata 24028maxresident)k
8inputs+544outputs (0major+9188minor)pagefaults 0swaps
In dftbplus-17.1/_build/prog/dftb+/_autotest/spinorbit/As4S4/output
:
*** Geometry step: 0
iSCC Total electronic Diff electronic SCC error
1 -0.16934620E+02 0.00000000E+00 0.61570042E+00
2 -0.16926831E+02 0.77892704E-02 0.63517628E+00
3 -0.16967559E+02 -0.40728812E-01 0.36731560E+00
4 -0.16987855E+02 -0.20295601E-01 0.91602230E-02
5 -0.16987866E+02 -0.11133467E-04 0.45408560E-02
6 -0.16987866E+02 -0.12033076E-06 0.16752061E-02
ERROR!
-> Failure in ZHEEVR 1
Setting OMP_NUM_THREADS
to 1 and trying again, just to make sure it's not a weird parallelism problem.
...
Succeeds at 1, 4, 8, 10, 11 threads. Fails at 12, 13, 24 threads.
:man_facepalming:
I think this might be chip-code-path-specific, too. Failed only on Legion's login nodes, succeeded on U-type nodes at same and higher thread counts. Since it was requested on Thomas, I'm testing it there now.
Okay, so all the implemented tests pass on Thomas on 24 threads. I had to remove the unimplemented tests from the list, because otherwise they're marked as failed :roll_eyes: and the tests all fail. Once you remove them, everything completes successfully, and then it moves onto another test step, which uses Python and numpy, which passes if you load the Python module.
I think I can sed out the unimplemented tests in the buildscript to make this work. The relevant file is: dftbplus-17.1/test/prog/dftb+/tests
, the relevant tests are sockets/H2O
and sockets/diamond
.
Ah, and the install prefix is set in make.config
but can be set on the command line, so something like this:
wget https://github.com/dftbplus/dftbplus/archive/17.1.tar.gz
module purge
module load gcc-libs
module load compilers/intel/2017
module load mpi/intel/2017
module load arpack-ng
module load python2
tar -xzf 17.1.tar.gz
cd dftbplus-17.1
cp sys/make.x86_64-linux-intel ./make.arch
./utils/get_opt_externals dftd3
./utils/get_opt_externals slakos
make LIB_BLAS=-mkl MKL_LIBDIR= OTHERLIBS=-qopenmp LNOPT= INSTALLDIR=$install_prefix
sed -i -e '/^sockets\/H2O/d' -e '/^sockets\/diamond/d' test/prog/dftb+/tests
make test
make INSTALLDIR=$install_prefix install
Have plugged into proper build script and installed on Thomas.
Checked back with originator of IN.
No response from INO, closing.
(IN:02543102)
From: http://www.dftbplus.org/download/
OpenMP only, needs BLAS and ARPACK-NG.
Looks like it'll need a new build of ARPACK-NG for compatibility: lib change between Intel 2015 vs 2017, giving the following error during compilation:
Currently working it through, but it looks like this should work: