March 2017 - firedrake - imperial.ac.uk

firedrake-install script fails using wrong C++ compiler
by Buesing, Henrik 09 Mar '17

09 Mar '17

Dear Firedrake developers, I'm trying to install Firedrake with the install script on CentOS Linux release 7.3.1611. I'm doing a "firedrake-install --no_package_manager --disable-ssh". The script fails when building pyop2/sparsity.cpp, because it did not find mpi.h. I notice that it uses gcc for compiling while it should use the appropriate MPI Wrapper function mpicxx. There are environment variables available in my system ($MPICC, $MPICXX, $MPIFC) pointing to the appropriate MPI Wrappers (mpicc, mpicxx, mpif90). How do I tell the install script to use these? I notice the option "--mpicc" for the C compiler, but how do I set the C++ (and Fortran) compiler? Thank you! Henrik -- Dipl.-Math. Henrik Büsing Institute for Applied Geophysics and Geothermal Energy E.ON Energy Research Center RWTH Aachen University ------------------------------------------------------ Mathieustr. 10 | Tel +49 (0)241 80 49907 52074 Aachen, Germany | Fax +49 (0)241 80 49889 ------------------------------------------------------ http://www.eonerc.rwth-aachen.de/GGE hbuesing(a)eonerc.rwth-aachen.de<mailto:hbuesing@eonerc.rwth-aachen.de> ------------------------------------------------------

1 0

Running firedrake with nohup
by William Booker 06 Mar '17

06 Mar '17

Dear Firedrakers, I'm trying to run some Firedrake runs on HPC, but I keep killing my run whenever I close the terminal I've ssh'ed in with. I was using nohup: nohup mpirun -n 16 python run.py > output.out Is there another way to get firedrake to run when I close the connection to my local computer? Thanks Will

4 3

Re: [firedrake] Nans and infs
by Andrew McRae 05 Mar '17

05 Mar '17

:) On 5 March 2017 at 16:36, Francis Poulin <fpoulin(a)uwaterloo.ca> wrote: > Hello Andrew, > > Quick check shows that it does work nicely using less than (lt). > > Thanks for encouraging me to write better code! > > Cheers, Francis > > > ------------------ > Francis Poulin > Associate Professor > Department of Applied Mathematics > University of Waterloo > > email: fpoulin(a)uwaterloo.ca > Web: https://uwaterloo.ca/poulin-research-group/ > Telephone: +1 519 888 4567 x32637 <(519)%20888-4567> > > ------------------------------ > *From:* Francis Poulin > *Sent:* Sunday, March 05, 2017 11:32 AM > *To:* firedrake(a)imperial.ac.uk > *Subject:* RE: [firedrake] Nans and infs > > Good question. > > I agree it's probably better to avoid evaluating a function that divides > by zero. > > In the expression I have a conditional expression that uses less than or > equal to something. It occurs to me that if I use only less than (but not > equal to) that should do the trick. > > I would guess that I should change le to lt but not sure. Do you think > this is what I should try? > > Cheers, > Francis > > > ------------------ > Francis Poulin > Associate Professor > Department of Applied Mathematics > University of Waterloo > > email: fpoulin(a)uwaterloo.ca > Web: https://uwaterloo.ca/poulin-research-group/ > Telephone: +1 519 888 4567 x32637 <(519)%20888-4567> > > ------------------------------ > *From:* firedrake-bounces(a)imperial.ac.uk [firedrake-bounces(a)imperial.ac.uk] > on behalf of Andrew McRae [A.T.T.McRae(a)bath.ac.uk] > *Sent:* Sunday, March 05, 2017 11:15 AM > *To:* firedrake(a)imperial.ac.uk > *Subject:* Re: [firedrake] Nans and infs > > Pardon the obvious question, but is there a reason you don't simply > rewrite the expression to avoid generating NaN/Inf in the first place? > > Replacing Inf with 0 also seems like a strange thing to do! > > Andrew > > On 3 March 2017 at 21:15, Francis Poulin <fpoulin(a)uwaterloo.ca> wrote: > >> Hello, >> >> I am doing a stability calculation for different profiles. >> >> The test case that I have used is Gaussian, where the velocity and >> gradient of PV is defined using something like this, >> >> Ub = Function(V).interpolate(Uj*exp(-(x[0]-Ly/2)**2)) >> dQb = Function(V).interpolate(-2.*Uj*exp(-(x[0]-Ly/2)**2)*(2.*(x[0]-Ly/2)**2 >> - 1.)) >> >> I have compared this with the results of using a spectral method and the >> results agree very well. >> >> I am now considering a less trivial example using something called a Bump >> function. Yes, technical name is Bump. Thanks to previous help I have >> figured out how to define it piecewise (thanks again). >> >> Ub = Function(V).interpolate(conditional(le(abs(x[0]-Ly/2.), 1.0), >> exp(-1./(1. -(x[0]-Ly/2)**2)), 0.0)) >> dQb = Function(V).interpolate(conditional(le(abs(x[0]-Ly/2.), 1.0), >> -2.*exp(-1./(1. -(x[0]-Ly/2)**2))*(3.*(x[0]-Ly/2.)**4 - >> 1)/((x[0]-Ly/2.)**2-1)**4, 0.0)) >> >> However, when I try this I get zero growth rate, which does not agree >> with the spectral method. >> >> Actually, in the spectral method we need to zero out the Nans and Infs in >> expressions, and that works out nicely. Is there an easy way to convert >> Nans and Infs to zero in interpolate? >> >> I tried using both a Krylov method and Lapack. The first gives zero, >> which I don't believe, and the second fails, I presume because I am >> dividing by zero somewhere. >> >> Any help would be greatly appreciated! >> >> Cheers, Francis >> >> >> >> ------------------ >> Francis Poulin >> Associate Professor >> Department of Applied Mathematics >> University of Waterloo >> >> email: fpoulin(a)uwaterloo.ca >> Web: https://uwaterloo.ca/poulin-research-group/ >> Telephone: +1 519 888 4567 x32637 <(519)%20888-4567> >> >> >

