# [CIG-MC] ConMan issues report

Matthew Knepley knepley at mcs.anl.gov
Wed Aug 24 17:44:16 PDT 2016

```On Wed, Aug 24, 2016 at 6:51 PM, Lyra, Wladimir <wladimir.lyra at csun.edu>
wrote:

> Dear all,
>
> Thanks for all the help. I downloaded ConMan via Thorsten Becker
> virtualbox and checked what a variable viscosity solution should be
> yielding.
>
>
> Based on that, I determined that my code (2D Cartesian finite difference,
> iterative with multigrid) is doing the correct thing but cannot go beyond a
> small variation in viscosity (grad(eta) ~ 10) before crashing (resolution
> Nx=Nz=256).
>
What does "crashing" mean? Non-convergence?

By Multigrid do you mean multigrid on the momentum operator and then some
iteration for mass conservation like Uzawa,
or a Schur complement method or preconditioner? For variable viscosity
using a Schur complement method, its important
to have a preconditioner for the Schur complement that is uniform in the
coefficient. You can get most of this by diagonal
weighting (
http://www.sciencedirect.com/science/article/pii/S003192010800191X) but you
can do even better by using a
mass matrix (http://arxiv.org/abs/1607.03936). FEM methods can have
excellent scaling even for very large viscosity
variations if you use a stable FEM discretization, as shown in
http://dl.acm.org/citation.cfm?id=2683624, which is a freely
available software package.

Matt

> Before abandoning this approach and move to using one of the several codes
> already available, I would like to understand why is it that this method is
> failing. In other words, I'm still puzzled as to how exactly were the
> software from the 90's able to solve the problem of variable viscosity. Is
> it better conservation in the finite element method as opposed to finite
> differences?
>
>
> I'm using a very high order FDM (6th order in space, 3rd in time) and I'm
> seeing results still similar to if I use low (2nd) order in space.  In
> the Blankenbach et al 1989 benchmark several codes were FD.
>
>
> Cheers,
>
>
>
> ==================================
>
> Dr Wladimir Lyra, Assistant Professor
>
> California State University, Northridge
>
> Dept of Physics and Astronomy
>
> 18111 Nordhoff St, Northridge, CA 91330
>
> Phone : +1 818 677 7464
>
>
>
>
> ------------------------------
> *From:* Shijie Zhong <shijie.zhong at colorado.edu>
> *Sent:* Thursday, August 18, 2016 10:40:20 AM
>
> *To:* cig-mc at geodynamics.org
> *Subject:* Re: [CIG-MC] ConMan issues report
>
>
> Just to add a comment on 2D direct vs iterative solvers. What Scott said
> is certainly true for most cases, especially for 1 time step Stokes flow
> solve. However, for time-dependent convection problems with time-dependent
> viscosity (due to temperature), iterative solvers may run faster than
> direct solvers due to the use of previous solutions as the initial guessed
> solution, especially if one has to compute for tens of thousands of time
> steps.
>
> Hope you all have had a nice summer.
>
>
> Shijie
>
>
>
> Shijie Zhong, Professor
> Department of Physics
> University of Colorado at Boulder
> Boulder, CO 80309
> Tel: 303-735-5095; Fax: 303-492-7935
>
>
> From: CIG-MC on behalf of John Naliboff
> Date: Thursday, August 18, 2016 at 11:32 AM
> To: "cig-mc at geodynamics.org"
> Subject: Re: [CIG-MC] ConMan issues report
>
> Hi Scott, Hi Thorsten,
>
> *Speed of ConMan vs ASPECT *
>
> Similar findings to Scott.  For 2-D models of lithospheric deformation, I
> have found that ASPECT is indeed slower than the code I used previously
> (Fortran, direct solver).  This of course changes in 3D.  Perhaps it would
> be worth having a direct solver option (MUMPS) in ASPECT for 2D problems?
>
> *ConMan development*
>
> Scott, thank you for the update.  If the community is interested, the new
> version of ConMan certainly seems worth making available rather than fixing
> up the old version.  Completely understand about the time issue regards to
> writing updated documentation.  I will speak to Louise and Lorraine and see
> what they think.  At minimum, ConMan is an excellent tool for teaching and
> benchmarking 2-D convection problems.
