[Tech] Backoff considered harmful?

[Michael Rogers]

Ed Tomlinson wrote:
> eg. if we are basing send rates on average recieve rates from nodes
> with much less output bw than input we may be wasting lots of bw...

I don't think our sending rate to peer X is based on peer X's sending 
rate to us. As far as I know, the bandwidth estimation code decides how 
many requests to accept in total by measuring the amount of outgoing 
bandwidth used per incoming request, including requests that are 
answered locally. That means if we have spare input bandwidth *and* we 
can answer most requests locally, we'll use the spare input bandwidth. 
But if we have to forward the majority of requests, we'll be limited by 
our output bandwidth and probably won't be able to make full use of our 
input bandwidth. But I don't really see a way out of that situation - 
asymmetric connections just weren't designed with p2p in mind.

> What would happen if we only sent to one peer at a time (queueing
> other messages) or what would happen if we used peek recieve
> rates instead of average?

I don't see how sending to one peer at a time would help - we're more 
likely to be limited by our output bandwidth than their input bandwidth.

Not quite sure what you're suggesting about measuring peak receive rates 
- we're not trying to measure the peer's "real" bandwidth, we're trying 
to measure how many requests it can handle in the long term. The 
low-level congestion control code should take care of short-term peaks 
and troughs in the available bandwidth.

Cheers,
Michael