GROMACS

Roland Schulz wrote:

• It seems odd that we automatically change nstlist to a value which then causes an error. It seems it would make sense that we change nstlist at maximum as much, so that the requested number of nodes give a valid domain decomposition.

Indeed, this can happen, IIRC I've seen it too.

it should be possible to the the effect of nstlist on DD and avoid failures (unlike avoiding slowdown which does happen too) - although it may require some complex logic (as the nstlist override happens before the DD initialization).

This message is not contradictory. It prints the maximum number of cells. In triclinic boxes with relatively large cut-offs, 2 domains in x or y might not works, whereas 1 or 3 will work. This is a bit counterintuitive, but unfortunately this is the case.
It would be nice to prevent the nstlist increase in such cases, but since that is done before the DD is initialized, this is hard. A better practical solution for this issue would be to allow one domain to utilize multiple GPUs. We are working on this.

I would agree with roland that the message itself is highly inconsistent, even though the code might not be. If we haven't done any work on this, could we at least update the error message to be more explicit and convey the message that even 2 nodes might not work in some cases?