>
> *ASPECT installation*
>
> Yes, it can definitely eat up a good deal of time.   However, for newer
> versions of Mac OS X (Yosemite, El Capitan) a  deal.II binary is now
> available.  Basically, the binary installs all of ASPECT’s dependencies and
> then all one has to do is install ASPECT through the terminal launched by
> the deal.ii application.  Quite straightforward.
>
> There are also a few new methods to streamline the deal.II installation on
> Linux (https://github.com/dealii/candi
> <https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_dealii_candi&d=DQMGaQ&c=Oo8bPJf7k7r_cPTz1JF7vEiFxvFRfQtp-j14fFwh71U&r=betAI_WKi5c6efJqRXXEeDH07uoTxe7Vp010C6-tIlg&m=7gDE9rf695IujYwNjjKMGT_i1usQL4cvGw0qTIBpMtA&s=MmU7xIE6J052duXWliqORBISB7m71wVHN3NXeVYt-2o&e=>,
> deal.II on home-brew).  I’m still installing piece by piece, but others in
> the group here at Davis have used “candi” with success.
>
> Cheers,
> John
>
> *************************************************
> Assistant Project Scientist, CIG
> Earth & Planetary Sciences Dept., UC Davis
>
>
> On Aug 17, 2016, at 4:36 PM, Scott King <sking07 at vt.edu> wrote:
>
> Circa my geodynamics class of two years ago conman was faster than 2D
> citcom or Aspect.   For 2D problems iterative solvers don't have a large
> enough number of degrees of freedom to overcome the number of iterations
> needed to converge.   3D of course is a different story.
>
> Scott
>
> Sent from my iPhone
>
> On Aug 17, 2016, at 7:12 PM, Thorsten Becker <twb at ig.utexas.edu> wrote:
>
> I'd be curious how the speed of Aspect in 2D benchmark tests compares with
> ConMan for high Ra cases. From Aspect cookbook cases my student ran a while
> back it seemed the Aspect was quite resource demanding for high Ra at
> comparable resolved structure, though there might have been some user
> error.
>
> Thorsten Becker - Jackson School of Geosciences, UT Austin
> <https://urldefense.proofpoint.com/v2/url?u=http-3A__www-2Dudc.ig.utexas.edu_external_becker_&d=DQMGaQ&c=Oo8bPJf7k7r_cPTz1JF7vEiFxvFRfQtp-j14fFwh71U&r=betAI_WKi5c6efJqRXXEeDH07uoTxe7Vp010C6-tIlg&m=7gDE9rf695IujYwNjjKMGT_i1usQL4cvGw0qTIBpMtA&s=Qq8DaVI_Y4CirsQwROSnrDPgBY2xn3zVnfpiZzVZyhI&e=>
>
> On Wed, Aug 17, 2016 at 5:44 PM, John Naliboff <jbnaliboff at ucdavis.edu>
> wrote:
>
>> Dear Lyra,
>>
>> Yes, we're running on a mac. I can try running on a linux if it's
>> simpler, but what is the testing/debugging that is required?
>>
>>
>> It is just that we have successfully compiled ConMan on linux, but not on
>> a mac recently.  The mac installation may work right away or additional
>> issues may arise.  I would try compiling on your mac first with the 64 bit
>> compiler, but then move to linux if that does not work.
>>
>> Hm, interesting. We are in fact building our own code for specific
>> problems in icy moons, and want to benchmark the code. We can get
>> the isoviscous Blankenbach tests right, but we're running into some
>> problems for variable viscosity (cases 2 and 3). We want to check the
>> output of a code that does these cases right to see where we're doing
>> wrong.
>>
>>
>> I know the isoviscous case has been tested successfully and I believe
>> cases 2 and 3 have as well, but I will need to check on this.
>>
>> Either ConMan or ASPECT will of course work for testing the Blankenbach
>> cases.  ASPECT will definitely compile on the mac, though.
>>
>> I should be able to get to the ConMan fix sometime tomorrow and will send
>> an email as soon as it is updated.
