Queues for processes like rred are not created by hostname but we
spawn at most CPU*2 queues to place items in. The problem is that we
then proceeded to limit it to at most 10 queues (via QueueHost::Limit)
again at the end of the method so that all items (after the first 10
queues are busy) are forcibly placed into a generic catch-all instance
which is bad because we don't keep all CPUs we have available busy and
worse we end up sheduling the most work to a single one while random
distribution was intended.