Load average accounts to the number of runnable threads (threads waiting in the runqueue). On a 2-core system a load of 2 would mean, that on average there is 1 process runnable per core, which would be optimum usage. If you have a higher load, than the os has to make more context switches, which would increase the system cpu-load. - Henri

Which doesn't helps me at all. I'm looking for a cause here. - Andreas Gohr

You should try to find out, how many threads you have running. Maybe your http-server just starts more threads when there are many requests? A load of 5 isn't really that bad. - Henri

Found the root of the problem myself now. It's Fast-CGIs hanging in state D. Now I need to find out why they hang. - Andreas Gohr

Ok. Good luck! Typically these processes are waiting for an interrupt (for instance because of disk io). If they stay in this state, it could be a lost interrupt or a deadlock or disk read error... - Henri