Thursday, August 27, 2009

Trip Report from SIGCOMM 2009 Barcelona ...

I am going to keep the identity of the writer of this secret, but here is what s/he had to say about the recent SIGCOMM conference. My comments are in [RHK: ... ] brackets. We will be reading several these papers as new additions to the reading list this semester:


Interesting points from SIGCOMM and associated workshops (as revealed by Student X):

I was at SIGCOMM etc. last week. Some high level points:

-Datacenter research increasing in prominence

[RHK: I included some papers on datacenter and enterprise networking as I think these are becoming increasingly important in the research community]

- Energy concerns increasing too - although people still confuse energy/power

- Still lots of research on wireless - two dedicated sessions vs. one for all other areas

- Lots of work on social networks - mostly focused on Facebook, although people are saying the focus should shift towards twitter.

Some more detailed interesting points below.


- Best paper award went to work on using freed up TV spectrum for WiFi type connectivity. Excellent idea. Not sure if lower frequency and greater propagation radius (compared w/ WiFi) would prove to be a fundamental challenge.
[RHK:I had planned to add this paper in the reading list!]

- Amin Vahdata's group presented PortLand as the next step to Fat Trees. Basically an L2 fabric that use a pseudo MAC to get L3 fabric advantages. Pseudo MAC has structure .... Also gives desired property of a datacenter network fabric - plug & play, scalable, small state, allow VM to move easily.

- MSR presented VL2 for datacenter architecture. Uses randomness to spread traffic around. Paper also includes datacenter traffic characterization - most flows are mice flows, most bytes are from elephant flows - maybe not surprising. They assume that data is spread everywhere in a datacenter so you can do fast VM migration for real.
[RHK: I've included both of these papers in the reading list this semester]

- MSR Asia presented BCube as next step for DCell. The idea is similar to DCell - make all routers equal.

- There was a paper on assigning dynamic application user id to de-anonymize the Internet - the specific goal was to catch and blacklist spammers. Not sure if de-anonymizing the Internet in general is a good idea.

- There was a paper on cleverly using routers as cache and thus generalizing CDN, WAN optimizer, proxy cache, etc. The presenter got some tough questions.

- Brighten gave an awesome talk on Pathlet Routing and handled all questions like a true pro - woot RAD Lab!

- Hari Balakrishnan et al. had a paper on cutting electric bill for Internet scale systems. Paper contains nice description of U.S. retail electricity mechanisms. The topic was well received I think. The proposal is to shift service requests to different geographic locations according to electric price differential. I think the scheme may have serious stability problems. Hari escaped SIGCOMM early before I got a chance to poke him in private.
[RHK: We are doing this paper too!]

- There was a paper in defense of wireless carrier sense, arguing that despite regular criticisms, carrier sense is sufficient in the average case. Not sure how relevant to RAD Lab, but interesting all the same.

- MSR also presented NetMedic for detail diagnosis in small enterprise networks. To me seems like X-Trace but at a different granularity and targeting a different problem space. Supposedly the most highly rated paper (not best paper award because the topic is not new?)
[RHK: I thought about including this paper with X-trace, but decided not to. This comes up later in the semester, so I may change my mind.]

- There was a highly mathematical paper on spatio-temporal compressive sensing, using the low rank nature of sparse/incomplete traffic matrices to interpolate the rest of the traffic matrix. I have trouble following all the linear algebra. Interesting because they used k-nearest-neighbors as one of the comparisons, but they never talk about it as a statistical machine learning algorithm.

- University College London presented a paper on rendevouz on a ring (ROAR). Trying to optimize datacenter application data replication/partition for something like search. Interesting because it uses a ring topology somewhat like Chord, and has some provable performance bounds.

- CMU gave the first Incast talk. Showed that default TCP misconfigured for datacenters - need to reduce TCP RTO min and use fine grain timers.
[RHK: We will be reading this paper!]

Workshop on Research on Enterprise Networks (WREN)

- For intrusion detection, flat policy for the entire network is not as good as dividing the network into a few subgroups and have policies for each. Performance is much better, but the overhead of per-host policy is still avoided.

- There was an idea on using Bloom filters in hardware to store routing state at edge routers. Avoids the need for ever larger amounts of fast memory to store routing state.

- There was a paper from Wisconsin/MSR looking at datacenter traffic characteristics. Interesting points include bi-modal distribution of packets - 10s of bytes and ~1500 bytes, on/off behavior, most discards happening at aggregation layer instead of edge or core.

- Berkeley Incast talk: the key point is that the story becomes much more complicated once you fix TCP misconfiguration.
[RHK: we will be reading this paper!]

WOSN - Workshop on Online Social Networks

- Private info leaks from OSN faster than water leaks from a fishing net - I guess we already know that. The question is do people care or do they just want their 5 min of Facebook fame ...

- There's an idea re owning VM with data instead of owning physical machines with data to ensure control over data in cloud environment.

- Youtube spend 5% of time on layout, 95% of time on scalability, according to one of their employee when questioning an auto-layout tool.

- A very small % of Flickr and Facebook users account for a very large % of all activity.

- OSN can design features that significantly increase user activity, e.g. Facebook bday reminder - I guess we already know that.

- There is a need for temporal metrics for fast changing networks a proposal for a distance-hop metric - I think we're recycling the space-time concept from relativity.

- Someone pointed to prior work in MSR that found people migrated from MySpace to Facebook due to perceptions of class - online networks are starting to resemble real life social networks.

- Someone asked how could/would/should a good Facebook app for the iPhone be different from the desktop version. The presenter didn't give a good answer.

- There was work on doing BlueTooth ad-hoc networks in the presence of WiFi infrastructure networks. The idea is to bypass WiFi network monitoring/limitations. Not sure if it's a good goal to achieve or not.

Other workshops:

- VISA — The First ACM SIGCOMM Workshop on Virtualized Infastructure Systems and Architectures

- PRESTO — The Second ACM SIGCOMM Workshop on Programmable Routers for Extensible Services of TOmorrow

- MobiHeld (formerly called MobiHand) — The First ACM SIGCOMM Workshop on Networking, Systems, Applications on Mobile Handhelds

SIGCOMM outrageous opinion session

- Proposal for anti-social networks, with explicit enemy edges and implicit friend edges, resulting in VERY interesting graph properties and network dynamics.

- Proposal for a green SIGCOMM next year in New Delhi. Key suggestions include serving only vegetarian food and turning off air conditioning. This one won the best outrageous opinion award.

- Proposal for Bollywood SIGCOMM next year - 2 min song and group dance sequence every 10 min of talk, alternated with 2 min fight sequence every 10 min of talk (audience participation required).

- Proposal for attracting more women to SIGCOMM by making the men more appealing. The ladies cheered wildly for this one.

- Proposal for resolving author list issues. The presenter's last slide was a pitch for writing single author papers such as "How to get your paper into SIGCOMM by Scott Shenker". Woot RAD Lab!