Issue with EPS_DEFAULT Sensitivity: Why Does 1.0E-12 Cause CPASSERT Failure?

[yis...@163.com]

Dear CP2K developers,

I would like to report a potential issue regarding the keyword EPS_DEFAULT. During my tests, I found that its value affects the calculation in the following way:

• With EPS_DEFAULT = 1.0E-12, the program fails with a “CPASSERT failed” error.

• With EPS_DEFAULT = 1.0E-11 or 1.0E-13, the calculation proceeds normally into the iteration process.

I have attached the input and output files for each test case. The naming convention is as follows:

• 1.0E-12_*: corresponds to EPS_DEFAULT = 1.0E-12

• 1.0E-11_*: corresponds to EPS_DEFAULT = 1.0E-11

• 1.0E-13_*: corresponds to EPS_DEFAULT = 1.0E-13

Thank you for your attention and for all your work on CP2K. I look forward to your feedback.

[Frederick Stein]

Hi,

This could be a bug. Did you compile CP2K yourself? If yes, with which compiler?

Best,

Frederick

[yis...@163.com]

I am using a supercomputer provided by a commercial supercomputing center. The CP2K software I am using was compiled by the center's staff. Is there anything I can help with?

[Frederick Stein]

If you use the module system, can you provide the result of `module ls` (or similar command to display the currently loaded modules). It would be great to have some more information on how CP2K was built. The behavior you observed appears as a bug in the code or the compiler/libraries.

[yis...@163.com]

Enter "module ls" in the command window, and its output is shown in the following figure.

[Frederick Stein]

Ok, serife-less fonts are not a good choice here... Please try `module list` after you have loaded all modules to run CP2K.

[yis...@163.com]

Regarding this issue, I've provided feedback to the administrators of the supercomputing center. The SCC admin has now shown me this diagram—I'm not sure if it fully addresses your question.

[yis...@163.com]

compiled toolchain using `--with-cmake=system --with-gcc=install --with-openmpi=install --with-sirius=no`. The host OS is Rocky Linux 8.10. 

[Frederick Stein]

Thank you, I will try to reproduce it.

[Frederick Stein]

I can reproduce it. I will investigate it further.

[yis...@163.com]

thank you very much

[Frederick Stein]

I can reproduce it with the same number of MPI ranks but a lower number resolves it. I guess you hit a rare corner case of the implementation. I still do not understand why it happens.

[yis...@163.com]

Thank you for your outstanding work. I was able to run CP2K smoothly with EPS_DEFAULT = 1.0E-13 as well. I'm reporting this bug in the hope that CP2K will become even more user-friendly.

[Frederick Stein]

I have opened an issue on Github.

[Frederick Stein]

Hi,

I found the culprit and will merge a bugfix today. It indeed seems to be a corner case of the Fortran standard/GCC compiler implementation.

Best,

Frederick

[yis...@163.com]

Thank you