[petsc-users] Floating point exception

Matthew Knepley knepley at gmail.com
Mon Apr 27 13:34:06 CDT 2015


On Tue, Apr 28, 2015 at 2:44 AM, Danyang Su <danyang.su at gmail.com> wrote:

> Hi Barry,
>
> I got the following arithemetic exception after the previous bug is fixed.
>
> Loaded symbols for /lib/x86_64-linux-gnu/libnss_files.so.2
> 0x00007f3b23b98f20 in __nanosleep_nocancel ()
>     at ../sysdeps/unix/syscall-template.S:81
> 81      ../sysdeps/unix/syscall-template.S: No such file or directory.
> (gdb) cont
> Continuing.
>
> Program received signal SIGFPE, Arithmetic exception.
> 0x00007f3b260df449 in gselim (A=0x1e4c580, x=0x1d30c40, n=9)
>     at par_relax.c:3442
> 3442               x[k] /= A[k*n+k];
> (gdb)
>
> I tried both PETSc 3.5.2 and 3.5.3 and they return the same error as shown
> above. For 3.5.3, i edited fp.c file and then configure and make.
>

It looks like there is a 0 on the diagonal of some matrix A. We cannot know
which one
without the stack trace. Can you send the whole trace?

  Matt


> Thanks,
>
> Danyang
>
> On 15-04-25 07:34 PM, Danyang Su wrote:
>
>> Hi All,
>>
>> The "floating point underflow" is caused by a small value divided by a
>> very large value. This result is forced to zero and then it does not report
>> any underflow problem. I just rerun this bad case to see if it still get
>> stuck later. This will take a while.
>>
>> Thanks for all your kindly reply,
>>
>> Danyang
>>
>> On 15-04-25 07:02 PM, Barry Smith wrote:
>>
>>>    Ok, you do have
>>>
>>> #ifndef PETSC_HAVE_XMMINTRIN_H
>>> #define PETSC_HAVE_XMMINTRIN_H 1
>>> #endif
>>>
>>>    so the change you made should cause it to stop trapping underflow
>>> exceptions.
>>>
>>>    Now in one email you reported a FPE within hypre, then I asked you to
>>> run with -start_in_debugger to determine where it happened exactly and then
>>> you reported the FPE happened in  user code (what seemed to be an underflow
>>> issue). Why is this? Can you not run it where it generated the FPE in hypre
>>> using the -start_in_debugger option?
>>>
>>>     Barry
>>>
>>> Perhaps you have multiple PETSC_ARCH or multiple PETSc installs to
>>> explain why you reported two different places where the exception occurred.
>>>
>>>  On Apr 25, 2015, at 8:31 PM, Danyang Su <danyang.su at gmail.com> wrote:
>>>>
>>>>
>>>>
>>>> On 15-04-25 06:26 PM, Matthew Knepley wrote:
>>>>
>>>>> On Sat, Apr 25, 2015 at 8:23 PM, Danyang Su <danyang.su at gmail.com>
>>>>> wrote:
>>>>>
>>>>>
>>>>> On 15-04-25 06:03 PM, Barry Smith wrote:
>>>>>     If this is what you got in your last run
>>>>>
>>>>>       at ../../gas_advection/velocity_g.F90:1344
>>>>> 1344                                        cinfrt = cinfrt_dg(i1) *
>>>>> diff(ic,idim)                                      !diff is a very small
>>>>> value, e.g., 1.0d-316
>>>>>     then it is still catching floating point underflow, which we do
>>>>> not want. This means either the change I suggested you make in the fp.c
>>>>> code didn't work or it actually uses a different floating point trap than
>>>>> that one.  BTW: absurd numbers like 1.0d-316 are often a symptom of
>>>>> uninitialized data; could that be a problem that diff is not filled
>>>>> correctly for all the ic, idim you are using?
>>>>>
>>>>>      This going round and round is very frustrating and a waste of
>>>>> time. You need to be more proactive yourself and explore the code and poke
>>>>> around to figure out how to solve the problem.
