In parallel MD simulations, the calculation of most interaction terms are spread over many computational nodes, but the calculation of colvars is not parallelized. Therefore, additional calculations are executed by the node calculating the colvars, and most importantly, additional communication is added between the first node and the other nodes. To mitigate that, atom groups should be kept relatively small (up to a few thousands, depending on the computational cost to simulate the system by itself). A test calculation with VMD can provide a crude estimate of the impact of a large colvar on a NAMD simulation.