Sorry to revive an old post but this seemed to me the most logical place to ask this question.
I was following @jalemkul 's excellent tutorials, some of the mdps (for example) of which set nstcomm to 1, but I get the grompp error you @alevilla mention. Setting nstcalcenergy to 1 definitely slows the simulation down (I’m not sure exactly by how much but simulations that ended overnight were now ending after Close Of Business on the next day, so it is a de facto 2x slowdown). Here @anapaularp mentions attempting nstcomm = 10 and @pszilard suggests one might not want to have (or should at least look out for) such a low value of nstcalcenergy. Given all that, I was wondering how one decides on a reasonable value for nstcomm? I suppose one can do it iteratively: pick a value, start the simulation watching the drift of the COM, extrapolating whether it might violate a reasonable bound for the planned duration of the simulation, and restarting at a new value, perhaps bisecting on an optimal value. But this is cumbersome and assumes one knows the duration a priori. Is there some recognized standard for this choice?