Session 1: Router Hardware

Rethinking Packet Forwarding Hardware 
By Martin Casado

- Key points
  - Design goals:  software router vs. hardware router
    - Customer hardware -> price dropping down, but not flexible
    - Software router: flexible but slow
    - Network processors/FPGAs: hard to program and limited port density

- Proposed approach:
  - Software: all forwarding alg.
  - Hardware: lookup table (TCAM), making cache forwarding
  - pkt fields, dependent state, output action
  - revoke flow if state changes
  - tunnel works in this model
  - Q: scalability? - require miss rate < 2%
  - Results show L2 forwarding hit rates close to 100%
                 L3 forwarding hit rates: modified FW > traditional 
			when cache sizes increase from 1k - 32k
  - Q: can this work today?
  - Open issues
  - Implications
    - still consistent with today's arch
    - run multiple alg.
    - use QoS/cache replacement for isolation

- Questions:
  Q: Comments on those anti-caching opinions
	- caching forwarding good at closed environment
        - not for backbone.
  Q: Shared cache?
  Q: How to support programmer to pass from one application pattern
	to another?
	- not a problem
  Q: How to set up key for cache
	TCAM has the flexible to cache attribute defined by SW
  Q: Comments on adv. over pure HW solutions
	caching scheme can't handle if states changes all the time.


API Design Challenges for Open Router Platforms on Proprietary Hardward
By Jeff Mogul

- Hardware router -- how to make it easy
- Answer following questions:

Q1: Why open router platforms?
	allow 3rd parties to extend.

Q2: What HW features are interesting?

Q3: What Software runs on Open-Router
	- layering in an O-R
	- examples of switchlets (sp. firewalls, sp. monitoring, 
		NAT, dyn. VLAN, OpenFlow, Click)
	- categories of switchlets (per-pkt, per-flow, control-plane func.)

Q4: Design challenges?
	- Resource mgmt.
	- controlled sharing
	- isolation bw switchlets.
	- portability
	- manageability

- What proposed -- API like tcamAddRow. etc.
- OpenFlow example and problems.

- Questions:
  Q: In long term, evolution path like 3D graphics?
  Q: Router vendors, whether is a trend to Open-Router.
  Q: Resource mgmt?
	- Refer to another paper
	- don't need to go too high up

Final Discussions on following topics:

- TCAM is costy, high power consumption, big, etc.
 	comments on where TCAM fits in?
	TCAM can be cheap.
	can parallel use to increase functionality
- In terms of control processing, re-writing SW, 
	what are the overheads?
- For quick updates, need to reorder the router hash table,
	a paper on alg. for organizing TCAM table,
	but there are tradeoffs.
- There are a range of things, what actually support?
- What have been done on consistency?