If you look at the second formula in the documentation (how the density is calculated) you will see that it involves an integral over the Brillouin zone. This is turned into a numerical integral, a.k.a. sum weighted by the k-grid volume element d**k** = Δk_{x} Δk_{y} Δk_{z}. The broadening should be chosen proportional to this value. If your system is long along one dimension then the Δk will be small along that direction given a constant k-point density as the Brillouin zone is inversely proportional to the length along that direction: for a cell with perpendicular lattice vectors you have Δk = 2π/(LN), where L is the length of the cell along that direction and N is the number of k-points. If you keep only one k-point along the C direction then the broadening should be proportional to Δk_{x} Δk_{y} = (2π)^{2}/(L_{x} N_{x} L_{y} N_{y})

As a rule of thumb you can use a broadening of 1000K for metals and 100K for semiconductors and if the metal has a hard time converging you can either 1) increase the k-point sampling (expensive) or 2) increase the broadening (cheap, but reduces quality of results due to artificial smearing)