Dear LAMMPS users/developers:
I am testing the performance of different acceleration packages of LAMMPS for a system using the hybrid/overlay style (pair_style hybrid/overlay buck/coul/long 8.0 12.0 table linear 23001 pppm) on our Intel Xeon E5-2680v4 HPC. The LAMMPS I am using is the latest development version: 18Sep2018. The details about the test cases are listed below:
-
Case 1 (worked as expected): use the standard LAMMPS without any acceleration packages
- pair_style hybrid/overlay buck/coul/long 8.0 12.0 table linear 23001 pppm
-
Case 2 (worked as expected): use the user-omp package. That gave me a few percents performance gain
- suffix omp
- package omp 1
- pair_style hybrid/overlay buck/coul/long 8.0 12.0 table linear 23001 pppm
-
Case 3 (worked as expected): use the user-intel package, Intel pair style for buck/coul/long and standard table potential. That gave me extra %16 performance gain compared to case 1
- package intel 0 omp 1
- pair_style hybrid/overlay buck/coul/long/intel 8.0 12.0 table linear 23001 pppm
-
Case 4 (failed): use the user-intel package, Intel pair style for buck/coul/long and omp style for the table potential. As per the official documentation, table potential supports omp style.
- Package omp 1
- package intel 0 omp 1
- pair_style hybrid/overlay buck/coul/long**/intel** 8.0 12.0 table**/omp** linear 23001 pppm
Case 4 was failed, giving the following error:
Hybrid pair style last /omp style table/omp
Last active /omp style is pair_style table/omp
ERROR: Requested neighbor pair method does not exist (…/neighbor.cpp:739)
Last command: run 300000
Regarding with the error “Requested neighbor pair method does not exist”, I am not sure why this happed. It is just not supported for a hybrid pair style or might be a bug, or I am using the table potential in a wrong way. Anyone can help? Many thanks.
Best wishes,
Feng