I am running a protein ligand complex simulation using Gromacs 2021 on a Windows Subsystem for Linux. My laptop has Nvidia GeForce RTX 3050 GPU. When i run the simulation of Lysozyme tutorial (as given in GROMACS tutorial) for 100ns, the expected finish time is showing approximately 1 week. I looked at the topology file to get an understanding of system size and found that my total system size is approximately 47500 including Solvent, ions, protein and ligand.
1) The "dt" defined in the mdp file is 2 fs and the number of steps (nsteps) is 50000000. I wanted to know if there is a way to speed up the process or is this the natural computation time that RTX 3050 provides? I went through other queries about the same issue and also I have worked with RTX 3080 Ti which completes a 100 ns simulation in approximately 30 - 40 min. So I assume that since 3050 also belongs to a similar class/family as that of 3080 Ti, it should atleast provide a better simulation timing (say 100 ns in 1-2 hours). I might be wrong about the technical aspect of GPU computation. Any help in this matter will be much appreciated.
2) Also, I wanted to know, since I am running these simulation in Windows Subsystem for Linux (WSL2), does that affect the computation speed of the GPU when MD Simulations are run using GROMACS?
I would appreciate if someone can help me out in this regard.
Thanks
Satyam
Dear Satyam Sangeet
No you cannot change the dt = 2 fs, that's force field dependent. If you go coarse grained (like MARTINI) then you may go for 20/30/40 fs, but as long as you stay in the GROMOS family as per the tutorial (I think it was a GROMOS united atom representation right?) then you generally wont have longer dt. You can (and generally will always) change the total number of steps nsteps to define the length of the simulations.
The results are quite bad for such a GPU. I think the problem is that you are not really using it. I know that back in the day WSL1 didn't let the users rely on GPUs, but this should have been fixed in WSL2.
What is the command you use to run the simulations? is the GPU installed in WSL2?
It's very likely that (i) the GPU is not installed in WSL2 (so the nvidia cuda drivers are missing), (ii) GROMACS has been compiled without GPU support, (iii) both the latter together.
Best
Nicola
Dear Nicola Piasentin
Thank you so much for your answer. I can assure you that the GPU is installed on WSL2. I have installed the latest CUDA drivers also. Moreover, GROMACS has been compiled with GPU support enabled.
Also the command that I am using to perform the production run is:
"gmx mdrun -v -deffnm md_0_1 -nb gpu".
Once I run "nvidia-smi" I can see that the GPU is being utilized. Also when I run "gmx --version" the "GPU support" tab is shown as "CUDA". Also the CUDA driver and runtime both are shown as version 12.1.
Initially I used to face the GPU issue because when I used to compile GROMACS with GPU support on and run "gmx --version" after the successful installation, I used to get "GPU Support" as disabled. Moreover, the "nb -gpu" tab would not work. But right now I can confirm that GROMACS is compiled with GPU and the GPU is being used as well. But I am not sure why the computational speed is so poor.
Best
Satyam
Dear Satyam Sangeet
Nice to know that everything is installed. I never used GPU on WSL, so I don't know how much further I can help you. Maybe be more specific in the command
gmx mdrun -nt #number_of_cores -v -nb gpu -gpu_id 0 -deffnm #nameofyourfiletpr
but I bet nothing will change.
Generally, things work better on a pure Ubuntu installation rather than on WSL. Here for examples are some results for GROMACS sims (only cpu here but it seems true for GPU as well)
https://www.phoronix.com/review/wsl2-ryzen9-5900x/2
At this point (just a speculation) maybe there is some windows feature that is blocking the performance of your GPU? I would seriously consider getting at least a partition on your laptop and install there Ubuntu. Keep in mind, anyway, that for long/large simulations a laptop is not good, generally speaking.
Best
Nicola
P.S. Did you compile gromacs or did you install precompiled binaries with something like $> sudo apt-get install gromacs?
Dear Nicola Piasentin
Thank you so much for your answer and pointing out the production run code to be more specific. I already tried the "-nt" option but as you correctly pointed out, nothing seems to change in the computation speed. I have compiled GROMACS from scratch.
I was also speculating that using the pure Ubuntu OS would be the right thing to do over here rather than WSL2.
Best
Satyam
- Badges
- Science topic