Mind the Gap - SIGCOMMconferences.sigcomm.org/sigcomm/2012/slides/... · Cambridge: Damon Wischik,...

Mind the Gap

Nick McKeownStanford University

Dean President

My problem

I get excited about improving the practice.

I am a networking researcher.

The Internet’s problem

Architecture: Seems impossible to change– Well-known shortcomings– Many great proposals– No way to demonstrate and test at scale– Huge installed base understandably reluctant to change

Industry: Strong vested interest in status quo– Very high margins as-is– Closed, proprietary solutions– Complex, brittle products– Innovator’s Dilemma

No path to change

My solution

I am done

I quit

Dean President

Why I love my job

I work with people much smarter than me.

I get to work on intellectually interesting ideas.

… that might positively change the practice.

Then, we try to actually change the practice.

Whatever it takes

• 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 tool

• …

The course of action

Is determined by the idea.

Is determined by what you want to change.

Some things that worked for me

Let me start with some motherhood and apple pie…

Look for blind spots

Question old school assumptions

Open your heart and mind to people who question assumptions

Question AssumptionsExample 1

“On the self-similar nature of Ethernet traffic”

Leland, Taqqu, Willinger, Wilson [Sigcomm ‘93+

Whether you agree with them or not – it got the whole community questioning assumptions.

Question AssumptionsExample 2

How big to make a backbone router buffer?

CB

Buffer Size

RTT ´C

RTT ´C

N

Th

rou

gh

pu

t

25,000

100%

2,500,00010Gb/s WAN

Number of packets

On-chip buffersSmaller designLower power

tW

ind

ow

Siz

e

Buffer

Buffer Size

RTT ´C

RTT ´C

N

Th

rou

gh

pu

t

log(W)

25,000~50

~ 90%

100%

2,500,00010Gb/s WAN

Number of packets

20 pkts

Integrated all-optical buffer [UCSB 2008]

On-chip buffersSmaller designLower power

People

Stanford: Guido Appenzeller, Neda Beheshti, Tim Roughgarden, Ashish Goel, Mihaela Enachescu

Toronto: Yashar Ganjali

Cambridge: Damon Wischik, Frank Kelly, Gaurav Raina

UMass: Don Towsley, Chris Hollot

UW Madison: Paul Barford, Joel Sommers

GA Tech: Amogh Dhamdhere, Constantine Dovrolis

UCSB: John Bowers, Emily Burmeister, Dan Blumenthal

Question AssumptionsExample 3

Myth 1: It is hard for switch/router hardware to maintain lots of queues.

– Since 1995, seen 10-15 ASICs do it easily

– Recently: 64x10Gb/s switch, 128k flows, 10% overhead.

Myth 2: You can’t build a large flow table– Next couple of years: Over 1Tb/s, 100,000s entries,

100s bits wide, several tables.

It doesn’t mean you have to; just says you can.

Some things that worked for me

Some surprising takeaways

Choosing research projects

• Pick a problem that is intellectually interesting.

• And improves the practice.

• And industry doesn’t like (yet).

PacketForwarding

PacketForwarding

PacketForwarding

PacketForwarding

PacketForwarding

Control

Control

Control

Control

Control

EthaneMartin Casado et al [Sigcomm ‘07]

Policy

“Laptops can’t accept incoming connections”

“A can’t talk to B”

Network Control Plane

Microsoft Security Group: “Come on in…”

Cisco Switching Group: “It will never work…”

Raw nerve.

We must be onto something.

Checklist

• Intellectually interesting.

• And improves the practice.

• And industry doesn’t like (yet).

Corollary: You can’t give stuff away

Example 1: Ethane (2007)

– Tried to give it away for free.

– Realized industry wasn’t ready to see it.

– Started Nicira.

Example 2: Network Memory (2001)

Example 3: Scalable Routers (1997)

Packet Buffer

10Gb/s10Gb/s

40 bytes every 32ns

Problem - Two memory operations every 32ns.- One DRAM operation about 50ns.- SRAM 75x price per bit, and 10x power.- 40Gb/s links coming; linecards with 16 x 10Gb/s links.- Thousands of queues.

DRAM

Network Memory

10Gb/s10Gb/s

Solution- SRAM cache size for zero misses: 2qblnq

- Practical for 160Gb/s linecard, 1024 queues.

SRAM SRAM

DRAM

Tails Heads

Blocks, bBlocks, b

1

q

ASIC

History

2001: Started on problem; industry discussion.

2003: Tried to give it away.

2004:

– 80% of world’s SRAM: Ethernet switches.

– Of those, 80% used by Cisco ($400M p.a.)

– Started Nemo.

2007:

– Savings estimate $150M p.a.

