Hi Greg,
thank you for that whole bunch of information! 
I have no redirecting surfaces, but I use (mk)illums to bundle groups of light sources for the indirect calculation. I do not know if this causes additional start-up processing.
The idea to deactivate the obstruction cache is quite promising for my case. I have one illuminated room, so all those light sources are not hidden by surfaces any way. That means that all that caching is for nothing and just wastes the start-up time in this special scenario.
I will try this as soon as the next image is done (I guess tomorrow) and report any success.
The idea to use -P is a good one - but I wonder how to do this when using ranimate to control the rendering. Maybe you remember, I am that crazy guy using ranimate to render stills - which is nice cause I can use it to continue after broken processes, distribute processes both on smp and on clusters, and all that... I will have a look at the ranimate documentation once again.
CU and thanks, Lars.
If you don't have any mirrors or prisms in your scene, then the most likely culprit is the source obstruction routine. Since there are no options controlling this calculation, the only option is to recompile the rendering tools setting -DSHADCACHE=0 in your rmake script. Again, the overall calculation might end up taking longer, but you will avoid the start-up time.
This reduced the start-up time from several hours to 15 seconds... still I have not seen the results of the rendering 
Did you check what happens if you have your 1800 sources in a large scene?
CU Lars.
My 1800 sources are in a large scene -- the octree is a few hundred megs, I think.
OK, that is strange. I try to list the more or less special characteristics.
- use of instances
- use of mesh primitives
- colorpict mappings
- lots of lights (about 1850)
- light sources are grouped into spheres, that are processed by mkillum. the illum alt material is an invisible trans type
The rpict process takes about 510MB, already being optimized for memory using instances as mentioned.
Should I get some debugging tools installed to check which functions use up all the time at the beginning of occluder-cached renderings?
CU Lars.
I have no redirecting surfaces, but I use (mk)illums to bundle groups of light sources for the indirect calculation. I do not know if this causes additional start-up processing.