1 0

Re: [firedrake] Nans and infs
by Andrew McRae 05 Mar '17

05 Mar '17

Pardon the obvious question, but is there a reason you don't simply rewrite the expression to avoid generating NaN/Inf in the first place? Replacing Inf with 0 also seems like a strange thing to do! Andrew On 3 March 2017 at 21:15, Francis Poulin <fpoulin(a)uwaterloo.ca> wrote: > Hello, > > I am doing a stability calculation for different profiles. > > The test case that I have used is Gaussian, where the velocity and > gradient of PV is defined using something like this, > > Ub = Function(V).interpolate(Uj*exp(-(x[0]-Ly/2)**2)) > dQb = Function(V).interpolate(-2.*Uj*exp(-(x[0]-Ly/2)**2)*(2.*(x[0]-Ly/2)**2 > - 1.)) > > I have compared this with the results of using a spectral method and the > results agree very well. > > I am now considering a less trivial example using something called a Bump > function. Yes, technical name is Bump. Thanks to previous help I have > figured out how to define it piecewise (thanks again). > > Ub = Function(V).interpolate(conditional(le(abs(x[0]-Ly/2.), 1.0), > exp(-1./(1. -(x[0]-Ly/2)**2)), 0.0)) > dQb = Function(V).interpolate(conditional(le(abs(x[0]-Ly/2.), 1.0), > -2.*exp(-1./(1. -(x[0]-Ly/2)**2))*(3.*(x[0]-Ly/2.)**4 - > 1)/((x[0]-Ly/2.)**2-1)**4, 0.0)) > > However, when I try this I get zero growth rate, which does not agree with > the spectral method. > > Actually, in the spectral method we need to zero out the Nans and Infs in > expressions, and that works out nicely. Is there an easy way to convert > Nans and Infs to zero in interpolate? > > I tried using both a Krylov method and Lapack. The first gives zero, > which I don't believe, and the second fails, I presume because I am > dividing by zero somewhere. > > Any help would be greatly appreciated! > > Cheers, Francis > > > > ------------------ > Francis Poulin > Associate Professor > Department of Applied Mathematics > University of Waterloo > > email: fpoulin(a)uwaterloo.ca > Web: https://uwaterloo.ca/poulin-research-group/ > Telephone: +1 519 888 4567 x32637 <(519)%20888-4567> > >

2 2

Nans and infs
by Francis Poulin 04 Mar '17

04 Mar '17

Hello, I am doing a stability calculation for different profiles. The test case that I have used is Gaussian, where the velocity and gradient of PV is defined using something like this, Ub = Function(V).interpolate(Uj*exp(-(x[0]-Ly/2)**2)) dQb = Function(V).interpolate(-2.*Uj*exp(-(x[0]-Ly/2)**2)*(2.*(x[0]-Ly/2)**2 - 1.)) I have compared this with the results of using a spectral method and the results agree very well. I am now considering a less trivial example using something called a Bump function. Yes, technical name is Bump. Thanks to previous help I have figured out how to define it piecewise (thanks again). Ub = Function(V).interpolate(conditional(le(abs(x[0]-Ly/2.), 1.0), exp(-1./(1. -(x[0]-Ly/2)**2)), 0.0)) dQb = Function(V).interpolate(conditional(le(abs(x[0]-Ly/2.), 1.0), -2.*exp(-1./(1. -(x[0]-Ly/2)**2))*(3.*(x[0]-Ly/2.)**4 - 1)/((x[0]-Ly/2.)**2-1)**4, 0.0)) However, when I try this I get zero growth rate, which does not agree with the spectral method. Actually, in the spectral method we need to zero out the Nans and Infs in expressions, and that works out nicely. Is there an easy way to convert Nans and Infs to zero in interpolate? I tried using both a Krylov method and Lapack. The first gives zero, which I don't believe, and the second fails, I presume because I am dividing by zero somewhere. Any help would be greatly appreciated! Cheers, Francis ------------------ Francis Poulin Associate Professor Department of Applied Mathematics University of Waterloo email: fpoulin(a)uwaterloo.ca Web: https://uwaterloo.ca/poulin-research-group/ Telephone: +1 519 888 4567 x32637

