There's two things termed backoff:
It used to be that when a node RNFed, the node it RNFed to would not send it another request for a while to keep it from being completely overloaded.
The backoff time increased with repeated
RNFs.
Currently, Freenet nodes have a limited capacity for requests, so they send other nodes "You can send me requests every N milliseconds". A node is backed off if it has been sent a request recently and therefore can't be sent another one yet.
The other one is
MRI. The two big effects of mRI seem to be
1) you need *lots* of nodes in your routing table to always have somewhere to send requests
2) routing is very irregular, because only nodes not within the mRI interval are considered for routing, so if you've just sent a request to the best place to send the current request, you'll have to send it elsewhere.
Current routing (29/06) in stable is almost degenerated to 'route to least busy', with slight preferences from
NGR but we're working on that.