>>>>>
>>>>>     Please email $PETSC_DIR/$PETSC_ARCH/include/petscvariables.h so I
>>>>> can see what FP trap is being used on your machine.
>>>>>
>>>>> Barry
>>>>> Do you mean $PETSC_DIR/$PETSC_ARCH/conf/petscvariables? Otherwise I
>>>>> cannot find this file.
>>>>>
>>>>> Its include/petscconf.h
>>>>>   Do I need to reconfigure PETSc after changing the code you mentioned?
>>>>>
>>>>> No, but you need to rebuild.
>>>>>
>>>> Yes, I have done 'gnumake'.
>>>>
>>>>>     Matt
>>>>>   Danyang
>>>>>
>>>>>
>>>>>
>>>>> On Apr 25, 2015, at 2:24 PM, Danyang Su <danyang.su at gmail.com> wrote:
>>>>>
>>>>>
>>>>>
>>>>> On 15-04-25 11:55 AM, Barry Smith wrote:
>>>>> On Apr 25, 2015, at 1:51 PM, Danyang Su <danyang.su at gmail.com>
>>>>>    wrote:
>>>>>
>>>>>
>>>>>
>>>>> On 15-04-25 11:32 AM, Barry Smith wrote:
>>>>>
>>>>>     I told you this yesterday.
>>>>>
>>>>>     It is probably stopping here on a harmless underflow. You need to
>>>>> edit the PETSc code to not worry about underflow.
>>>>>
>>>>> Edit the file /home/dsu/Soft/PETSc/petsc-3.5.2/src/sys/error/fp.c and
>>>>> locate
>>>>>
>>>>> #elif defined PETSC_HAVE_XMMINTRIN_H
>>>>>      _MM_SET_EXCEPTION_MASK(_MM_MASK_INEXACT);
>>>>> #else
>>>>>
>>>>> change it to
>>>>>
>>>>> #elif defined PETSC_HAVE_XMMINTRIN_H
>>>>>      _MM_SET_EXCEPTION_MASK(_MM_MASK_INEXACT | _MM_MASK_UNDERFLOW);
>>>>> #else
>>>>>
>>>>>    Then run make gnumake in the PETSc directory to compile the new
>>>>> version. Now link and run the program again with -fp_trap and see where it
>>>>> gets stuck this time.
>>>>>
>>>>>     Did you do this?
>>>>>
>>>>>    Barry
>>>>>
>>>>> Yes, I did change the code in fp.c and run 'make gnumake' in the PETSc
>>>>> directory. I just did a double check and ran make gnumake again and got the
>>>>> following information this time.
>>>>>
>>>>>
>>>>> dsu at nwmop:~/Soft/PETSc/petsc-3.5.2$
>>>>>    make gnumake
>>>>> Building PETSc using GNU Make with 10 build threads
>>>>> ==========================================
>>>>> make[1]: Entering directory `/home/dsu/Soft/PETSc/petsc-3.5.2'
>>>>> make[1]: Nothing to be done for `all'.
>>>>> make[1]: Leaving directory `/home/dsu/Soft/PETSc/petsc-3.5.2'
>>>>> =========================================
>>>>>
>>>>>
>>>>> Then I recompiled the codes, ran with -fp_trap and still got the
>>>>> following error
>>>>>
>>>>> Backtrace for this error:
>>>>> Note: The EXACT line numbers in the stack are not available,
>>>>> [2]PETSC ERROR:       INSTEAD the line number of the start of the
>>>>> function
>>>>> [2]PETSC ERROR:       is given.
>>>>> [2]PETSC ERROR: [2] PetscDefaultFPTrap line 379
>>>>> /home/dsu/Soft/PETSc/petsc-3.5.2/src/sys/error/fp.c
>>>>>        INSTEAD the line number of the start of the function
>>>>> [3]PETSC ERROR:       is given.
