Why Spectre noise-aware PLL simulation give huge PSF output and taking long time?

we are trying to follow the spectre RF noise-aware PLL sim flow. Using MMSIM71 and IC5141_USR5(with latest hotfix), and the PLL is running at 2.5GHz.

It goes well at the beginning, the divider ,CP and VCO model was extracted sucessfully, though we can not view it since the model was encrypted.  We don't see much problem since it is a general PLL structure and the flow is very straightfoward too.   BTW, We set the ChargPump load voltage (Vload) to 0.45~0.75.

The transient simulation also goes well at first, untill the chargepump load voltage  gradually reached 0.45V, ie, the lower boundary of the voltage range that we set previously  in CP model extraction.  Now the simulation become very slow, the simulation step is in 0.01 fs. It takes 6 hours to simulate 3us.  And after the simulation run to 60us, the result psf data has become 65G bytes, which is not possible to be handled by wavescan.  Basides, my PLL will settle only after 100us. 

For VCO CMI model, we set the "sampling point per second" to 50, and "inject jitter" to 100us. 

Any clue any what's happening?  Please kindly help, thanks a lot.