March 2018 - firedrake - imperial.ac.uk

firedrake-install currently failing due to network issues at Argonne
by Lawrence Mitchell 29 Mar '18

29 Mar '18

Dear firedrakers, As of early this morning, firedrake-install/update have not been able to complete: failing when installing PETSc. This is, unfortunately, due to circumstances beyond our control. Argonne, where PETSc is hosted, has taken down one of the servers necessary in this process due to a "cyber incident". It is unclear when it will be back on line at this point. Please see https://lists.mcs.anl.gov/pipermail/petsc-users/2018-March/034836.html for more details. I will update when we expect installation to be working properly again. Apologies! Lawrence

1 1

Current build failures for Firedrake
by Lawrence Mitchell 23 Mar '18

23 Mar '18

Dear all, some of you who are attempting to install firedrake at the moment may be noticing that it doesn't work. Sorry! This was my fault. I am trying to figure out why it is that the PETSc builds fail. I'll update when I've fixed things. Cheers, Lawrence

1 1

Re: [firedrake] Firedrake test script not advancing
by Andrew McRae 21 Mar '18

21 Mar '18

Owen, I'm guessing you don't care about running in parallel, at least not immediately, so don't let this stop you from getting started. If it really bugs you, make a 16.04 virtual machine instead. This is the most recent LTS (long term support) Ubuntu release, and gets maintenance updates for 5 years from April 2016, by which time you'll have graduated. Hopefully ;) On the other hand, 17.10 only gets updates for 9 months from October 2017. (I'm assuming you're running a VM just for Firedrake, not for anything else you do) Andrew On 19 March 2018 at 18:02, Lawrence Mitchell < lawrence.mitchell(a)imperial.ac.uk> wrote: > > > > On 19 Mar 2018, at 17:52, Ham, David A <david.ham(a)imperial.ac.uk> wrote: > > > > I’m in favour of dumping OpenMPI. It also causes us parallel test issues > cf https://github.com/firedrakeproject/firedrake/issues/1199 > > > > Are there any downsides with going to mpich? > > All the ubuntu packages that link against MPI are built against OpenMPI. > We need to be very careful that when we install on such systems that we > don't accidentally end up getting the wrong one. > > Lawrence > _______________________________________________ > firedrake mailing list > firedrake(a)imperial.ac.uk > https://mailman.ic.ac.uk/mailman/listinfo/firedrake >

2 1

Firedrake test script not advancing
by Owen Pembery 19 Mar '18

19 Mar '18

Dear Firedrakers, I've just installed Firedrake on an Ubuntu virtual machine, and I'm finding that the firedrake test script seemingly stops advancing (sometimes at different points in the tests). By `stops advancing' I mean `stays on the same test for at least an hour'. I've attached the console output for a couple of attempts running the test script. Is there something I'm doing wrong, is this a bug, or is it an issue with the virtual machine? (Or something else?) Thanks, Owen

