Thanks! I'll try out the new run parameters and get back to you.

Have you tried the new velocity Verlet integrator (VV2)? The last time I simulated a water box with it, I got a large persistent net dipole that wouldn't go away, and very small fluctuations.