>>>>> [3]PETSC ERROR: [3] PetscDefaultFPTrap line 379
>>>>> /home/dsu/Soft/PETSc/petsc-3.5.2/src/sys/error/fp.c
>>>>> [2]PETSC ERROR: User provided function() line 0 in Unknown file
>>>>> trapped floating point error
>>>>> [3]PETSC ERROR: User provided function() line 0 in Unknown file
>>>>> trapped floating point error
>>>>>
>>>>>
>>>>>      This is different then what you sent a few minutes ago where it
>>>>> crashed in hypre.
>>>>>
>>>>>      Anyways you need to use the -start_in_debugger business I sent in
>>>>> the previous email to see the exact place the problem occurs.
>>>>>
>>>>> Here is the information shown on gdb screen
>>>>>
>>>>> Program received signal SIGFPE, Arithmetic exception.
>>>>> 0x00000000006c2bef in velocity_g (l_sufx=1, suffix=..., nmax=12,
>>>>> njamxc=34,
>>>>>       cinfradx=..., radial_coordx=.FALSE., _suffix=3)
>>>>>       at ../../gas_advection/velocity_g.F90:1344
>>>>> 1344                                        cinfrt = cinfrt_dg(i1) *
>>>>> diff(ic,idim)                                      !diff is a very small
>>>>> value, e.g., 1.0d-316
>>>>> (gdb)
>>>>>
>>>>> After type cont on gdb screen, I got error information as below
>>>>>
>>>>> [1]PETSC ERROR: *** unknown floating point error occurred ***
>>>>> [1]PETSC ERROR: The specific exception can be determined by running in
>>>>> a debugger.  When the
>>>>> [1]PETSC ERROR: debugger traps the signal, the exception can be found
>>>>> with fetestexcept(0x3d)
>>>>> [1]PETSC ERROR: where the result is a bitwise OR of the following
>>>>> flags:
>>>>> [1]PETSC ERROR: FE_INVALID=0x1 FE_DIVBYZERO=0x4 FE_OVERFLOW=0x8
>>>>> FE_UNDERFLOW=0x10 FE_INEXACT=0x20
>>>>> [1]PETSC ERROR: Try option -start_in_debugger
>>>>> [1]PETSC ERROR: likely location of problem given in stack below
>>>>> [1]PETSC ERROR: ---------------------  Stack Frames
>>>>> ------------------------------------
>>>>> [1]PETSC ERROR: Note: The EXACT line numbers in the stack are not
>>>>> available,
>>>>> [1]PETSC ERROR:       INSTEAD the line number of the start of the
>>>>> function
>>>>> [1]PETSC ERROR:       is given.
>>>>> [1]PETSC ERROR: [1] PetscDefaultFPTrap line 379
>>>>> /home/dsu/Soft/PETSc/petsc-3.5.2/src/sys/error/fp.c
>>>>> [1]PETSC ERROR: User provided function() line 0 in Unknown file
>>>>> trapped floating point error
>>>>> [0]PETSC ERROR: *** unknown floating point error occurred ***
>>>>> [0]PETSC ERROR: The specific exception can be determined by running in
>>>>> a debugger.  When the
>>>>> [0]PETSC ERROR: debugger traps the signal, the exception can be found
>>>>> with fetestexcept(0x3d)
>>>>> [0]PETSC ERROR: where the result is a bitwise OR of the following
>>>>> flags:
>>>>> [0]PETSC ERROR: FE_INVALID=0x1 FE_DIVBYZERO=0x4 FE_OVERFLOW=0x8
>>>>> FE_UNDERFLOW=0x10 FE_INEXACT=0x20
>>>>> [0]PETSC ERROR: Try option -start_in_debugger
>>>>> [0]PETSC ERROR: likely location of problem given in stack below
>>>>>
>>>>> Thanks,
>>>>>
>>>>> Danyang
>>>>> Thanks,
>>>>>
>>>>> Danyang
>>>>>
>>>>> On Apr 25, 2015, at 1:05 AM, Danyang Su <danyang.su at gmail.com>
>>>>>    wrote:
>>>>>
>>>>> Hi Barry and Satish,
>>>>>
>>>>> How can I get rid of unknown floating point error when a very small
>>>>> value is multiplied.
