Hi, as this is our paper that generated this thread, I just want to clarify one point. It is obvious that there is a need for locality in P2P, as P2P does not take into account any constraints of the underlay network. We have shown (as others before us) that this lack of locality awareness create a huge increase of traffic on some links the ISPs would like to see less loaded. What are those links, how to define locality (AS level, ISP level, fine grain sub-ISP level, predefined static locality definition, etc.) is orthogonal to our study.
We have shown that whatever the definition of locality is, if you don't take it into account, then you will unnecessarily overload some links. One of our main contribution is that BitTorrent can efficiently keep the traffic local without significantly impacting the end-users download completion time, and this result still hold with locality values much higher that what was previously evaluated. Even if that may seem straight forward to some, I can give you tons of arguments why it could be different. Our result is the mere (yet another proof) of the fantastic efficiency of BitTorrent to replicate fast contents even in case there are complex constraints to handle. Believe me or not, there are still persons discussing the simple fact that BitTorrent can achieve good utilization of the upload of peers with a large number of peers. You will find as a side result that in our experiments with 10 000 peers arriving in a flash crowd of 60 seconds (quite challenging, isn't it, at least to anybody who ever try to build a fully distributed system to replicate contents), BitTorrent is only 40% slower than an equivalent client-server architecture with a seed having the aggregate upload capacity of all peers (1.6 Gbit/s in our experiments). Thus BT is only 40% slower than the optimal, taking into account that I consider optimal a completely different scenario, which is quite unfair for BitTorrent. I am still amazed when I find experimentally such results. Arnaud. jul wrote: > On Dec 4, 2008, at 11:50 AM, Matthew Kaufman wrote: > > >> Having built multi-Gbps nationwide ISP networks myself, I can tell you >> that going on an AS basis is useless. >> > > Thus why I listed it second. But the fact is people use it just like > they use prefixes which are to some people considered useless too. > What used to be Joost did locality by being only prefix-aware and they > admitted it sucked. Then there is ONO which creates large amounts of > DNS traffic, etc... Personally I don't do much locality since I use > modulo-arithmetic based routing, I don't care too much about the > physical coordinates... > > >> If I have excess capacity to another ISP over a settlement-free >> peering >> link that comes out of the data center where I've homed the ATM >> traffic >> coming up from your DSL line, I'd much prefer that your BitTorrent >> client send lots of data to users of that ISP than send them to people >> on my network that are at the extreme other end of my national or >> international backbone net, perhaps over an ATM or MPLS tail that I >> pay >> burst charges for when you're doing that. >> >> P4P provides more data, assuming I'm willing to disclose those sorts >> of >> things to the public, but it still isn't sufficient *and* if I tell >> you >> I prefer that you get data from X but you find that you get the file >> you >> want in 1/20th the time by getting it from Y, I probably can't >> convince >> you anyway. >> >> Matthew Kaufman >> >> jul wrote: >> >>> P4P or pure AS caching. >>> >>> ~J >>> >>> On Dec 4, 2008, at 7:50 AM, [EMAIL PROTECTED] >>> <mailto:[EMAIL PROTECTED]> wrote: >>> >>> >>>> P2P locality seems to me a red herring since most major ISPs: >>>> >>>> * have markets in several countries and geographic areas >>>> * lease physical facilities or IP networks from and to other >>>> ISPs; a very complex mapping >>>> * keep their L2 and L3 network maps and leasing costs very secret >>>> for competitive reasons. >>>> >>>> >>>> Having worked many years for a leading ISP with global reach, I >>>> guess >>>> locality would have seemed hard to define to our team. >>>> Unless the p2p protocol measures the latency, bandwidth and the >>>> number of IP hops itself. >>>> I have not seen any protocol that can also measure ownership of >>>> links >>>> by ISPs and numbers of hops between ISPs, their $$$ costs and >>>> populate the routing table with such metrics. >>>> Please share if you know such a technology. >>>> >>>> Henry >>>> >> _______________________________________________ >> p2p-hackers mailing list >> p2p-hackers@lists.zooko.com >> http://lists.zooko.com/mailman/listinfo/p2p-hackers >> > > _______________________________________________ > p2p-hackers mailing list > p2p-hackers@lists.zooko.com > http://lists.zooko.com/mailman/listinfo/p2p-hackers > -- Arnaud Legout, Ph.D. INRIA Sophia Antipolis - Planète Phone : 00.33.4.92.38.78.15 2004 route des lucioles - BP 93 Fax : 00.33.4.92.38.79.78 06902 Sophia Antipolis CEDEX E-mail: [EMAIL PROTECTED] FRANCE Web : http://www-sop.inria.fr/planete/Arnaud.Legout/index.html _______________________________________________ p2p-hackers mailing list p2p-hackers@lists.zooko.com http://lists.zooko.com/mailman/listinfo/p2p-hackers
