[aspect-devel] "Floating point exception" error from spherical harmonic functions of C++ boost library

Shangxin Liu sxliu at vt.edu
Fri Feb 12 18:51:03 PST 2016


Hi;

I think previously I've run into this kind of FPE problem but haven't make
a detailed test in DEBUG mode. Recently, from my tests of several spherical
shell convections in DEBUG mode, I'd like to report this error:

When I ran the 3D spherical shell convection from cookbooks, the job is
forwarding normally step by step. However, when I change the initial
condition of this prm file to harmonic perturbation using a single
degree-order spherical harmonic functions from boost library, the job
always gets the following "Floating point exception" error regarding the
boost math:

[br001:20888] *** Process received signal ***

[br001:20888] Signal: Floating point exception (8)

[br001:20888] Signal code:  (7)

[br001:20888] Failing at address: 0x1151d26

[br001:20888] [ 0] /lib64/libpthread.so.0(+0xf500) [0x7fdc123e1500]

[br001:20888] [ 1]
/home/shangxin/aspect_public/aspect/build/aspect(_ZN5boost4math18tgamma_delta_ratioIeeNS0_8policies6policyINS2_14default_policyES4_S4_S4_S4_S4_S4_S4_S4_S4_S4_S4_S4_EEEENS0_5tools12promote_argsIT_T0_ffffE4typeES8_S9_RKT1_+0x85)
[0x1151d26]

[br001:20888] [ 2]
/home/shangxin/aspect_public/aspect/build/aspect(_ZN5boost4math6detail25spherical_harmonic_prefixIeNS0_8policies6policyINS3_14default_policyES5_S5_S5_S5_S5_S5_S5_S5_S5_S5_S5_S5_EEEET_jjS7_RKT0_+0x11f)
[0x1151705]

[br001:20888] [ 3]
/home/shangxin/aspect_public/aspect/build/aspect(_ZN5boost4math6detail20spherical_harmonic_rIeNS0_8policies6policyINS3_14default_policyES5_S5_S5_S5_S5_S5_S5_S5_S5_S5_S5_S5_EEEET_jiS7_S7_RKT0_+0xb8)
[0x115138d]

[br001:20888] [ 4]
/home/shangxin/aspect_public/aspect/build/aspect(_ZN5boost4math20spherical_harmonic_rIddNS0_8policies6policyINS2_14default_policyES4_S4_S4_S4_S4_S4_S4_S4_S4_S4_S4_S4_EEEENS0_5tools12promote_argsIT_T0_ffffE4typeEjiS8_S9_RKT1_+0x3e)
[0x1150fb5]

[br001:20888] [ 5]
/home/shangxin/aspect_public/aspect/build/aspect(_ZN5boost4math20spherical_harmonic_rIddEENS0_5tools12promote_argsIT_T0_ffffE4typeEjiS4_S5_+0x41)
[0x1150994]

[br001:20888] [ 6]
/home/shangxin/aspect_public/aspect/build/aspect(_ZNK6aspect17InitialConditions20HarmonicPerturbationILi3EE19initial_temperatureERKN6dealii5PointILi3EdEE+0x3b5)
[0x11638f5]

[br001:20888] [ 7]
/home/shangxin/aspect_public/aspect/build/aspect(_ZN6aspect9SimulatorILi3EE48set_initial_temperature_and_compositional_fieldsEv+0x355)
[0x12ad7f3]

[br001:20888] [ 8]
/home/shangxin/aspect_public/aspect/build/aspect(_ZN6aspect9SimulatorILi3EE3runEv+0x392)
[0x13019e4]

[br001:20888] [ 9]
/home/shangxin/aspect_public/aspect/build/aspect(main+0x606) [0x10b40b3]

[br001:20888] [10] /lib64/libc.so.6(__libc_start_main+0xfd) [0x7fdc1205ecdd]

[br001:20888] [11] /home/shangxin/aspect_public/aspect/build/aspect()
[0x10b2329]

[br001:20888] *** End of error message ***

-----------------------------------------------------------------------

mpirun noticed that process rank 0 with PID 20888 on node br001 exited on
signal 8 (Floating point exception).

-----------------------------------------------------------------------

Here I used the public ASPECT version from git hub and didn't modify
anything. From my experience, in RELEASE mode, this FPE error doesn't
matter when I ran the code for only one time step but will cause the
time-dependent job to hang up at some time step. However, If I switch off
the "ASPECT_USE_FP_EXCEPTIONS" when compile ASPECT, this problem doesn't
appear any more.

I further checked the harmonic_perturbation.cc code carefully and didn't
find any operations like division by zero. Currently I'm using the boost
version 1.49.0. I don't know whether this version has a bug in its
spherical harmonic functions.

I wonder whether anyone else here working on spherical harmonics ran into
this error before (Rene and Jacky?). And which boost version are you using?

Best,
Shangxin
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.geodynamics.org/pipermail/aspect-devel/attachments/20160212/504cf2f7/attachment.html>


More information about the Aspect-devel mailing list