University of Maryland's old HPC platform deepthought 2 (composed with old intel chips, each node has 20 cpus) was permanently shut down in September 2022, and abaqus 2022 has been migrated to a new HPC platform Zaratan (All AMD EPYC chips with 128 cpus on each node https://hpcc.umd.edu/hpcc/zaratan.html). Since then, the parallelization of abaqus explicit jobs is not as fast as it was on deepthought 2 when utilizing 2 or more nodes, sometimes even running on a single node is not working, either.
The abaqus job clearly cannot be properly parallelized and accelerated on AMD chips. Since I have tested both my job and the benchmark job e14_droptest_v0 from Dassault SIMULIA, which is also claimed that has been tested in the following two white papers released by AMD.
https://www.amd.com/system/files/documents/amd-epyc-7003-sb-hpc-simulia-abaqus-explicit.pdf
https://www.amd.com/system/files/documents/amd-epyc-7003-sb-simulia-abaqus.pdf
Could this issue be caused by the wrong system configuration?
What is the necessary tuning to tune the AMD chips work efficiently to parallelize abaqus jobs?
Quick additions added here:
Found someone encountered similar problems before when running abaqus jobs on AMD EPYC chips. And they were lucky enough to have their consultant company solve their problem. But how? Why not AMD provide me with an official answer?
http://www.totalcae.com/learn/tuning-amd-epyc-systems-for-cae-applications/