>>>>>
>>>>> e.g.,
>>>>> cinfrt_dg(i1) and diff(ic,idim) are  1.0250235986806329E-008
>>>>> 8.6178408169776945E-317 respectively,
>>>>>
>>>>> cinfrt = cinfrt_dg(i1) * diff(ic,idim)
>>>>>
>>>>> I get the following error when run with "-fp_trap -start_in_debugger".
>>>>>
>>>>> Backtrace for this error:
>>>>> *** unknown floating point error occurred ***
>>>>> [2]PETSC ERROR: The specific exception can be determined by running in
>>>>> a debugger.  When the
>>>>> [2]PETSC ERROR: debugger traps the signal, the exception can be found
>>>>> with fetestexcept(0x3d)
>>>>> [2]PETSC ERROR:  cinfrt_dg(i1),diff(ic,idim) 1.0250235986806329E-008
>>>>>  8.6178408169776945E-317
>>>>> where the result is a bitwise OR of the following flags:
>>>>> [2]PETSC ERROR: FE_INVALID=0x1 FE_DIVBYZERO=0x4 FE_OVERFLOW=0x8
>>>>> FE_UNDERFLOW=0x10 FE_INEXACT=0x20
>>>>> [2]PETSC ERROR: Try option -start_in_debugger
>>>>> [2]PETSC ERROR: likely location of problem given in stack below
>>>>>
>>>>> Thanks,
>>>>>
>>>>> Danyang
>>>>>
>>>>> On 15-04-24 01:54 PM, Danyang Su wrote:
>>>>>
>>>>> On 15-04-24 01:23 PM, Satish Balay wrote:
>>>>>
>>>>>    c           4   1.0976214263087059E-067
>>>>>
>>>>> I don't think this number can be stored in a real*4.
>>>>>
>>>>> Satish
>>>>>
>>>>> Thanks, Satish. It is caused by this number.
>>>>>
>>>>> On Fri, 24 Apr 2015, Danyang Su wrote:
>>>>>
>>>>>
>>>>> On 15-04-24 11:12 AM, Barry Smith wrote:
>>>>>
>>>>> On Apr 24, 2015, at 1:05 PM, Danyang Su <danyang.su at gmail.com>
>>>>>    wrote:
>>>>>
>>>>> Hi All,
>>>>>
>>>>> One of my case crashes because of floating point exception when using 4
>>>>> processors, as shown below. But if I run this case with 1 processor, it
>>>>> works fine. I have tested the codes with around 100 cases up to 768
>>>>> processors, all other cases work fine. I just wonder if this kind of
>>>>> error
>>>>> is caused because of NaN in jacobi matrix, RHS or preconditioner?
>>>>>
>>>>>       Yes, almost for sure it is one of these places.
>>>>>
>>>>>       First run the bad case with -fp_trap  if all goes well you'll
>>>>> see the
>>>>> function where the FPE is generated. Then run also with
>>>>> -start_in_debugger
>>>>> and
>>>>> type cont in all four debugger windows. When the FPE happens the
>>>>> debugger
>>>>> should stop showing exactly where the FPE happens.
