[BUG]lmp MD very slow with offline installer compared with compiled myself #730
Description
Summary
The lmp command in the offline installer seems to be very slow compared with the lammps compiled manually.
We have try to test the speed of lammps of water model (compressed and not compressed) on aliyun GPU-V100
lammps NPT simulation 10000 steps. unit: timesteps/second
| which lammps | speed of compressed model | speed of not compressed mode |
|---|---|---|
| deepmd-kit-2.0.0.b1-cuda10.1_gpu-Linux-x86_64.sh | 9.825 | 12.485 |
| deepmd-kit-2.0.0.b1-cuda11.0_gpu-Linux-x86_64.sh | 9.694 | 12.422 |
| lammps by @denghuilu , cuda11.0 tf2.4 | 165.754 | 23.405 |
Ubuntu 18.04 aliyun GPU-V100 single node.
lammps-dp compiled manually and lammps in offline installer mentioned above.
lammps input: a trivial NPT simulaiton script pair_style deepmd graph.pb
case1: download offline installer and run lmp command.
case2: compile tensorflow 2.4 from source (as well as libtensorflow_cc) lammps dp and compile lammps_mpi
case3: download offline installer and use this env (this means use the libtensorflow_cc.so libtensorflow_framework.so in the installer when linking) to compile deepmd-kit and lammps .
Further Information, Files, and Links
2,0 model and compress version , lammps in.lammps , conf.lmp
https://deepmd-kit.oss-cn-beijing.aliyuncs.com/water_wanghan_models.tar.gz