Yes, the problem was the g_ewald estimation, probably a sign problem.
I cannot set the dielectric constant to 1.0 for my simulations, I
would get a repulsive gravitational potential instead of an attractive
one.
The coulomb potential is defined as
Phi_C = C*q_i*q_j/eps/r
whereas the gravitational potential is defined as
Phi_G = -G*m_i*m_j/r
so I need to get the sign (-1) into the equation. Since I cannot
change C, I had to change the dielectric constant epsilon...
As I said, not pretty but working 
Andreas