We've had a similar problem, something about the AFP process pegging itself. It'll run the procs to 50 or 60 percent when then should be at 3 or 4.
Stopping the service and starting it again seems to do the trick. Sorry I don't know a better solution; we're still wrestling with it.