The MLFMM method depends on the matrix being sparse, so that coupling between some parts of the matrix and others can be described by coupling between groups of elements rather than between individual elements The message seems to say that when the matrix is 20% full this becomes less valid and also perhaps the computational advantage to this method is reduced.

It looks like you have a problem where too many of the parts couple to lots of the other parts significantly rather than being able to be regarded as parts of small groups interacting with other groups.