3 6

Poisson equation; homogeneous vs nonhomogeneous
by Eric M 19 Mar '18

19 Mar '18

My impression from reading the manual is that for a Poisson problem with non-homogeneous boundary conditions, the conversion to a homogeneous problem takes place and the boundary integrals are taken care of. So why are my results much better when converting the problem to a homogeneous problem myself? Here are 2 different Poisson codes (degree 3 Lagrange elements) and their outputs: nref=7 for j in range(nref): mesh = UnitSquareMesh(2**(j+1), 2**(j+1)) mesh.coordinates.dat.data[:, 0] = 1.0 - mesh.coordinates.dat.data[:, 0] V = FunctionSpace(mesh, "Lagrange", 3) u = TrialFunction(V) v = TestFunction(V) x, y = SpatialCoordinate(mesh) f = Function(V) f.interpolate(-(2+x**2+y**2)*exp((x**2+y**2)/2)) g=Function(V) g.interpolate(exp((x**2+y**2)/2)) a = (dot(grad(v), grad(u)) ) * dx L = f * v * dx bcs = [DirichletBC(V, g, (1, 2, 3, 4))] u = Function(V) solve(a == L, u, bcs=bcs) #u=u+g #this method works better here; but not for time marching L2error[j]=sqrt(assemble(dot(u - g, u - g) * dx)) H1error[j]=sqrt(assemble(dot(grad(u-g),grad(u-g))*dx)) #L=f*v*dx-dot(grad(g),grad(v))*dx #bcs = [DirichletBC(V, 0, (1, 2, 3, 4))] #if only want one boundary specified: for example use (1,) I think output: L2 errors n= 0 0.00043533120955450153 n= 1 2.6722563512459733e-05 n= 2 1.6269810018255959e-06 n= 3 1.6694766615405184e-07 n= 4 1.1739111863202767e-06 n= 5 7.280913725610821e-06 n= 6 2.0327589385362215e-05 L2 convergence rate 4.025983134615211 4.037789103155838 3.2847295743419402 -2.813855433236994 -2.632796250527484 -1.48124771918039 H1 errors n= 0 0.005033670038757814 n= 1 0.0007002176561563797 n= 2 9.299757283782469e-05 n= 3 1.2089747635515502e-05 n= 4 7.532429204089909e-06 n= 5 3.3080917711483075e-05 n= 6 9.162160804557016e-05 H1 convergence rate 2.8457353045032385 2.912538471487578 2.9434089338107663 0.682597016344028 -2.134812144094874 -1.4696886262688682 Now making the substition u=w+g and solving for w (which has homogeneous boundary condition) ... a = (dot(grad(v), grad(u)) ) * dx L=f*v*dx-dot(grad(g),grad(v))*dx bcs = [DirichletBC(V, 0, (1, 2, 3, 4))] u = Function(V) solve(a == L, u, bcs=bcs) u=u+g output: L2 errors n= 0 0.00043533082808728233 n= 1 2.672295570362407e-05 n= 2 1.6257969941867753e-06 n= 3 1.0013920117782886e-07 n= 4 6.2201879296317444e-09 n= 5 3.878122859276369e-10 n= 6 2.4212950203939354e-11 L2 convergence rate 4.025960697004482 4.03886055546273 4.0210683688015765 4.008904872857674 4.003527754961185 4.001507729908028 H1 errors n= 0 0.0050336743129552495 n= 1 0.0007001812028177674 n= 2 9.279064302630619e-05 n= 3 1.1994409777711576e-05 n= 4 1.527217787021703e-06 n= 5 1.9276952639853355e-07 n= 6 2.4216339384015273e-08 H1 convergence rate 2.845811638200329 2.915677094480334 2.9516171647844374 2.9733844513983505 2.9859569015421905 2.9928242994816663 Obviously the latter problem gives better results. Why is this?

2 1

Workshop on Debugging Numerical Software 4th and 5th Jun 2018, Bath
by Eike Mueller 17 Mar '18