Checklist

• Intellectually interesting.

• And improves the practice.

• And industry doesn’t like (yet).

Corollary: You can’t give stuff away

Example 1: Ethane (2007)

Example 2: Network Memory (2001)

Example 3: Scalable Routers (1997)

1 2 3 4 5 6 7 8 9 10 111213141516

17 1819 20 2122 232425262728 29303132

2km

History

1997

– State of the art was 40Gb/s per rack.

– Tried to give it away.

– “Leave it to us”.

– Started Abrizio.

1999

– All outsourced solutions.

Checklist

• Intellectually interesting.

• And improves the practice.

• And industry doesn’t like (yet).

I put everything in public domain

Industry– Invests huge amounts to develop and sell products.

– Patents protect their ideas, giving them confidence to invest.

University research– Serves society at large.

– Stay ahead by running fast, not by protecting.

– Makes it easier to work with industry.

Everything in public domain since 1999.

Why I decided to come back

By 2005Many smart people calling for change

DaveClark

LarryPeterson

TomAnderson

Scott Shenker

JonTurner

PeterFreeman

GuruParulkar

Overcoming the Internet Impasse through Virtualization FIND, GENI, …

100x100, 4D

HuiZhang

SandyFraser

JenniferRexford

AlbertGreenberg

Vision

Software Defined Networks

Martin Casado

Scott ShenkerTeemu Koponen

Guru Parulkar+ many (brave) students

Vertically integrated

Closed, proprietary

Slow innovation

Small industry

Specialized

Operating

System

Specialized

Hardware

AppAppAppAppAppAppAppAppAppAppApp

Specialized

Applications

Horizontal

Open interfaces

Rapid innovation

Huge industry

Microprocessor

Open Interface

LinuxMacOS

Windows(OS)

or or

Open Interface

Vertically integrated

Closed, proprietary

Slow innovation

AppAppAppAppAppAppAppAppAppAppApp

Horizontal

Open interfaces

Rapid innovation

ControlPlane

ControlPlane

ControlPlane

or or

Open Interface

Specialized

Control

Plane

Specialized

Hardware

Specialized

Features

Merchant

Switching Chips

Open Interface

Checklist

• Intellectually interesting.

• And improves the practice.

• And industry doesn’t like (yet).

What’s next?

Making Networks Work

An intellectual framework for verifying, troubleshooting and

debugging SDNs

With SDN we can:1. Formally verify that our networks are

behaving correctly.

2. Identify bugs, then systematically track down their root cause.

• Ensuring correctness [Frenetic][HFT][Netcore]Nate Foster, Andrew Ferguson, Mike Freedman, Jen Rexford, Rob Harrison, Dave Walker, ++

• Software Fault Localization [W3]Scott Shenker, Colin Scott, Kyriakos Zarifis, Andreas Wundsam.

• Checking behavior [NICE]Marco Canini, Daniele Venzano, Peter Peresini, DejanKostic, Jen Rexford.

• Checking Invariants [VeriFlow]Ahmed Khurshid, Wenxuan Zhou, Matthew Caesar, P. Brighten Godfrey

• Consistent updatesMark Reitblatt, Rick McGeer, ++

• …

Scott Shenker at 1st ONS in 2011 “The Future of Networking and the Past of Protocols”

Software Defined Network (SDN)

Global Network View

Network Virtualization

PacketForwarding

PacketForwarding

PacketForwarding

PacketForwarding

PacketForwarding

Network OS

Abstract Network View

Control

Programs

f View( )Control

Programs

f View( )Control

Programs

f View( )

Software Defined Network (SDN)

Global Network View

Network Virtualization

PacketForwarding

PacketForwarding

PacketForwarding

PacketForwarding

Abstract Network View

Control

Programs

f View( )Control

Programs

f View( )Control

Programs

f View( )firewall.c

…

if( pkt->tcp->dport == 22)

dropPacket(pkt);

…

PacketForwarding

Network OS1. <Match, Action>2. <Match, Action>3. <Match, Action>4. <Match, Action>5. <Match, Action>6. …7. …

1. <Match, Action>2. <Match, Action>3. <Match, Action>4. <Match, Action>5. <Match, Action>6. …7. …

1. <Match, Action>2. <Match, Action>3. <Match, Action>4. <Match, Action>5. <Match, Action>6. …7. …

1. <Match, Action>2. <Match, Action>3. <Match, Action>4. <Match, Action>5. <Match, Action>6. …7. …1. <Match, Action>

2. <Match, Action>3. <Match, Action>4. <Match, Action>5. <Match, Action>6. …7. …

How do other industries do it?

