---------------
Minutes for IETF 74 meeting.

* Agenda bashing

* Volker introduced the WG
   - Resurrected and re-chartered
   - Went through the new charter
   - Looking for new research work and existing WGs and BoFs
     related to p2prg

* Enrico Marocco, Mythbustering P2P Traffic Localization

Distinguished the term "localization".
Huge interest in academia and industry
Caching is hard to do w/ p2p localization Traffic shaping possible, but not good for user interests Discussed how peer selection is done in p2p applications -
  agnostic to network topology and depends on tit-for-tat in
  BitTorrent, for instance.  RTT, jitter estimation are other
  techniques.
p2p localization wants to do better than average.
Research papers show it is effective, at least in field trials.
Don't know how it will be in the wild.
This draft takes a look at various issues discussed in many fora.
This draft needs to be moved in a community effort from this group.  Please provide input.
Outlined the myths -- 6 of them.
Went through the myths:
1) Reduced cross-domain traffic: Comcast, TU-Berlin, etc. show
  reduction in outgoing traffic and inbound traffic.  This myth is
  beleived to be confirmed.
2) Increased application performance: 4 references show that
  this is true.  2 references show some disadvantages.
3) ...
Time running out -- author wants to get a sense from the room on if this work is worth pursuing.

Stas Khirman: Yes, useful work.  Good to have info in one single
  place. P2P people mainly interested in application performance enhancement Richard Yang: Application performance improvement achievable.  Have a student doing "adaptive localization".


* Henning Schulzrinne, Security Issues and Solutions in P2P systems
   for RT communications.

Focus is on RT communications.

Went through attacker motivation: could be political, financial, or
for fun.

Attacker resources is primarily minting identities.

Attack timing: cracker joins, proves herself and then goes rogue.

Reviewed p2p for RT communications.
  - map names to other identifiers.
  - provide computational services (proxying and relaying)
  - store data

Went through file sharing vs. RT communications slide.

Admission control

In telecom systems, it is hard to identify malicious peers.  Rogue
file detection is much easier.

RT systems need just enough node to carry the load for data
computation.

Conclusions: RT systems are different than file sharing:
- more than just key – value mapping
- identity scarcity is crucial
- reputation system unlikely to work
- avoid “avoiding centralization at all cost” makes the problem much easier

Vijay: this is useful, Bell Labs worked on voice mail (issue with 
replication, encryption)


* Steven LeBlond, Reducing BitTorrent Traffic at the Internet Scale

2 open questions about BT locality
  - how far can we push it
  - how many savings can you get?

Real experiments with BT clients: Reduction in inter-domain traffic, 
performance penalty, up to 10000 peers, Mininova: 790k torrent files + 
some others: 200k active torrents

Small modification in clients: partition merging to control inter-AS 
connections

Explained parameters and metrics

Distribution of peers onto ASes: large reduction of inter-AS traffic for 
large ASes: good for ISPs

Performance slowdown: peers in big ASes run into partitioning problems: 
PM algorithm required to solve this à both ISPs and peers benefit
41% saving of inter-AS traffic
(0,05% of torrent require 50% of the bandwidth)

Comment Richard Y.: Different costs for inter-AS and subscriber links


* Zoran Despotovic: An empirical study of peer distributions in 
BitTorrent swarms

SmoothIT - EU Project

Measurement of P2P traffic: first determine distribution of peers in ASes

For many ASs it’s not enough to use local knowledge - wide network view 
is required for localization: ALTO should take this into account
Enrico: ALTO is not localization only

Stas K.: avoid the impression that localization doesn’t help, because 
most peers are in the large ASes.

Question: What happens to those ASes where there aren't too many
peers?  What algorithms will they use?

Can larger ASes live with the localization algorithm that works
in smaller ASes?


* Stas Khirman, Distribution of content popularity (caching efficiency 
estimation)

Scraped three trackers - TPB, Demonoid, and Mininova.  TPB has
1.7M torrents, Demonoid has 11K and Mininovas has 4K.

Distribution of peers to torrents: depends on type of content and ways 
of content promotion (community, professional content)

Geographical distribution of example movie (using whois): extend media 
distribution to countries where the movie is not distributed in other ways

AS distribution: peers per AS

Henning: would be interesting and helpful to align the results of the 
last three presentations to gain a strong data point, and a standardized 
way of presenting these results

Stevens: Issue with Demonoid trackers having several ports

Richard Y.: All studies so far focused on ASes.  Would be good to get
data per ISP, for instance.
Stefano: anyway - localization has to take into account infrastructure, 
independent of AS ownership



* Richard Yang, Efficient p2p design using in-network data lockers.

We need to focus on how to reduce the traffic on the last mile.

For downlink we do not want to reduce the traffic.  We are only
interested in the uplink, which is usually the bottleneck.

Some problems of P2P Cache Architecture such as
-	relation between user and cache,
-	tight coupling between cache and appl protocol:
-	diversity problem – caching a constant moving target

Went through their basic idea of data lockers: need a decoupled
forwarding and signaling arch.

Data Locker Access Protocol (LAP), API implementation, Resource Model 
based on weights given to different applications

Integration into ALTO / P4P and preliminary evaluation: peers inside an 
ISP (partially) have data locker capabilities

Question: LAP will be implemented at the client level?
Richard: Yes.
Q: Did you consider solutions that would be transparent to the clients?
Richard: No, do you have a solution?
Reinaldo: Why would a client upload to a data locker?
Richard: Avoid access congestion on last mile.  Avoids messing with
  tit-for-tat systems since I still get credit.

Yushun: Microsoft working on something similar, block transfer protocol,
Penno: incentive such as tit-for-tat still works, as credits are personal
Yushun: different parametrization of caches required for live streaming
Richard Y. Caches do not decide about lifetime; it’s decided end-to-end


* Christophe Diot, Nano Data Centers

Use resources on GW for P2P content services: Make use of GW uptime

Showed figures on GW uptime: Reason is VoIP functionality

Add memory: e.g. 10G flash memory
Comment ?: memory constrained only?
Christophe: BW, memory and CPU constraints: There’s an ISP trend towards 
high end RGW (many service such as beackup, home server, …)

Pros: ISP friendly and secure, large capacity at the edge of the 
network, energy savings

Cons: management, cost of gateways
Comment: ? Is there problem of malware on RGW: yes

Henning: 90% energy savings not just shifted from servers to end users 
bill?
Christophe showed some figures comparing energy consumption of servers 
and RGWs

Paper available: P2P Live-TV. Experimental evaluations (320 nodes, 
planetlab) of network traffic

Comment Aaron F.: Comparison with reliable centralized equipment

Comment Stas K.: Cable networks have different uplink characteristics 
than DSL and FTTH