>>>>>
>>>>>      Barry
>>>>>
>>>>> Hi Barry,
>>>>>
>>>>> If run with -fp_trap -start_in_debugger, I got the following error
>>>>>
>>>>> [0]PETSC ERROR: *** unknown floating point error occurred ***
>>>>> [0]PETSC ERROR: The specific exception can be determined by running in
>>>>> a
>>>>> debugger.  When the
>>>>> [0]PETSC ERROR: debugger traps the signal, the exception can be found
>>>>> with
>>>>> fetestexcept(0x3d)
>>>>> [0]PETSC ERROR: where the result is a bitwise OR of the following
>>>>> flags:
>>>>> [0]PETSC ERROR: FE_INVALID=0x1 FE_DIVBYZERO=0x4 FE_OVERFLOW=0x8
>>>>> FE_UNDERFLOW=0x10 FE_INEXACT=0x20
>>>>> [0]PETSC ERROR: Try option -start_in_debugger
>>>>> [0]PETSC ERROR: likely location of problem given in stack below
>>>>> [0]PETSC ERROR: ---------------------  Stack Frames
>>>>> ------------------------------------
>>>>> [0]PETSC ERROR: Note: The EXACT line numbers in the stack are not
>>>>> available,
>>>>> [0]PETSC ERROR:       INSTEAD the line number of the start of the
>>>>> function
>>>>> [0]PETSC ERROR:       is given.
>>>>> [0]PETSC ERROR: [0] PetscDefaultFPTrap line 379
>>>>> /home/dsu/Soft/PETSc/petsc-3.5.2/src/sys/error/fp.c
>>>>> [0]PETSC ERROR: User provided function() line 0 in Unknown file trapped
>>>>> floating point error
>>>>>
>>>>> Program received signal SIGABRT: Process abort signal.
>>>>>
>>>>> Backtrace for this error:
>>>>> #0  0x7F4FEAB1C7D7
>>>>> #1  0x7F4FEAB1CDDE
>>>>> #2  0x7F4FE9E1AD3F
>>>>> #3  0x7F4FE9E1ACC9
>>>>> #4  0x7F4FE9E1E0D7
>>>>> #5  0x7F4FEB0B6DCB
>>>>> #6  0x7F4FEB0B1825
>>>>> #7  0x7F4FEB0B817F
>>>>> #8  0x7F4FE9E1AD3F
>>>>> #9  0x6972C8 in tprfrtlc_ at tprfrtlc.F90:2393 (discriminator 3)
>>>>> #10  0x4C6C87 in gcreact_ at gcreact.F90:678
>>>>> #11  0x707E19 in initicrt_ at initicrt.F90:589
>>>>> #12  0x4F42D0 in initprob_ at initprob.F90:430
>>>>> #13  0x5AAF72 in driver_pc at driver_pc.F90:438
>>>>>
>>>>> I checked the code at  tprfrtlc.F90:2393,
>>>>>
>>>>>           realbuffer_gb(1:nvars) = (/time,(c(ic),ic=1,nc-1),     &
>>>>>                                      (cx(ix),ix=1,nxout)/)
>>>>>
>>>>> All the values (time, c, cx) are reasonable, as shown below. The only
>>>>> possibility is that realbuffer_gb is in declared as real*4 if using
>>>>> sing
>>>>> precision output while time, c, cx are declared in real*8. I have a
>>>>> lot of
>>>>> similar data conversion from real*8 to real*4 output, other code does
>>>>> not
>>>>> return error.
>>>>>
>>>>>    time   0.0000000000000000
>>>>>    c           1   9.9999999999999995E-008
>>>>>    c           2   3.1555251077549618E-003
>>>>>    c           3   7.1657814842179362E-008
>>>>>    c           4   1.0976214263087059E-067
>>>>>    c           5   5.2879822292305797E-004
>>>>>    c           6   9.9999999999999964E-005
>>>>>    c           7   6.4055731968811337E-005
>>>>>    c           8   3.4607572892578404E-020
>>>>>    cx           1   3.4376650636008101E-005
>>>>>    cx           2   7.3989678854017763E-012
>>>>>    cx           3   9.5317170613607207E-012
>>>>>    cx           4   2.2344525794718353E-015
>>>>>    cx           5   3.0624685689695889E-008
>>>>>    cx           6   1.0046157902783967E-007
>>>>>    cx           7   1.5320169154914984E-004
>>>>>    cx           8   8.6930292776346176E-014
>>>>>    cx           9   3.5944267559348721E-005
>>>>>    cx          10   3.0072645866951157E-018
>>>>>    cx          11   2.3592486321095017E-013
>>>>>
>>>>> Thanks,
>>>>>
>>>>> Danyang
>>>>>
>>>>>
>>>>> I can check all the entries of jacobi matrix to see if the value is
>>>>> valid,
>>>>> but this seems not a good idea as it takes a long time to reach this
>>>>> point. If I restart the simulation from a specified time (e.g., 7.685
>>>>> in
>>>>> this case), then the error does not occur.
