Sunday, September 02, 2007

Today at KAIST in Daejeon, Korea - Workshop ReportTrip Report - Late August/Early Sep 2007, Japan & Korea

Trip Report - Late August/Early Sep 2007, Japan & Korea

I've put my comments on the CCR feedback site for
this years SIGCOMM papers, so this is my trip report
for the Kyoto part of the trip and you can get papers fom here too:

Also see Pan Hui's SIGCOMM Mobiarch workshop paper.


my talk:
H0: what if use haggle fwd for synch comms?
(compare to MANET (dongman lee,, asks) - proactive, reactive, location based)
KAIST results show can always do better!

H1: how to control storage needs of DO for forwarding:
Me: have to implement forgetting (cache), and assume DOs within community for
community relevance has good locality

Mia's talk: chris anderson, "The Long Tail"
(based on IMC student award paper)

Yangwoo Ko - late binding and use of
app. info for deciding how/when/who to forward
definitely haggle relevant!
SGR addresses H0, above!
next steps: use name service (NO/DO like)

Bodizar Radunovic, MC2!
so what do you do about lost tokens (e.g. nodes leave system) or new nodes?
plus do you have any ideas how to stop fake credit tokens?
for one idea (near end) on how to deal with churn...

Levy walk nature of human mobility and its impact on DTN/MANET performance"
Kyunghan Lee (KAIST)
has some very large (well 50+ node) measurement using fine grain GPS, then does
model fitting - shows truncated heavy tail (Levy) is good for people, better than Brownian, but also RWP and Levy yield same gain in forwarding, so
also haggle relevant! - concludes: shows DTN is MANET with higher delay,
but higher throughput! very nice!
see also
Andrew Edwards, who gave a talk at MSR while ago about the albatrosses
experiment in Antarctica, is now in Canada. Nature paper (to appear in next 8 weeks)
shows it isnt a good model after all!

"On the Optimal Resource Allocation in Multi-hop Wireless Networks"
Hyang-Won Lee
and this one too!
capacity as resource, but interference is simple model


"R-BGP: Staying Connected in a Connected World"
Bruce Maggs, as per NSDI paper - jolly good.

"Lessons learnt from large scale P2P systems and how they apply to the Future Internet"
Pablo Rodriguez, ex Inktomi, MSR-C and now at TIS (telefonica's new lab in Barcelona)
now: P2P from an ISP (and telco)'s perspective... ... ...message: its useful, not just for piracy.
But 2 key problems:- interference between P2P overlay traffic demand, and underlying TE and
peering/customer provider relationships between ISPs. Point of view is that there wasn't a way to bill for
ip multicast multiplier, but there might be a way to bill for p2p = by volume.

Note: sender gets money by number of recipients of content.
If use 1-1, easy - if do multicast, get a log cost.
if do P2P, get log net cost and reduced sender/storage costs.
log cost means you need a recursive tree charge based on sub tree.
reduced data center cost is just too bad for CDNs:) - one point about p2p v. multicast:-
p2p requires end user to increase symmetry of uplink -
you don't really need to charge by volume precisely - you could charge by time the uplink is full -
the end user could have this offset against content charge made by the content provider - value chain is
1. telco charges source for attachment rate
2. if net does packet copy, each branch point gets charged by telco for number of branches, recursively
3. leaves get to pay this charge, and have it subtracted from content lehal/drm fee

2bis: if leave does p2p, uplink, they get charged by net
3bis but get it discounted against content fee (since the uplink means that
they've succesfully re-sold content to some other receivers who are
paying for content but not for net yet


rest of talk on ip multicast tv - as per recent work

ed Power-Delay Tradeoff in Wireless Ad Hoc Networks Using Opportunistic Routing"
Won-Yong Shin (KAIST) - info theory analysis to find the advantage of fading in opprtunistic networks,
of getting capacity greater than the gupta/kumar model (e.g. from diversity) - excellent stuff! basically seems
to show that the Tse 2007 idea happens for free in a DTN...assumes only CSI available only at receiver:-
Opportunistic routing based on Biswas&Morris - packet proceeds to next hop closest to receiver; so change this to send and chosenext tx in next microcell to be any node with valid reception - gives diversity thru randomness - gain is ln(N) for N nodes per cell.

"Study on network size estimation schemes for peer-to-peer networks"
Hosik Cho (Seoul National University): reminds me of one of the microsoft research things
done by richard black and austen donnelly to do network discovery (think its in vista now)+and (sue moon points out) also multicast group size estimation.

"Spamscatter: Characterizing Internet Scam Hosting Infrastructure"
Geoff Voelker

All good talks - very active and very timely research, and good awareness of current hot topics and techniques to tackle them!

No comments: