PeerWise Discovery and Negotiation of Faster Paths Cristian
Lumezanu, Dave Levin, and Neil Spring (University of Maryland)

Summary: One of the problems for overlays is that measurements and probing is
not scalable.  They assert that there should be some incentives built into the
peer selection, e.g., in AS-level peering, where there are mutual benefits,
although in this case the benefits could encompase things like latency to the
destinations, and the amount of traffic.  The evaluate PeerWise for fairness,
scalability, and benefit to the participants.

Q & A:

Q: How do you measure the mutual advantage

A: It is the number of shortest paths available through the participants.

Q: It may not be the best model, for example, if I communicate
heavily.

A: I agree.  We need better metric.

Q: What happens when a peer is no longer beneficial.

A: The peer can terminate the peering.

Q:  Although PeerWise is scalable at the same time it may not give best paths.

A: We use the inaccuracies of embeddings measure the inefficiencies of paths and
attempt to choose the best ones.

Q: Is the lack of mutual benefitial relationships the real reason why Overlays
are not popular.

Ans: There are other reasons, one is that we do not clear about benefits, and
second is the ISPs limit the routing between peers.

Q:  What are the timescales of peering. 

A: We expect these to be long lived. 

Q: Some times the benefit will be balanced over time.

Related Q: Tit-for-tat in bit-torrent works because the benefit is
immediate. 

A: I feel that for it to work in the real network, there has to be a
way to balance long term benefit, and also the peering needs to be
long term

Q: Have you considered transitive benefit?

A: No, we have limited it to peer-wise benefit

Q: Will you have a situation where the benefits are limited to
high-end and lower-end users, as in bit-torrent?

A: We believe it would be the rise of the middle-class. 

Q:  I think you cannot make that conclusion from your dataset.

A: Yes, that is true, but it holds for our dataset.