>>>>>
>>>>> Would you please give me any suggestion on debugging this case?
>>>>>
>>>>> Thanks and Regards,
>>>>>
>>>>> Danyang
>>>>>
>>>>>
>>>>> timestep:    2730 time: 7.665E+00 years   delt: 1.000E-02 years iter:
>>>>> 1
>>>>> timestep:    max.sia: 0.000E+00 tol.sia: 0.000E+00
>>>>> timestep:    2731 time: 7.675E+00 years   delt: 1.000E-02 years iter:
>>>>> 1
>>>>> timestep:    max.sia: 0.000E+00 tol.sia: 0.000E+00
>>>>> timestep:    2732 time: 7.685E+00 years   delt: 1.000E-02 years iter:
>>>>> 1
>>>>> timestep:    max.sia: 0.000E+00 tol.sia: 0.000E+00
>>>>> timestep:    2733 time: 7.695E+00 years   delt: 1.000E-02 years iter:
>>>>> 1
>>>>> timestep:    max.sia: 0.000E+00 tol.sia: 0.000E+00
>>>>> timestep:    2734 time: 7.705E+00 years   delt: 1.000E-02 years iter:
>>>>> 1
>>>>> timestep:    max.sia: 0.000E+00 tol.sia: 0.000E+00
>>>>> Reduce time step for reactive transport
>>>>> timestep:    2734 time: 7.700E+00 years   delt: 5.000E-03 years iter:
>>>>> 1
>>>>> timestep:    max.sia: 0.000E+00 tol.sia: 0.000E+00
>>>>> Reduce time step for reactive transport
>>>>> timestep:    2734 time: 7.697E+00 years   delt: 2.500E-03 years iter:
>>>>> 1
>>>>> timestep:    max.sia: 0.000E+00 tol.sia: 0.000E+00
>>>>> [1]PETSC ERROR: --------------------- Error Message
>>>>> --------------------------------------------------------------
>>>>> [1]PETSC ERROR: Floating point exception
>>>>> [2]PETSC ERROR: --------------------- Error Message
>>>>> --------------------------------------------------------------
>>>>> [2]PETSC ERROR: Floating point exception
>>>>> [2]PETSC ERROR: Vec entry at local location 0 is not-a-number or
>>>>> infinite
>>>>> at end of function: Parameter number 3
>>>>> [2]PETSC ERROR: See
>>>>> http://www.mcs.anl.gov/petsc/documentation/faq.html
>>>>>
>>>>> for trouble shooting.
>>>>> [2]PETSC ERROR: Petsc Release Version 3.5.2, Sep, 08, 2014
>>>>> [2]PETSC ERROR: [1]PETSC ERROR: Vec entry at local location 0 is
>>>>> not-a-number or infinite at end of function: Parameter number 3
>>>>> [1]PETSC ERROR: See
>>>>> http://www.mcs.anl.gov/petsc/documentation/faq.html
>>>>>
>>>>> for trouble shooting.