>>
>> Cheers,
>> John
>>
>> *************************************************
>> Assistant Project Scientist, CIG
>> Earth & Planetary Sciences Dept., UC Davis
>>
>>
>> On Aug 17, 2016, at 3:35 PM, Lyra, Wladimir <wladimir.lyra at csun.edu>
>> wrote:
>>
>> Dear John
>>
>> *If you are running on a mac, more testing/debugging may be required.*
>>
>> Yes, we're running on a mac. I can try running on a linux if it's
>> simpler, but what is the testing/debugging that is required?
>>
>> *A bit off topic, but is there a particular reason you choose ConMan?
>> ASPECT might be a better choice if you would like to run in parallel, do 3D
>> simulations, use adaptive mesh refinement, etc. There is nothing you can do
>> ConMan that cannot be  easily duplicated in ASPECT.  Happy to help you get
>> setup with ASPECT if you would like to try it out.*
>>
>> Hm, interesting. We are in fact building our own code for specific
>> problems in icy moons, and want to benchmark the code. We can get
>> the isoviscous Blankenbach tests right, but we're running into some
>> problems for variable viscosity (cases 2 and 3). We want to check the
>> output of a code that does these cases right to see where we're doing
>> wrong.
>>
>> Did you benchmark ASPECT against the Blankenbach problems as well?
>>
>> Cheers,
>>
>> ==================================
>> Dr Wladimir Lyra, Assistant Professor
>> California State University, Northridge
>> Dept of Physics and Astronomy
>> 18111 Nordhoff St, Northridge, CA 91330
>> Phone : +1 818 677 7464
>>
>>
>> ------------------------------
>> *From:* John Naliboff <jbnaliboff at ucdavis.edu>
>> *Sent:* Wednesday, August 17, 2016 2:38:18 PM
>> *To:* cig-mc at geodynamics.org
>> *Subject:* Re: [CIG-MC] ConMan issues report
>>
>> Dear Leondardo,
>>
>> There is indeed an issue with the GitHub version of ConMan, thank you for
>> the reminder ;)
>>
>> We will push a patch to github in the next few days and let you know when
>>
>> On your end, what operating system are you trying to install ConMan on?
>> The updated version of ConMan should compile and run successfully on linux
>> if you use the gfortran64 compiler and associated Makefile.
>>
>> The code does compile successfully with the 32-bit compiler, but will
>> crash upon execution.  This seems consistent with your error?
>>
>> If you are running on a mac, more testing/debugging may be required.
>>
>> A bit off topic, but is there a particular reason you choose ConMan?
>> ASPECT might be a better choice if you would like to run in parallel, do 3D
>> simulations, use adaptive mesh refinement, etc.
>>
>> There is nothing you can do ConMan that cannot be  easily duplicated in
>> ASPECT.  Happy to help you get setup with ASPECT if you would like to try
>> it out.
>>
>> Anyhow, we will get ConMan updated in the next day or two.
>>
>> Cheers,
>> John
>>
>> *************************************************
>> Assistant Project Scientist, CIG
>> Earth & Planetary Sciences Dept., UC Davis
>>
>>
>> On Aug 16, 2016, at 4:34 PM, Leonardo S. Cassará <
>> leonardo10 at astro.ufrj.br> wrote:
>>
>> Dear ConMan developers and maintainers,
>>
>>
>> I am writing on mine and Professor Wladimir Lyra’s behalf, currently
>> researching on icy convection.
>>
>> conman.git and via the direct tar linkhttps://geodynamics.org/ci
>> g/software/conman/ConMan-2.0.0.tar.gz
>> <https://urldefense.proofpoint.com/v2/url?u=https-3A__geodynamics.org_cig_software_conman_ConMan-2D2.0.0.tar.gz&d=DQMFaQ&c=Oo8bPJf7k7r_cPTz1JF7vEiFxvFRfQtp-j14fFwh71U&r=betAI_WKi5c6efJqRXXEeDH07uoTxe7Vp010C6-tIlg&m=-SjoKD3I2cgpC0-lR2OFWlbaIaO4ml-VdNEZ7nzJf4E&s=x9fwriz_-ajoIC28JmsCNTLf3nT679rikpXcvR-YamI&e=>.
>> However, both ways gave us issues that I would like to address here:
>>
>> 1- When installing via the git repository, the src directory lacks the
>> libs folder, which we had to take from the installation via tar in order to
>> compile.
