Transcript www.tml.tkk.fi

Internet Indirection Infrastructure
Ion Stoica, Daniel Adkins, Shelley Zhuang,
Scott Shenker and Sonesh Surana
Presented by Essi Vehmersalo
Background
●
●
●
Current point-to-point communication abstraction
is not very well suited for mobility, multicast or
anycast
Solutions to decouple sending from receiving
–
Network level ->general, but require communitywide
consensus and commitment
–
Application level -> more deployable, but several
disjoint solutions exist (redundancy)
I3 attempts to provide a general but deployable
rendezvous based application level overlay
Service Model
●
Sources send packets to identifiers (id, data) and
receivers express interest in packets sent to an
identifier by inserting triggers (id, addr) to system
●
Abstracts the number and location of receivers
●
Further generalizations:
–
Inexact matching -> Anycast
–
Stack of identifiers -> Enables end-hosts to control
routing
Service Model (cont.)
Examples of i3 Usage
●
Enables for example
–
Service composition
–
Heterogeneous
multicast
–
Server selection
–
Large scale multicast
Design Issues
●
Overlay of i3 servers storing triggers and
forwarding packets, Chord used
●
Robustness by soft-state and replicating triggers
●
Triggers may be public or private
●
Routing efficiency: caching i3 server addresses,
choosing private triggers of nearby servers
Design Issues (cont.)
●
Caching triggers between i3 servers
●
Scalability and incremental deployment
●
Legacy applications can be supported with i3
proxy running on the host
Security
●
●
●
Problems and solutions of i3:
–
Eavesdropping -> using private triggers
–
Trigger hijacking -> indirection with private trigger
–
DoS against end-host or infrastructure (hierarchy of
triggers with leaves pointing to either victim or root)
-> challenges to ensure return routability of addresses,
limiting resource use of triggers, loop detection
Additional protection against denial of service
attacks
Using HIP with i3?
Simulation Results
●
●
●
Ratio of inter-node latencies of i3 and IP was
used as a metric (latency stretch)
Sampling triggers lowers latency stretch
considerably (in practice 16-32 samples enough)
Optimizations with considering network distance
in Chord routing improve performance 2-3 times
Conclusions and Further Questions
●
●
●
●
Provides powerful and flexible communication
abstraction
Still too early to evaluate the limitations and
expressiveness of i3
Economical model is still unclear
Embedding information into identifiers? Location
information? Hash of public key?