>>>>> [1]PETSC ERROR: Petsc Release Version 3.5.2, Sep, 08, 2014
>>>>> [1]PETSC ERROR: ../min3p_thcm_petsc_dbg on a linux-gnu-dbg named nwmop
>>>>> by
>>>>> dsu Thu Apr 23 15:38:52 2015
>>>>> [1]PETSC ERROR: Configure options PETSC_ARCH=linux-gnu-dbg
>>>>> --with-cc=gcc
>>>>> --with-cxx=g++ --with-fc=gfortran --download-fblaslapack
>>>>> --download-mpich
>>>>> --download-mumps --download-hypre --download-superlu_dist
>>>>> --download-metis
>>>>> --download-parmetis --download-scalapack
>>>>> [1]PETSC ERROR: #1 VecValidValues() line 34 in
>>>>> /home/dsu/Soft/PETSc/petsc-3.5.2/src/vec/vec/interface/rvector.c
>>>>> ../min3p_thcm_petsc_dbg on a linux-gnu-dbg named nwmop by dsu Thu Apr
>>>>> 23
>>>>> 15:38:52 2015
>>>>> [2]PETSC ERROR: Configure options PETSC_ARCH=linux-gnu-dbg
>>>>> --with-cc=gcc
>>>>> --with-cxx=g++ --with-fc=gfortran --download-fblaslapack
>>>>> --download-mpich
>>>>> --download-mumps --download-hypre --download-superlu_dist
>>>>> --download-metis
>>>>> --download-parmetis --download-scalapack
>>>>> [2]PETSC ERROR: #1 VecValidValues() line 34 in
>>>>> /home/dsu/Soft/PETSc/petsc-3.5.2/src/vec/vec/interface/rvector.c
>>>>> [2]PETSC ERROR: [1]PETSC ERROR: #2 PCApply() line 442 in
>>>>> /home/dsu/Soft/PETSc/petsc-3.5.2/src/ksp/pc/interface/precon.c
>>>>> [1]PETSC ERROR: #2 PCApply() line 442 in
>>>>> /home/dsu/Soft/PETSc/petsc-3.5.2/src/ksp/pc/interface/precon.c
>>>>> [2]PETSC ERROR: #3 KSP_PCApply() line 230 in
>>>>> /home/dsu/Soft/PETSc/petsc-3.5.2/include/petsc-private/kspimpl.h
>>>>> #3 KSP_PCApply() line 230 in
>>>>> /home/dsu/Soft/PETSc/petsc-3.5.2/include/petsc-private/kspimpl.h
>>>>> [1]PETSC ERROR: #4 KSPInitialResidual() line 63 in
>>>>> /home/dsu/Soft/PETSc/petsc-3.5.2/src/ksp/ksp/interface/itres.c
>>>>> [2]PETSC ERROR: #4 KSPInitialResidual() line 63 in
>>>>> /home/dsu/Soft/PETSc/petsc-3.5.2/src/ksp/ksp/interface/itres.c
>>>>> [1]PETSC ERROR: #5 KSPSolve_GMRES() line 234 in
>>>>> /home/dsu/Soft/PETSc/petsc-3.5.2/src/ksp/ksp/impls/gmres/gmres.c
>>>>> [2]PETSC ERROR: #5 KSPSolve_GMRES() line 234 in
>>>>> /home/dsu/Soft/PETSc/petsc-3.5.2/src/ksp/ksp/impls/gmres/gmres.c
>>>>> [2]PETSC ERROR: #6 KSPSolve() line 459 in
>>>>> /home/dsu/Soft/PETSc/petsc-3.5.2/src/ksp/ksp/interface/itfunc.c
>>>>> [1]PETSC ERROR: #6 KSPSolve() line 459 in
>>>>> /home/dsu/Soft/PETSc/petsc-3.5.2/src/ksp/ksp/interface/itfunc.c
>>>>> ^C[mpiexec at nwmop] Sending Ctrl-C to processes as requested
>>>>> [mpiexec at nwmop] Press Ctrl-C again to force abort
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> --
>>>>> What most experimenters take for granted before they begin their
>>>>> experiments is infinitely more interesting than any results to which their
>>>>> experiments lead.
>>>>> -- Norbert Wiener
>>>>>
>>>> <petscconf.h>
>>>>
>>>
>>
>


-- 
What most experimenters take for granted before they begin their
experiments is infinitely more interesting than any results to which their
experiments lead.
-- Norbert Wiener
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20150428/408b946b/attachment-0001.html>


More information about the petsc-users mailing list