Summer Planning Session
Date: 5/4/01
Attendee: Randy Katz, Ion Stoica, Yan Chen, Sharad Agarwal, Morley Mao,
Shelley Zhuang, Weidon Cui, Bhaskar Raman, Matt Caesar, Jimmy Shih, Ben
Zhao, Lakshmi Subramanian, Sridhar Machiraju, Mukund Seshadri, Tim Sohn,
Keith Sklower

Yan:
1. Wide-area Network Measurement and Monitroing Project (with Morley,
Weidong)
1.1 Real data (NLANR, NIMI) analysis of the iso-bar clustering
  a. Classical K-means
  b. Stat learning approaches such as Gaussian mixture model, kernel
     methods, etc
  c. Stationarity (global vs. local, macro vs micro)
1.2 Wide-area distance estimation service
  a. service model (dynamic clients)
  b. feature vector extraction
  c. hierarchical clustering for scalability
  d. performance comparison with existing works
2. Content Distribution Network with Application-level Overlay Multicast
2.1 Use distributed location services (Tapestry) for better scalability,
     ease-of-management and DoS attack-resilience
2.2 QoS service placement on overlay network
2.3 Adapataion with monitoring service of proj 1
2.4 Adapatation through application-level semantics
3. For both projects, wide-area network testbed and evaluation

Sharad & Lakshmi:
Infer topology properties between service providers.
Allow better monitoring placement.
Where to place services? Exploit topology information. E.g local
preferences of each service provider.
How to set local preferences to better support overlay network?
What new algorithms would this information enable?

Sharad:
I will look at inter-domain routing (BGP) and how to classify routes into
the spectrum between bogus routes and low-latency/high-bandwidth routes.
The fidelity of the algorithm will vary when BGP routers do this
individually, to when all the BGP routers in one domain cooperate, to when
BGP routers across multiple domains cooperate, to a centralized broker
aggregating and distributing this information across multiple domains.
Staying to the right of this spectrum of cooperation will allow one to
better utilize inter-domain connections for both normal traffic (by
removing bogus routes and picking overall better routes) and for specific
QoS traffic (by picking routes with specific required properties). This
discovery of route properties can be done anywhere in the spectrum between
passive techniques to aggressive techniques. Passive techniques include
looking at a random sample of packets and looking for specific flows.
Aggressive techniques include dropping specific/random packets, tagging
certain packets and injecting new probing traffic. Staying to the left of
this spectrum is more desirable, and more achievable if we stay to the
right of the cooperation spectrum.

Morley:
- Light-weight, scalable verification mechanisms in the network for
   detecting  malicious end hosts and routers/overlay nodes.
- Verify QoS guarantees offered by service providers in the overlay
   networks.
- Verification between overlay providers and interconnection providers.
   Use it to build a more robust network.
- Summer: Working at ATT on CDN brokering, for building overlay network.
   Summer project to evaluate the system.

Shelley:
- Design and implement efficient communication services on top of a
   lookup service.
- Traffic engineering in wireless roof-top networks.

Weidong:
1. design and implement a framework for negotiation systems
  a) Who will negotiate?  What will be negotiated?  Where to negotiate? How
     to negotiate?
  b) How to evaluate the negotiation system?
  c) related work includes: SLA, user preference, mechanism design, etc.
2. continue cs268 project "efficient bandwidth allocation for secondary
    paths"
  a) Inside a core network, for guaranteed services, we may need to set up
     secondary paths to recover from the failure of primary paths.
  b) A generic way is to allocate the same bandwidth for a secondary path
     as that of a primary path. but it's not efficient.
  c) The problem is how much bandwidth in total should be allocated for
     secondary paths on a link such that the probability that the failure
     of primary paths is fatal is under some limit.
  d) Challenges: We don't know the failure model very well because no ISP
     publishes this kind of information.  We need a scalable way to
     calculate the bandwidth.
  e) Failure models: there are some ways to hack out data regarding failure
     model.
3. collaborate with Yan and Morley on the mornitoring project
  a) my focus will be on clustering the sites based on latency
  b) also, contribute my knowledge on stat to other problems, e.g.,
     distance estimation.
4. networking PRELIM in Fall!

Bhaskar:
Performance and fault tolerance issues.
Testbed on Millennium.
Undergraduates to work on composing services.
We can hire (pay) some undergraduates.
SUPERB program ­ Randy will follow up

Lakshmi:
Working on two problems with Ion.
1. QoS over overlay network. Better than best effort. Distribute loss
across multiple flows. Using rate control. Collaboration with Hari at
MIT. Flow control between overlay and interconnecting layers.
2. Routing verification. Passive verification model. Inter-service
providers.

Jimmy:
Testbed for value-based pricing.
Congestion pricing simulation with more users.
Collaboration with UC Davis.
Congestion pricing with video servers and video phones.

Mukund:
Working with Yatin with application level multicast.
Integration between different types of multicast. Protocol translation.
Performance guarantee across different types of multicast.

Sridhar:
Class project with Mukund: Providing QoS without maintaining per flow
state. Native multicast. Peering between multicast domains. Measure MSDP
protocol. (BGP for multicast)

Themes:
  Service pacement.
  Monitoring versus verification.
Related Work:
   Reliable network overlay at MIT
Services:
   Multicast
   Content distribution
   Content naming system
   Monitoring/measurement service
    Where to place monitors or other service infrastructure
    Distance, # of hops, latency, bandwidth, estimation service
Methodology:
   OceanStore: event-based simulator
    wide area deployment. Washington, Wisconsin, MIT.
    local node/physical machine.
    Summer/late summer.
   Chord - simulation on Millennium clusters.
    Simulate delays/large scale number of machines
   Millennium-based wide area network emulator.
    Keith: port Dummy Net from FreeBSD to Linux