Hi,
I think it is very different to say that P2P locality does not work (or
at least does not optimally work)
with current proposed solutions or current distribution of peers
capacity, and that P2P locality does not work.
I agree that the former might be true on some scenarios, not the later.
My understanding of P2P locality is that it aims to reduce traffic on
specific links, often referred to inter-ISP links, which is
misleading as it is so easy to come up with an ISP that don't want to
reduce traffic on those links. There are probably as
many special cases as ISPs.
Improving download completion time of peers might be a nice side effect
of P2P locality but it is not a target
(combining caches and P2P locality is an interesting way to also
improved peers download completion time)
The P4P solution is very interesting from an architectural point of
view, as it gives freedom to adapt locality
to the need of each ISP. However, the field test is
somewhat limited as it focuses on Comcast and a single large torrent (as
described in the SIGCOMM paper). As Richard said
depending on the provisioning of peers you may observe different results
in terms of download completion time. But, no matter the distribution of
peers bandwidth is, you still reduce traffic on the targeted links.
In the case of Ono, the authors never shown that Ono reduces inter-ISP
traffic, but that it reduces the AS paths between peers.
This is very different. Instead, Ono focuses on improving peers download
completion time. In that case, the distribution of peers bandwidth
really matters.
Coming back to the question "Does it make sense to do P2P locality?"
my answer is yes. In our work "pushing bittorrent locality to the limit"
http://hal.inria.fr/inria-00343822/en/
we show measuring 200K torrents, 6M peers and 9K ASes that the potential
reduction of traffic on inter-AS links
is 40% (more than a 4PB of data reduction on inter-AS links on a few
hours) taking into account all those
tons of small torrents for which you cannot gain anything because you
never have two peers within the same ISP.
One can still argue that inter-AS links is not always what you want to
optimize,
as long as you want to reduce traffic on specific links, it will not
change much our results.
We show that increasing or decreasing the peers download completion time
will depend on where the bottlenecks are.
Therefore, the network bottlenecks will have a large impact, in
particular when they are on the links on which
you want to reduce the traffic (this is the case we evaluated), but the
access speed of peers will also matter.
Regards,
Arnaud.
_______________________________________________
alto mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/alto