>>
>> 2- After compiling (using the git repository installation + libs folder
>> from tar), we could not run the code by following the instructions. Here is
>> the output of our trial:
>>
>> :~/ConMan/conman/cookbook1\$ ../src/conman.pic < run.bb1a50
>> Segmentation fault (core dumped)
>>
>> We are either missing something or there is an installation problem that
>> we alone cannot solve. I hope this report helps the ConMan team to make the
>> code even better. We appreciate any attention given to this case.
>>
>>
>> Leonardo S. Cassara
>>
>> _______________________________________________
>> CIG-MC mailing list
>> CIG-MC at geodynamics.org
>> http://lists.geodynamics.org/cgi-bin/mailman/listinfo/cig-mc
>> <https://urldefense.proofpoint.com/v2/url?u=http-3A__lists.geodynamics.org_cgi-2Dbin_mailman_listinfo_cig-2Dmc&d=DQQFaQ&c=Oo8bPJf7k7r_cPTz1JF7vEiFxvFRfQtp-j14fFwh71U&r=betAI_WKi5c6efJqRXXEeDH07uoTxe7Vp010C6-tIlg&m=-SjoKD3I2cgpC0-lR2OFWlbaIaO4ml-VdNEZ7nzJf4E&s=8j8Dyb0qeP6hSoLJ6cuNztb7mVC47UxBlIzSvYo3Pt8&e=>
>>
>>
>>
>> _______________________________________________
>> CIG-MC mailing list
>> CIG-MC at geodynamics.org
>> http://lists.geodynamics.org/cgi-bin/mailman/listinfo/cig-mc
>> <https://urldefense.proofpoint.com/v2/url?u=http-3A__lists.geodynamics.org_cgi-2Dbin_mailman_listinfo_cig-2Dmc&d=DQMGaQ&c=Oo8bPJf7k7r_cPTz1JF7vEiFxvFRfQtp-j14fFwh71U&r=betAI_WKi5c6efJqRXXEeDH07uoTxe7Vp010C6-tIlg&m=7gDE9rf695IujYwNjjKMGT_i1usQL4cvGw0qTIBpMtA&s=Dt3qsJqmThDJs02n4DyPUhMh9zkU4UgHq9sKgD4dpxY&e=>
>>
>
> _______________________________________________
> CIG-MC mailing list
> CIG-MC at geodynamics.org
> http://lists.geodynamics.org/cgi-bin/mailman/listinfo/cig-mc
> <https://urldefense.proofpoint.com/v2/url?u=http-3A__lists.geodynamics.org_cgi-2Dbin_mailman_listinfo_cig-2Dmc&d=DQMGaQ&c=Oo8bPJf7k7r_cPTz1JF7vEiFxvFRfQtp-j14fFwh71U&r=betAI_WKi5c6efJqRXXEeDH07uoTxe7Vp010C6-tIlg&m=7gDE9rf695IujYwNjjKMGT_i1usQL4cvGw0qTIBpMtA&s=Dt3qsJqmThDJs02n4DyPUhMh9zkU4UgHq9sKgD4dpxY&e=>
>
> _______________________________________________
> CIG-MC mailing list
> CIG-MC at geodynamics.org
> http://lists.geodynamics.org/cgi-bin/mailman/listinfo/cig-mc
> <https://urldefense.proofpoint.com/v2/url?u=http-3A__lists.geodynamics.org_cgi-2Dbin_mailman_listinfo_cig-2Dmc&d=DQMGaQ&c=Oo8bPJf7k7r_cPTz1JF7vEiFxvFRfQtp-j14fFwh71U&r=betAI_WKi5c6efJqRXXEeDH07uoTxe7Vp010C6-tIlg&m=7gDE9rf695IujYwNjjKMGT_i1usQL4cvGw0qTIBpMtA&s=Dt3qsJqmThDJs02n4DyPUhMh9zkU4UgHq9sKgD4dpxY&e=>
>
>
>
> _______________________________________________
> CIG-MC mailing list
> CIG-MC at geodynamics.org
> http://lists.geodynamics.org/cgi-bin/mailman/listinfo/cig-mc
>

--
What most experimenters take for granted before they begin their
experiments is infinitely more interesting than any results to which their