I’ve watched through Nick McKeown’s keynote at SIGCOMM conference in Helsinki. It’s not boring and I made the following notes and links to the relevant places.
Nick McKeown – ACM SigComm12 Keynote
The Internet Problem
- plenty of shortcoming in current protocols
- plenty of available research
- installed base reluctant to change
- strong vested interest in status quo
- high margin business resists change
- closed proprietary solution,
- complex brittle products
Love his job because
- * works on intellectually interesting ideas
- * might positively change practice
- * Then try to actually change the practice by what ever method – prove a theorem, write a paper, build a demo, talk to lots of industry people, write a standard, give lots of talks, write a blog, start a company, build an open source whatever it takes
If you want to change the practice writing a paper is not enough, you might have to several of these to make things change.
Some things that worked for me
- Look for blind spots – turn over stones that have been left unturned. The manner matters but aggression
- Questions assumptions!!!
For example – How big to make a backbone router buffer ? * “The size of a router is limited by three things – the memory bandwidth, the memory bandwidth and the memory bandwidth.” * Ask five people and got give mutually exclusive answers. * Then a careful analysis of what buffers need to look like including the impact of deep buffers on power and chip performance.
Myth 1 – It’s hard for a switch/router hardware to maintain lots of queues
Today 64 x 10Gbps switch with128K flows at 8% CPU overhead.
Myth 2 – You can’t build a large flow table. In design today, 1Tbps with 100,000s of entries, hundreds of bits wide with multiple/several tables. The hard part is the management of the state, not the management of queues or flows.
TCAM designers/industry are way ahead of our requirements. Waiting for us to catch up and present the need/demand to buy new TCAM solutions.
Somethings That Worked For Me
- Pick a problems that’s intellectually interesting
- improve the practice
- pick a problem that industry doesn’t like(my emphasis)
Talk to chip manufacturers, if they liked it then he would cancel he project. If it’s interesting and exciting to vendors then they will implement it themselves. If they hate it, then it’s worth doing.
Cisco Switching Group in 1997 rejected the premise of flow networking. In fact “yelling” and hit a raw nerve therefore might be onto something.
Was it a stupid idea, or was it a new idea ?
10Gigabit per second buffer are unique because the arrival time of the smallest Ethernet frame is less than the access time of DRAM http://www.youtube.com/watch?v=Ho239zpKMwQ&t=24m30s
Like mainframes where the arrival of the microprocessor led away from a closed stack to an open platform, the merchant switch silicon moves away fron closed network stacks of proprietary hardware, specialized control planes and specialised features.
The focus now shifts to using software to control and program networks. Starts by focussing on software quality and software engineer.
Making networks work
Why debugging networks is hard Complex interactions
- - multiple protocols on a switch/rtouers
- - between state on different switches and routers.
- - Multiple uncoordinated writers of state
- - Operators can’t observe or control all state in the network.
- - Networks are kept working by Masters of Complexity in the operational arena.
- - no verification of network designs and architectures.
- - Networks are kept working by “Masters of Complexity”
- - Phenomonal people who can keep networks working.
Debugging networks is Yo Yo Ma “You are on your own Mate”
SDN will: – formally verify that networks are behaving correctly – identify bugs then systematically track down their root cause.
Then he walks briefly through an interpretation of the SDN and it’s elemental pieces.
Then leads in to the concept of static network checking with Head Space Analysis.
Header Space Analysis
In todays networks, simple questions are hard to answer: – can host A talk to host N – What are all the packet header from A that can reach B ? – Are there any loops in the network ? – can Group X be proven to be isoloated from Group Y ? – What happens when I change this configuration, are the conditions met ?
can build an external application that looks at the configuration of the network ?
We can test conditions such as “Can Workstations reach the Web Server for applications ?
Then he walks through a brief tutorial of header space analysis and how it works.
I wonder what would happen with firewalls or load balancers ? Can HSA handle this ?
Improves the practice of network implementation.
Through the practice of Software Defined Networks: * Allow a stronger intellectual foundation to networking * Allow us to define the right abstractions * Transfer technology from research to vendor faster in both directions.
Great quote at the end:
We tend to be too brutal on new ideas and too generous to old assumptions, and . We need to be generous to new ideas and more brutal with old ideas. Lets question the foundations and old ideas but lets embrace new ideas and don’t dismiss them.