Mdrun goes slower with OpenCL

User discussions
1

GROMACS version: 2022.4
GROMACS modification: No

Hi,

I am new to gromacs. So far, I successfully compiled and am able to run gromacs. For my first step, I am following ‘lysozyme in water’ tutorial (http://www.mdtutorials.com/).

I ran into some wired situation, so I came here.

The command, “gmx mdrun -deffnm md_0_1” runs much slower with GPU OpenCL acceleration. I mean when log shows GPU is detected and used, it runs slower.

The node used is headless, with Intel Core integrated 530 GPU.

My current goal is to run gromacs with mpi. But I am wondering what I should do now.

I appreciate any suggestion.

regards

2

Hi! Such low-end integrated GPU is unlikely to offer any noticeable speedup. It’s somewhat strange that it is much slower than the CPU-only run, but not being faster would be expected. So, it would probably be better to disable GPU support in your case.

3

Hi,

Thank you for your reply.

I thought even a low-end GPU would be something like “better than nothing”, but it is not. I will go without GPU.

it goes ca. 2500 to 8500 seconds.

regards

4

A few tips that might improve your performance:

  • Try -pme cpu to keep the PME computation on the CPU.
  • You will also likely get better performance by running fewer threads on the CPU, e.g. on a 4-core/8-threads CPU run 3 or 6 threads. This leaves more of the power budget of the chip to be used by the GPU.
5

Thank you for your suggestion.
-pme cpu runs ca. 4300sec.

mpiexec gmx_mpi ... runs ca. 3000sec without gpu. Or to be accurate, with OpenCL, with no compatible gpu.

gmx mdrun -nt 6 ... runs 8400 sec, -nt 7 also ca 8400 sec
gmx mdrun -nt6 -pme cpu runs 4300 sec

So far, OpenCL with no compatible GPU is faster than any other settings.

Without openCL, make check takes longer and several tests fail.

6

Hi

I also tried mpi.

Core i6, 8 cores, 2500sec

Core i6, 8 cores plus Core i3 4 cores 2500sec

For the time being, I go without GPU nor mpi.