On Fri, 19 Sep 2014 00:27:50 +0200 Jan Blechta <[email protected]> wrote:
> Yes, after many trials using > > $ cd test/unit/io/python > $ while true; do git clean -fdx && mpirun -n 3 xterm -e gdb -ex r -ex > q -args python -m pytest -sv; done > # when it hangs and you interrupt it, it asks for confirmation for > # quitting, so you type n and enjoy gdb... > > I've seen a situation when 2 processes deadlocked on > HDF5Interface::close_file() in DOLFIN with backtrace like > > # MPI barrier > ... > # MPI close > # HDF5 lib calls > H5FClose() > dolfin::HDF5Interface::close_file() > dolfin::HDF5File::close() > dolfin::HDF5File::~HDF5File() > dolfin::HDF5File::~HDF5File() > # smart ptr management > # garbage collection > > while 3rd process is waiting far away. Isn't it strange that > destructor is there twice in stacktrace? (The upper one is on '}' > line which I don't get.) What does it mean? Probably just code generation artifact - nothing harmful, see http://stackoverflow.com/a/15244091/1796717 Jan > > Jan > > > On Thu, 18 Sep 2014 16:20:51 +0200 > Martin Sandve Alnæs <[email protected]> wrote: > > > I've added the mpi fixes for temppath fixture and fixed > > some other related issues while at it: When parameterizing > > a test that uses a temppath fixture, there is a need for > > separate directories for each parameter combo. > > A further improvement would be automatic cleaning of old tempdirs, > > but I leave that for now. > > > > I've pushed these changes to the branch > > aslakbergersen/topic-change-unittest-to-pytest > > > > The tests still hang though, in the closing of HDF5File. > > > > Here's now to debug if someone wants to give it a shot: > > Just run: > > mpirun -np 3 python -m pytest -s -v > > With gdb: > > mpirun -np 3 xterm -e gdb --args python -m pytest > > then enter 'r' in each of the three xterms. > > > > You may have to try a couple of times to get the hanging behaviour. > > > > Martin > > > > On 18 September 2014 13:23, Martin Sandve Alnæs <[email protected]> > > wrote: > > > > > Good spotting both of you, thanks. > > > > > > Martin > > > > > > On 18 September 2014 13:01, Lawrence Mitchell < > > > [email protected]> wrote: > > > > > >> On 18/09/14 11:42, Jan Blechta wrote: > > >> > Some problems (when running in a clean dir) are avoided using > > >> > this (although incorrect) patch. There are race conditions in > > >> > creation of temp dir. It should be done using atomic operation. > > >> > > > >> > Jan > > >> > > > >> > > > >> > ================================================================== > > >> > diff --git a/test/unit/io/python/test_XDMF.py > > >> > b/test/unit/io/python/test_XDMF.py index 9ad65a4..31471f1 > > >> > 100755 --- a/test/unit/io/python/test_XDMF.py > > >> > +++ b/test/unit/io/python/test_XDMF.py > > >> > @@ -28,8 +28,9 @@ def temppath(): > > >> > filedir = os.path.dirname(os.path.abspath(__file__)) > > >> > basename = os.path.basename(__file__).replace(".py", > > >> > "_data") temppath = os.path.join(filedir, basename, "") > > >> > - if not os.path.exists(temppath): > > >> > - os.mkdir(temppath) > > >> > + if MPI.rank(mpi_comm_world()) == 0: > > >> > + if not os.path.exists(temppath): > > >> > + os.mkdir(temppath) > > >> > return temppath > > >> > > >> There's still a race condition here because ranks other than zero > > >> might try and use temppath before it's created. I think you want > > >> something like the below: > > >> > > >> if MPI.rank(mpi_comm_world()) == 0: > > >> if not os.path.exists(temppath): > > >> os.mkdir(temppath) > > >> MPI.barrier(mpi_comm_world()) > > >> return temppath > > >> > > >> If you're worried about the OS not creating files atomically, you > > >> can always mkdir into a tmp directory and then os.rename(tmp, > > >> temppath), since posix guarantees that renames are atomic. > > >> > > >> Lawrence > > >> _______________________________________________ > > >> fenics mailing list > > >> [email protected] > > >> http://fenicsproject.org/mailman/listinfo/fenics > > >> > > > > > > > > _______________________________________________ > fenics mailing list > [email protected] > http://fenicsproject.org/mailman/listinfo/fenics _______________________________________________ fenics mailing list [email protected] http://fenicsproject.org/mailman/listinfo/fenics