2 2

Point evaluations of a function
by William Booker 02 Mar '17

02 Mar '17

Dear Firedrakers, This is a somewhat similar query to what Onno asked earlier this week about evaluating a function at a point in Firedrake. It's a bit easier because the point evaluation comes at the boundary of the domain, so we're currently including it through ds(1) integrals. The attached script is a Crank-Nicholson weak form for equations 11-14 in the attached pdf. The point evaluation comes in at equation 28 where the boundary inflow evolution can be combined with the diffusion equation. The current script works when theta = 0, so that the non-linear parts are known so we can use a linear solver. When we choose a non zero theta, we get the following error for a nonlinear solve: h_solver.solve() File "/usr/not-backed-up/firedrake/2017-02-28/firedrake/src/firedrake/firedrake/variational_solver.py", line 223, in solve solving_utils.check_snes_convergence(self.snes) File "/usr/not-backed-up/firedrake/2017-02-28/firedrake/src/firedrake/firedrake/solving_utils.py", line 160, in check_snes_convergence %s""" % (snes.getIterationNumber(), msg)) firedrake.exceptions.ConvergenceError: Nonlinear solve failed to converge after 0 nonlinear iterations. Reason: Inner linear solve failed to converge after 0 iterations with reason: DIVERGED_NANORINF We believe it may be due to having the unknown boundary flux evaluation. What can we do to fix this for a nonzero theta? Thanks Will

2 1

Re: [firedrake] problems with notebook
by David Ham 01 Mar '17

01 Mar '17

Hi Francis, That's a good idea. We should have a firedrake-update --environment to also upgrade all the python packages in the virtualenv. Cheers, David On Wed, 1 Mar 2017 at 10:58 Francis Poulin <fpoulin(a)uwaterloo.ca> wrote: > Hello Lawrence, > > Yes, that did the trick. Sorry for the bother. > > I realize that normally, we might not want to update everything, including > ipython, but is there an option if we wanted to update everything? Might > be helpful for people like me who don't know exactly what to update, like > ipykernel. > > Not to see if I can do the exercises in the notebook. > > Cheers, Francis > > ------------------ > Francis Poulin > Associate Professor > Department of Applied Mathematics > University of Waterloo > > email: fpoulin(a)uwaterloo.ca > Web: https://uwaterloo.ca/poulin-research-group/ > Telephone: +1 519 888 4567 x32637 <(519)%20888-4567> > > > ________________________________________ > From: firedrake-bounces(a)imperial.ac.uk [firedrake-bounces(a)imperial.ac.uk] > on behalf of Lawrence Mitchell [lawrence.mitchell(a)imperial.ac.uk] > Sent: Wednesday, March 01, 2017 10:38 AM > To: firedrake(a)imperial.ac.uk > Subject: Re: [firedrake] problems with notebook > > > On 1 Mar 2017, at 10:34, Francis Poulin <fpoulin(a)uwaterloo.ca> wrote: > > > > Thanks for the suggestion. I thought that firedrake-update would update > ipython, but I guess not. I will try and remember do that from time to > time. > > Yes, although it runs the upgrade cycle for direct firedrake components, > we don't upgrade every package installed inside the virtualenv (since this > would take too long, and we don't directly control all of that). > > > As you can see below, I have now joined the modern times and have > version 5.3.0, but it still seems to have problem with updating the > kernel. I did try it again with jupyter and it still failed. > > > > Actually, when I try ipython notebook it fails as well. > > > > Should I remove ipython and try reinstalling it? > > > Oh, sorry, I think the error was not directly ipython. Please also try: > > $ pip install -U jupyter > $ pip install -U ipykernel > > and then go round again... > > Thanks, > > Lawrence > _______________________________________________ > firedrake mailing list > firedrake(a)imperial.ac.uk > https://mailman.ic.ac.uk/mailman/listinfo/firedrake > > _______________________________________________ > firedrake mailing list > firedrake(a)imperial.ac.uk > https://mailman.ic.ac.uk/mailman/listinfo/firedrake > -- Dr David Ham Department of Mathematics Imperial College London

