simtk.org

Hi,

I'm an experienced GROMACS user and I'm having some problems getting GROMACS-openmm working.

First, I cannot get the CUDA version to work. I get the following error:

cudaMalloc CUDAStream::Allocate failed out of memory

As far as I can tell, I have the CUDA driver properly installed and my paths setup correctly. I can run some of the NVIDIA samples, so I think the libraries should be installed correctly.

Second, if I rename the openmm cuda dylib so that I am using the reference CPU library, I get a warning that all of my Born radii are 1e-6 and my simulation quickly blows up. I have copied the params.agp file to the appropriate location. I notice that both the windows and linux versions include a modified grompp binary, while the Mac version contains neither a binary or source modification to grompp. Is a modification necessary to read in the appropriate Born radii and surface tensions?

Finally, although I realize this is a preview release, the current method of distributing this package is very confusing. I think it would be much more convenient to release a complete binary package for each platform. I have built GROMACS literally dozens of times on different platforms and I have no idea what I have done wrong in this case. I've tried several times and have followed the instructions exactly and I still don't have it working.

Any help you can provided would be very much appreciated.

Cheers,
Justin

Hi Justin,

The message means that your GPU doesn't have enough memory to simulate the system. How many atoms does it include? And what GPU do you have? High end ones will typically have 1 GB or more of memory, while low end ones might only have 256 MB.

I've done a lot of work recently to optimize the memory use of OpenMM, so the next release will be able to handle larger systems. For the moment, though, it's fairly limited, especially if you don't have a high end GPU.

Regarding the params.agb file, it only includes Born radii for the Amber96 force field. If you want to use a different one, you need to modify it and add in the atom types for your force field. (Only the first two entries in each row, atom type and radius, are used, so you can omit the other columns.)

> the current method of distributing this package is very confusing.

Agreed, and we definitely want to improve it! I can assure you, though, that you didn't make any mistakes in installing it. It's working, but producing a (not as clear as it could be) error message due to running out of memory.

Peter