Abaqus 2019 Bug: Catastrophic performance degradation of solver with user subroutines

Hello!

I've discovered a bug in Abaqus 2019.HF6 (fix pack 1947) which results in catastrophic performance degradation of solver with user subroutines in Abaqus 2019 comparing to Abaqus 2016. All what you need is to create a model that uses a user subroutine which allocates and deallocates some memory and makes some non-trivial calculations. Below I give two examples for HETVAL and UMAT, in which performance degradation factors are 37 and 59 respectively.


EXAMPLE OF HEAT-TRANSFER PROBLEM WITH USER SUBROUTINE HETVAL

Steps to reproduce the bug:

1. Rename model_hetval.txt to model_hetval.py (sorry, it's not possible to attach a *.py file here).
2. Rename hetval_bug_demonstration.txt to hetval_bug_demonstration.for (sorry, it's not possible to attach a *.for file here).
3. Run the python script model_hetval.py to create the model.
4. Locate user subroutine file hetval_bug_demonstration.for in job properties and run Job-1 on 4 CPUs.
5. Check job execution time in job monitor after the job completes (i.e. wallclock time).
6. Perform the above steps for Abaqus 2016.HF7 and Abaqus 2019.HF6 and compare job execution time.

In my case the job runs about 95 seconds in Abaqus 2016 and about 3565 seconds in Abaqus 2019. Thus, performance degradation factor in this case is about 37.


EXAMPLE OF STATIC PROBLEM WITH USER SUBROUTINE UMAT

Steps to reproduce the bug:

1. Rename model_umat.txt to model_umat.py (sorry, it's not possible to attach a *.py file here).
2. Rename umat_bug_demonstration.txt to umat_bug_demonstration.for (sorry, it's not possible to attach a *.for file here).
3. Run the python script model_umat.py to create the model.
4. Locate user subroutine file umat_bug_demonstration.for in job properties and run Job-1 on 4 CPUs.
5. Check job execution time in job monitor after the job completes (i.e. wallclock time).
6. Perform the above steps for Abaqus 2016.HF7 and Abaqus 2019.HF6 and compare job execution time.

In my case the job runs about 48 seconds in Abaqus 2016 and about 2875 seconds in Abaqus 2019. Thus, performance degradation factor in this case is about 59.


SYSTEM INFORMATION

The bug has been found in Abaqus 2019.HF6 (fix pack 1947).
Abaqus version without this bug: Abaqus 2016.HF7 (fix pack 1646).
OS version: Windows 10 Pro 1803, 64-bit.
FORTRAN Compiler: Intel Fortran 16.0 Update 1
C++ Compiler/Linker: Visual Studio 2015 Update 3 for Abaqus 2019 and Visual Studio 2012 for Abaqus 2016.


I really hope this will be fixed in some fix pack of the 2019 release...

Best regards,
Mikhail Kozlov