17 Mar '18

Dear Firedrakers, the workshop below might be of interest for some of you. It will take place in Bath on Mon 4th, Tue 5th June 2018, the same week as the Firedrake conference. It would be great to have some representation from the Firedrake community at our workshop. Eike ========= Dear all, we would like to invite you to a two day === Workshop on Debugging Numerical Software === at the University of Bath (UK) on Mon 4th -Tue 5th June 2018 The goal of the workshop is to bring together research software engineers from industry and academia to talk about this important but - as we feel - often neglected topic. As I’m sure you all know from long experience, a lot of time “writing” code is actually spent hunting for (sometimes very elusive) bugs, yet there seems to be limited general advise or consensus on how to make this often arduous task easier. Instead of concentrating on the technicalities, specific debuggers or particular programs, the workshop aims to explore common themes, discuss general strategies and share experiences. Although the focus will be on numerical software, developers from other discipplines are very welcome. The workshop will consist of the following elements: * TALKS by specialists from industry and government organisations (Numerical Algorithms Group [NAG], Allinea/ARM, CRAY, Met Office) * an interactive BUG HUNTING session * small group DISCUSSION SESSIONS on topics such as - teaching debugging - what makes debugging (numerical software) hard? - are there any general principles/strategies to be used? - using tools (or is adding print statements enough?) - classification of bugs - is it possible to write bug-free code? - ensuring correctness of numerical codes in the presence of rounding errors. - debugging codes with varying output (e.g. stochastic simulations and Monte Carlo) We will aim to produce some sort of tangible output, such as speed blog(s) and/or a whitepaper on debugging numerical software. We also hope that the workshop might also lead to the development of new training material on the topic. Further details and a preliminary schedule can be found on the workshop webpage https://sites.google.com/site/debuggingnumericalsoftware/ *** Registration *** To express your interest in the workshop, please fill out the following google form: https://sites.google.com/site/debuggingnumericalsoftware/registration To allow planning of catering etc. we ask you to ===== REGISTER BY TUE, 1ST MAY ===== Please note that the number of spaces is limited and filling out the form will NOT confirm your place yet. We will get in touch with you as soon as possible to confirm your attendance. DO NOT MAKE ANY TRAVEL/ACCOMMODATION ARRANGEMENTS BEFORE YOU RECEIVED AN EMAIL FROM US CONFIRMING YOUR ATTENDANCE. *** Bug hunting session *** One of the central parts of the workshop will be an interactive bug hunting session. For this we would like the participants to share interesting bugs they have encountered (and maybe even successfully fixed). We will then debug and investigate those bugs in small groups to form the basis for subsequent discussions. Ideally, participants should extract their bug and bring it with them in the form of a simple program which can be shared with other participants. Please provide further details on any bugs you can bring on the registration form and make sure that you will bring a laptop. *** Accommodation *** Please arrange your own accommodation in Bath. There is a large number of hotels for different budgets in the city centre, see section “Accommodation” on the webpage for a small selection. Since Bath is a popular tourist destination we advise you to book early (but not before your place is confirmed!) *** Travel arrangements *** Bath is very easy to reach by train, it’s only 15 minutes from Bristol and on the main train line to London Paddington (which can be reached in 1.5 hours from Bath). Please see the “Travelling to Bath” section on the webpage for further details. Information on how to get to the University can be found on the “Workshop Location” section on the same webpage. *** Workshop fee and financial support *** No fee is required to attend the workshop. Tea/coffee and lunch will be provided free of charge on both days and there will be a conference dinner on Monday evening in the Aqua Italia restaurant in Bath which is paid for by our sponsors. We will not normally cover costs for travel and accommodation. However, to allow junior participants such as PhD students and RSEs in the early stages of their career to attend, we will be able to provide a small number of travel bursaries. To apply for this financial support, please email us the following information (2-3 sentences for each bullet point) by 1 May 2018: * your current status and stage of career * your main interests * your motivation for attending the workshop * brief justification why travel support would be important for you Please also arrange for your supervisor or manager to send us an email to confirm your status and to provide support for your application. Any decisions on support will be at the discretion of the organisers. The workshop is organised by James Grant as local RSE in Bath and Eike Mueller as a new Fellow of the Software Sustainability Institute. It has been made possible through generous financial support from the Software Sustainability Institute (www.software.ac.uk<http://www.software.ac.uk/>) and the Institute of Mathematical Innovation in Bath (http://www.bath.ac.uk/imi/) James and Eike

