Here are my notes from A New Way to look at Networking
by Van Jacobson.
Packet switching
old new
focus on paths ====> focus only on endpoints
conversations ====> packets
Transitivity makes routing emerge
they encapsulated the data and made a standard addressing structure to
glue together networks
TCP/IP too successful???
Connected - binary either your in our your out
Connecting requires - globally unique address that is stable minutes to
hours
Connection is heavyweight
point to point ( conversation ) is supported, but TCP/IP hates broadcast
Problem space for TCP/IP
Connecting users who have access to one machine
circuit and packet switched networks focus on getting two machines to
talk
New Problem Space
One user has tons of machines
A machine wants to read a named chunk of data ( web page, email, etc )
This is dissemination not conversation
Dissemination
Data is important, not the channel
Data must be secured, not the channel
Dissemination must be efficient, which it isn't with communication
Currently users have to do the plumbing to make dissemination
Protocol and medium agnostic - IP, zeroconf, multicast, pigeons)
Anything with the data can respond
Data is signed and optionally secured so receiver can verify integrity
Example
REQ http://nytimes.com/today
RESP http://nytimes.com/today
http://nytimes.com/20060830/index.html
<namemac>
<datamac>
<html>
...
</html>
Design philosophy
Data has a name, not a location
Integrity and trust are derived from data, not the channel
Anything that moves in time or space is a pipe
Communication
User communicates intent, so network can optimize
Request / response model gives user fine grained control over incoming
traffic QoS
Network is content focused, not conversation focused, so popular content
won't generate congestion
No distinction between bits on disc, on wire, or in memory
Nodes are nameless - wireless & sensor nets can use simple local
protocols like - flooding, proximity, diffusion, etc
Opportunistic transport - planes, trains, automobiles, as well as mobile
phones on people
Security
Trust and data integrity are baked in - solves Phishing, Pharming, and
Spam
Trust is associated with data not SSL channels
Security of country improves, can't disrupt carrier pigeon
Already solved
PGP
Repositories - "instant PKI", "UT Austin Snakebite"
Data location - URL, search, epidemics, directed diffusion, filtered
'small world'
Names and Naming
Augment names with time/version - cacheable, stable
Integrity preserving data segmentation - keep responses idempotent and
small
Nicknames and intentional names - "all the open doors in building A"
Hard Stuff
Incentive structure ( flow & congestion control, sharing &
redistribution incentives ) - leechers
Miscreant & freeloader detection - anonymity
Redistirbution - content routing, storage replacement strategies,
liability issues
Corrupted, invalid, or expired content - hints up the line
Conclusion
Telephones - plumbing at wire level - phone number is a routing program
IP - plumbing at the data level - URL is a endpoint from another
endpoint
Dissemination Based Arch - push data plumbing down - Decentralized named
content
Ad-hoc dissemination overlays - Akami CDN, BitTorrent, Sonos mesh, Apple
Rendezvous
by Van Jacobson.
Packet switching
old new
focus on paths ====> focus only on endpoints
conversations ====> packets
Transitivity makes routing emerge
they encapsulated the data and made a standard addressing structure to
glue together networks
TCP/IP too successful???
Connected - binary either your in our your out
Connecting requires - globally unique address that is stable minutes to
hours
Connection is heavyweight
point to point ( conversation ) is supported, but TCP/IP hates broadcast
Problem space for TCP/IP
Connecting users who have access to one machine
circuit and packet switched networks focus on getting two machines to
talk
New Problem Space
One user has tons of machines
A machine wants to read a named chunk of data ( web page, email, etc )
This is dissemination not conversation
Dissemination
Data is important, not the channel
Data must be secured, not the channel
Dissemination must be efficient, which it isn't with communication
Currently users have to do the plumbing to make dissemination
Protocol and medium agnostic - IP, zeroconf, multicast, pigeons)
Anything with the data can respond
Data is signed and optionally secured so receiver can verify integrity
Example
REQ http://nytimes.com/today
RESP http://nytimes.com/today
http://nytimes.com/20060830
<namemac>
<datamac>
<html>
...
</html>
Design philosophy
Data has a name, not a location
Integrity and trust are derived from data, not the channel
Anything that moves in time or space is a pipe
Communication
User communicates intent, so network can optimize
Request / response model gives user fine grained control over incoming
traffic QoS
Network is content focused, not conversation focused, so popular content
won't generate congestion
No distinction between bits on disc, on wire, or in memory
Nodes are nameless - wireless & sensor nets can use simple local
protocols like - flooding, proximity, diffusion, etc
Opportunistic transport - planes, trains, automobiles, as well as mobile
phones on people
Security
Trust and data integrity are baked in - solves Phishing, Pharming, and
Spam
Trust is associated with data not SSL channels
Security of country improves, can't disrupt carrier pigeon
Already solved
PGP
Repositories - "instant PKI", "UT Austin Snakebite"
Data location - URL, search, epidemics, directed diffusion, filtered
'small world'
Names and Naming
Augment names with time/version - cacheable, stable
Integrity preserving data segmentation - keep responses idempotent and
small
Nicknames and intentional names - "all the open doors in building A"
Hard Stuff
Incentive structure ( flow & congestion control, sharing &
redistribution incentives ) - leechers
Miscreant & freeloader detection - anonymity
Redistirbution - content routing, storage replacement strategies,
liability issues
Corrupted, invalid, or expired content - hints up the line
Conclusion
Telephones - plumbing at wire level - phone number is a routing program
IP - plumbing at the data level - URL is a endpoint from another
endpoint
Dissemination Based Arch - push data plumbing down - Decentralized named
content
Ad-hoc dissemination overlays - Akami CDN, BitTorrent, Sonos mesh, Apple
Rendezvous
0 Comments:
Post a Comment
Links to this post:
Create a Link
<< Home