Changing PETSc parameters between solves causes segfault

[Former user]

I have a code like this

dolfin_set("Krylov monitor convergence",true); //default is false

// lots of supposedly unrelated code

pde1 = LinearPDE(...);
pde2 = LinearPDE(...);

// lots of supposedly unrelated code

while (...)
    for(a few times)
        pde1->solve()

    pde2->solve()

which works just fine. However, if I change it to

// lots of supposedly unrelated code

pde1 = LinearPDE(...);
pde2 = LinearPDE(...);

// lots of supposedly unrelated code

while (...)
    for(a few times)
        pde1->solve()

    dolfin_set("Krylov monitor convergence",true); //default is false
    pde2->solve()
    dolfin_set("Krylov monitor convergence",false); //default is false

then I get a segmentation fault message from PETSc that comes from one of these two lines

    dolfin_set("Krylov monitor convergence",true); //default is false
    pde2.solve()

Is there anything special that I should do to make this work?

[Niclas Jansson]

For which version of PETSc?

[Former user]

PETSc version 3.7.3 from the FEniCS Ubuntu PPA.

[Former user]

Sorry, please ignore my previous reply, I was looking at the wrong thing :D This happens on beskow with their petsc libraries. I cannot check the exact version now because beskow is down for maintenance.

[Former user]

Ok, it happens using module cray-petsc/3.6.1.0 on Beskow.

[Niclas Jansson]

Ok, but can you reproduce the problem on another platform? If not, you have to check that you have the correct version of all the aux. modules loaded on the Cray.

[Former user]

I can try to do that. Meanwhile, this is the list of modules I'm using

#!

Currently Loaded Modulefiles:
  1) modules/3.2.6.7                                                          12) Base-opts/1.0.2-1.0502.57047.1.1.ari                                     23) pdc-utils/0.0
  2) nodestat/2.2-1.0502.58998.2.7.ari                                        13) craype-network-aries                                                     24) cray-tpsl/1.5.2
  3) sdb/1.0-1.0502.58450.3.27.ari                                            14) craype/2.2.1                                                             25) boost/1.63.0-gcc-4.9.1
  4) alps/5.2.3-2.0502.9295.14.14.ari                                         15) craype-haswell                                                           26) cmake/3.7.1
  5) lustre-cray_ari_s/2.5_3.0.101_0.35.1_1.0502.8640.15.1-1.0502.18911.12.4  16) cray-mpich/7.0.4                                                         27) cray-petsc/3.6.1.0
  6) udreg/2.3.2-1.0502.9889.2.20.ari                                         17) slurm/slurm-14.11.9                                                      28) intel/14.0.4.211
  7) ugni/6.0-1.0502.10245.9.9.ari                                            18) openssh/5.3p1-with-pam-gsskex-20100124                                   29) cray-libsci/13.0.1
  8) gni-headers/4.0-1.0502.10317.9.2.ari                                     19) openafs/1.6.17-3.0.101-0.35.1_1.0502.8640-cray_ari_s                     30) pmi/5.0.5-1.0000.10300.134.8.ari
  9) dmapp/7.0.1-1.0502.10246.8.47.ari                                        20) heimdal/1.5.2                                                            31) atp/1.7.5
 10) xpmem/0.1-2.0502.57015.1.15.ari                                          21) snic-env/1.0.0                                                           32) PrgEnv-intel/5.2.56
 11) hss-llm/7.2.0                                                            22) slurm-utils/0.1

which I load with

#!

module load cray-tpsl/1.5.2 boost cmake cray-petsc/3.6.1.0
module swap PrgEnv-cray PrgEnv-intel

Do you see anything wrong with this?

[Niclas Jansson]

How do you run configure? Properly or forcing autoconf to pick up PETSc?

[Former user]

This is what I have in config.log:

#!

$ ./configure --prefix=/cfs/nobackup/m/mleoni/mySoftware/fenics-hpc/install --with-pic --enable-function-cache --enable-optimize-p1 --disable-boost-tr1 --with-parmetis --with-petsc --enable-mpi --enable-mpi-io --disable-progress-bar --disable-xmltest --enable-ufl --with-gts

[Niclas Jansson]

Ok, seems to be correct. And I assume you get something like

Cray PETSc ..... yes

when you run it? Then we need to try on a different platform to rule out if this is a Cray issue or (most likely) and issue in our interface.

[Former user]

Yes, I get that in the configure output. I will try to reproduce it on another architecture!

[Former user]

I haven't been able to reproduce this bug regularly, and in several cases the code seemed to work as expected. We can mark this as "Invalid" for now and perhaps get back to it later if the problem shows up again.