1 0

Firedrake can now install icepack for you.
by Ham, David A 15 Mar '18

15 Mar '18

Dear Firedrakers, We’ve just added icepack to the list of external packages that you can add to your Firedrake install. You can install icepack by adding --install icepack to the firedrake-install or firedrake-update. For example: firedrake-update --install icepack Icepack is Dan Shapero’s glacier and ice sheet modelling package built on Firedrake (https://icepack.github.io) We’re really excited to make installing icepack easy for users. Other packages which we’ve bundled into the firedrake installer are gusto, thetis, and pyadjoint. You can always see the current list of packages by typing: firedrake-update --install-help If you maintain a released user package built on Firedrake then we’d also love to make it available by the same mechanism. The requirements are: 1. The package must be released and ready for users, including suitable documentation. 2. The package must be actively maintained. 3. The package must be in a publicly accessible git repository. 4. The package must be a Python package installable in editable mode (i.e. "pip install -e ." must work in the top directory of the package). We look forward to adding more packages to the installer. Regards, David PS: While I’ve got your attention, don’t forget to register and submit an abstract to Firedrake ’18. We’re really looking forward to meeting our user community and hearing what Firedrake does (and doesn’t!) do for you. https://www.firedrakeproject.org/firedrake_18.html

1 0

openMPI 3.0 testing issues
by Jack Betteridge 14 Mar '18

14 Mar '18

Hi, I am trying to fix my firedrake install. I appear to be failing all of the parallel tests (200 or so). I have narrowed the problem down to the way pytest is calling MPI and have a minimal failing example: from subprocess import check_call from mpi4py import MPI call = ["mpiexec", "-n", "1", "hostname"] check_call(call) comment out mpi4py line and it runs fine In anticipation of XY problem, I'm just trying to pass the firedrake tests. (this is openMPI 3.0.0 on my Arch Linux install, but verified failing on another system also) I'm not yet on the mailing list, but this is cross posted in slack. Cheers, Jack

2 1

Reminder: Firedrake user and developer workshop 7-8 June 2018
by Lawrence Mitchell 09 Mar '18

09 Mar '18

Dear all, a friendly reminder that we are holding the second Firedrake user and developers workshop this summer on 7 & 8 of June at Imperial. If you are doing exciting things with Firedrake, we'd love to hear from you. It is also an opportunity to harangue us in person about missing features. A reminder that both abstract submission and early registration closes on: 10th April For more details, see https://www.firedrakeproject.org/firedrake_18.html Cheers, Lawrence

1 0

From libadjoint to pyadjoint
by Bock, Andreas 09 Mar '18

09 Mar '18

Hi, After having installed `pyadjoint` by running `firedrake-update --install pyadjoint` I am getting `AttributeError: 'CoordinatelessFunction' object has no attribute 'block_variable'` on all calls to `solve` or `assemble` e.g.: ``` solve(F == 0, self.I, bcs=bcs) File "/home/bock/bin/firedrake/src/pyadjoint/fenics_adjoint/solving.py", line 28, in solve block = SolveBlock(*args, **kwargs) File "/home/bock/bin/firedrake/src/pyadjoint/fenics_adjoint/solving.py", line 112, in __init__ self.add_dependency(c.block_variable) File "/home/bock/bin/firedrake/src/firedrake/firedrake/function.py", line 275, in __getattr__ return getattr(self._data, name) AttributeError: 'CoordinatelessFunction' object has no attribute 'block_variable' ``` >From the error above it seems that the constellation of my imports is incorrect, and that the (I believe) correct `Function` in firedrake_adjoint/types/function.py is not used. I have been playing around with different combinations of `from {firedrake_adjoint, pyadjoint} import *` without success. Can you please point me in the direction of documentation that can help me modify my code to cater for the new change? `firedrake-status`: ``` --------------------------------------------------------------------------- |Package |Branch |Revision |Modified | --------------------------------------------------------------------------- |COFFEE |master |5bb1b30 |False | |FInAT |master |ef12759 |False | |PyOP2 |master |8a918cfc |False | |fiat |master |4a201c7 |False | |firedrake |master |3cf27a5f |False | |h5py |firedrake |c1e4a81 |False | |libspatialindex |master |4768bf3 |True | |petsc |firedrake |3c589c124b|False | |petsc4py |firedrake |3f09f6d |False | |pyadjoint |master |89b15bd |False | |tsfc |master |e6882a1 |False | |ufl |master |bd0376da |False | --------------------------------------------------------------------------- ``` <mailto:firedrake%40imperial.ac.uk> Best regards, Andreas Bock

2 2