Some preliminary findings..
Changing the kernel seems to have a positive effect on the load. I changed from 2.6.27.46 to 2.6.27.54 (sorry, im bound by locally available kernels due to a kernel patch we created to fix some NFS problems in the linux kernel. Patch should be available in the stock linux kernel as of 2.6.36, but we dont have that kernel locally available yet), At similar user levels the load is now back to what it was. It doesnt however influence the context switches. Still very high. Here's a graph of a server where I changed the kernel and nothing else:
load: http://grab.by/7Odt (i rebooted the server with a new kernel at the point where the high load drops to 0)
cs: http://grab.by/7Odw
Here's a server with the old kernel, but imap-login { service_count=0 }. It seems this doesnt have much impact. cs are lower (but higher than 1.2), probably because user levels havent gone up to normal levels yet due to loadbalancer distribution issues after a reboot)
load: http://grab.by/7OdL cs: http://grab.by/7OdN
Ive now also set up a server with new kernel and service_count = 0, just to cover all the bases. But not enough data on that server yet. It takes a while for user levels to creep up once a server gets rebooted as users are sticky to a server once they're linked to it.
Cor