1 0

problems with notebook
by Francis Poulin 01 Mar '17

01 Mar '17

Hello again, I was looking over the firedrake notebooks (thanks for putting those together!) and decided to try running the one that involves adjoint. https://nbviewer.jupyter.org/github/firedrakeproject/firedrake/blob/master/… I did install scipy and adjoint, as requested, but get an error when I try plotting u_init axes = plot(u_init); The error is copied below. Is there something else that I need to install for this to work? Cheers, Francis --------------------------------------------------------------------------- CompilationError Traceback (most recent call last) <ipython-input-11-ab90b4141777> in <module>() 1 u_init, p_init = w.split() ----> 2 axes = plot(u_init); 3 #axes.set_aspect("equal") 4 #axes.set_title("Velocity") 5 #axes = plot(p_init) /home/fpoulin/software/firedrake/src/firedrake/firedrake/plot.pyc in plot(function_or_mesh, num_sample_points, axes, plot3d, **kwargs) 293 coords = function.ufl_domain().coordinates.dat.data_ro 294 X, Y = coords.T --> 295 vals = np.asarray(function.at(coords)) 296 C = np.linalg.norm(vals, axis=1) 297 U, V = vals.T /home/fpoulin/software/firedrake/src/firedrake/firedrake/function.pyc in at(self, arg, *args, **kwargs) 532 else: 533 p_result = np.empty(value_shape, dtype=float) --> 534 single_eval(points[i:i+1], p_result) 535 l_result.append((i, p_result)) 536 except PointNotInDomainError: /home/fpoulin/software/firedrake/src/firedrake/firedrake/function.pyc in single_eval(x, buf) 510 def single_eval(x, buf): 511 """Helper function to evaluate at a single point.""" --> 512 err = self._c_evaluate(self._ctypes, 513 x.ctypes.data_as(POINTER(c_double)), 514 buf.ctypes.data_as(POINTER(c_double))) /home/fpoulin/software/firedrake/src/PyOP2/pyop2/utils.pyc in __get__(self, obj, cls) 62 if obj is None: 63 return self ---> 64 obj.__dict__[self.__name__] = result = self.fget(obj) 65 return result 66 /home/fpoulin/software/firedrake/src/firedrake/firedrake/function.pyc in _c_evaluate(self) 460 @utils.cached_property 461 def _c_evaluate(self): --> 462 result = make_c_evaluate(self) 463 result.argtypes = [POINTER(_CFunction), POINTER(c_double), POINTER(c_double)] 464 result.restype = c_int /home/fpoulin/software/firedrake/src/firedrake/firedrake/function.pyc in make_c_evaluate(function, c_name, ldargs) 610 ["-I%s/include" % d for d in get_petsc_dir()], 611 ldargs=ldargs, --> 612 comm=function.comm) /home/fpoulin/software/firedrake/src/PyOP2/pyop2/compilation.pyc in load(src, extension, fn_name, cppargs, ldargs, argtypes, restype, compiler, comm) 328 raise CompilationError("Don't know what compiler to use for platform '%s'" % 329 platform) --> 330 dll = compiler.get_so(src, extension) 331 332 fn = getattr(dll, fn_name) /home/fpoulin/software/firedrake/src/PyOP2/pyop2/compilation.pyc in get_so(self, src, extension) 166 Unable to compile code 167 Compile log in %s --> 168 Compile errors in %s""" % (e.cmd, e.returncode, logfile, errfile)) 169 else: 170 cc = [self._cc] + self._cppargs + \ CompilationError: Command "['mpicc', '-std=c99', '-fPIC', '-Wall', '-march=native', '-O3', '-fno-ivopts', '-I/home/fpoulin/software/firedrake/src/firedrake/firedrake', '-I/usr/include', '-I/home/fpoulin/software/firedrake/lib/python2.7/site-packages/petsc/include', '-o', '/tmp/pyop2-cache-uid1000/7137338dc043a2dcfd35bcc4ecc64e90_p6844.so.tmp', '/tmp/pyop2-cache-uid1000/7137338dc043a2dcfd35bcc4ecc64e90_p6844.c', '-shared', '-L/usr/lib', '-lspatialindex_c', '-Wl,-rpath,/usr/lib']" return error status 1. Unable to compile code Compile log in /tmp/pyop2-cache-uid1000/7137338dc043a2dcfd35bcc4ecc64e90_p6844.log Compile errors in /tmp/pyop2-cache-uid1000/7137338dc043a2dcfd35bcc4ecc64e90_p6844.err ------------------ Francis Poulin Associate Professor Department of Applied Mathematics University of Waterloo email: fpoulin(a)uwaterloo.ca Web: https://uwaterloo.ca/poulin-research-group/ Telephone: +1 519 888 4567 x32637

2 10