Design Philosophy
==================

* Written in 1988, around 15 years after ideas underlying the Internet were first developed
* Mostly a retrospective, but a very influential paper
* Written to teach newcomers to the IETF/IRTF why the Internet was the way it was
* Been revisited many times since.

Key ideas in DP paper
==================
* The main idea: Spelling out the design goals of the Internet and how it influenced the design of protocols

* What were these goals?
--> "Effective technique for multiplexed utilization of existing interconnected networks"
(originally, the ARPANET and the ARPA packet radio network; foresight: "there would be other networks").

* Why packet switching? Well understood from the days of the ARPANET

* Other goals in priority order:
--> Resilience to networks or gateways being lost (military origins)
--> Generality to support multiple types of communications (The designers did not what would run on the Internet.)
--> Variety of networks (wanted to get everyone on the Internet)
--> Distributed management (no single controlling entity)
--> Cost effective
--> Host attachment with low effort
--> Resource accounting
Importantly, commercial utility was not a concern back then (actually until the mid 1990s).
Interesting difference from the telcos. Extremely well-engineered, single entity control, performance guarantees.

* Question: What did they miss? Performance guarantees, security, mobility, hosts may not be computers anymore.

* How did they do (on whatever they cared about)?
--> Surprisingly well!
--> It's quite hard to build something when you don't know what it will be used for (generality), yet ...
--> the Internet is good enough ("only just works") for almost everything.
--> Unforunately, it's optimal for almost nothing.
--> This is the principle of best effort: think of it as postal service without insurance
--> Important to keep this in mind before critiquing the design. It was good for what they cared about.
--> Doesn't mean we shouldn't think about new ways of doing things, which may be better for what _we_ now care about.
--> E.g., Performance, security, mobility, etc.

The goal of resilience
===================
* Only a network partition should prevent communication. Communication should
  continue under all other circumstances.
* Need to protect state associated with a conversation.
* What if we used hop-by-hop reliable delivery? What if some hops dies and
  takes the state to its grave?
* The development of the "fate sharing" principle: OK to lose state if entity
  associated with state dies. Very democratic in some sense. Minimizes
  collateral damage from routers dying.
* Resilience is still second to interconnection. So there are no failure
  reporting mechanisms assumed of the end hosts.

The goal of generality
===================
* Need to support multiple different services
* Example from 1988: file transfer, remote login, the cross-network debugger
  (unreliable), and audio
* Early observation: Reliability hurts latency because a reliable, in-order
  service results in HoL blocking.
* Resulted in the use of datagrams/packets as the underlying building block
* Use raw datagrams for good delay characteristics
* Use TCP layered on datagrams for reliability
* Btw, not all networks are built this way. E.g., Infiniband is a network
  technology that guarantees reliability, even if the application does not want
  it. 

The goal of network variety
===================
* The Internet had to easily incorporate existing local networks.
* Helped expedite the growth of the Internet.
* Examples: Satellite, radio, LAN, serial links
* How did they achieve this? Minimum assumptions on what the network provides so that everyone could interoperate.
* End-to-end argument: Place functionality within the switches/routers only if
  it's necessary and sufficient. Otherwise, leave it at the end hosts.
* What was the network not expected to do?: Reliable delivery, prioritization, broadcast, multicast. 

Other goals
=================
* Distributed management lead to two-tier routing
* Cost was sidelined: Retransmissions have to go through the whole network.
* Accounting: again sidelined.

Summary: Goals and their consequences
=================
* Multiplexed utilization of existing networks => packet switched networks connected by packet forwarding gateways
* Resilience => Stateless switches, fate sharing
* Generality => Datagrams as a building block
* Network variety => Minimum assumptions of the network; move everything into hosts; end-to-end principle 
* Distributed management => Two-tier routing
* Cost/accounting => Relatively sidelined (retransmissions have to go through the whole network)

Aside:
=================
David Clark has much more to say on this topic.
Highly recommend this book: https://mitpress.mit.edu/books/designing-internet
and this video: https://www.youtube.com/watch?v=qX-ojw1gLmE
For a history of how the ARPANET---and then the Internet---came to be:
https://www.amazon.com/Where-Wizards-Stay-Up-Late/dp/0684832674