Making ASICs Work

$10B tool businesssupports a

$250B chip industry

Specification

Functional Description (RTL)

Testbench & Vectors

Functional Verification

Logical Synthesis

Static Timing

Place & Route

Design Rule Checking (DRC)

Layout vsSchematic (LVS)

Layout Parasitic Extraction (LPE)

Manufacture& Validate

100s of Books

>10,000 Papers

10s of Classes

Making Software Work

Static Code Analysis

Invariant Checker

Interactive Debugger

Model Checking

Run-time Checker

Specification

TestbenchFunctional Description

(Code)

$10B tool businesssupports a

$300B S/W industry

100s of Books

>100,000 Papers

10s of Classes

Making Networks Work (Today)

traceroute, ping, tcpdump, SNMP, Netflow

…. er, that’s about it.

Why debugging networks is hard

Complex interaction – Between multiple protocols on a switch/router.

– Between state on different switches/routers.

Multiple uncoordinated writers of state.

Operators can’t…– Observe all state.

– Control all state.

Networks are kept working by

“Masters of Complexity” A handful of books

Almost no papers

No classes

Philosophy of Making Networks Work

YoYo“You’re On Your Own”

Yo-Yo Ma“You’re On Your Own, Mate”

With SDN we can:1. Formally verify that our networks are

behaving correctly.

2. Identify bugs, then systematically track down their root cause.

Software Defined Network (SDN)

Global Network View

Network Virtualization

PacketForwarding

PacketForwarding

PacketForwarding

PacketForwarding

Abstract Network View

Control

Programs

Control

Programs

Control

Programs

firewall.c

…

if( pkt->tcp->dport == 22)

dropPacket(pkt);

…

PacketForwarding

Network OS

1. <Match, Action>2. <Match, Action>3. <Match, Action>4. <Match, Action>5. <Match, Action> 6. …7. …

Three of our projects

1. Static Checking [HSA]“Independently checking correctness”

2. Automatic Testing [ATPG]“Is the datapath behaving correctly?”

3. Interactive Debugging [ndb]“Finding bugs, and their root cause, in an operational network”

Peyman Kazemian

Hongyi ‘James’

Zeng

GeorgeVarghese(UCSD)

NikhilHandigol

BrandonHeller

VimalJeyakumar

DavidMazières

1. Static checkingIndependently checking correctness

Motivations

In today’s networks, simple questions are hard to answer:

– Can host A talk to host B?

– What are all the packet headers from A that can reach B?

– Are there any loops in the network?

– Is Group X provably isolated from Group Y?

– What happens if I remove a line in the config file?

57

Software Defined Network (SDN)

Global Network View

Network Virtualization

PacketForwarding

PacketForwarding

PacketForwarding

PacketForwarding

Abstract Network View

Control

Programs

Control

Programs

Control

Programs

PacketForwarding

Network OS1. <Match, Action>2. <Match, Action>3. <Match, Action>4. <Match, Action>5. <Match, Action> 6. …7. …

1. <Match, Action>2. <Match, Action>3. <Match, Action>4. <Match, Action>5. <Match, Action> 6. …7. …

1. <Match, Action>2. <Match, Action>3. <Match, Action>4. <Match, Action>5. <Match, Action> 6. …7. …

1. <Match, Action>2. <Match, Action>3. <Match, Action>4. <Match, Action>5. <Match, Action> 6. …7. …

1. <Match, Action>2. <Match, Action>3. <Match, Action>4. <Match, Action>5. <Match, Action> 6. …7. …

1. <Match, Action>2. <Match, Action>3. <Match, Action>4. <Match, Action>5. <Match, Action> 6. …7. …

1. <Match, Action>2. <Match, Action>3. <Match, Action>4. <Match, Action>5. <Match, Action> 6. …7. …

1. <Match, Action>2. <Match, Action>3. <Match, Action>4. <Match, Action>5. <Match, Action> 6. …7. …

1. <Match, Action>2. <Match, Action>3. <Match, Action>4. <Match, Action>5. <Match, Action> 6. …7. …

Static Checker

1. <Match, Action>2. <Match, Action>3. <Match, Action>4. <Match, Action>5. <Match, Action> 6. …7. …

1. <Match, Action>2. <Match, Action>3. <Match, Action>4. <Match, Action>5. <Match, Action> 6. …7. …

1. <Match, Action>2. <Match, Action>3. <Match, Action>4. <Match, Action>5. <Match, Action> 6. …7. …

1. <Match, Action>2. <Match, Action>3. <Match, Action>4. <Match, Action>5. <Match, Action> 6. …7. …

1. <Match, Action>2. <Match, Action>3. <Match, Action>4. <Match, Action>5. <Match, Action> 6. …7. …

“A can talk to B”

“Guests can’t reach PatientRecords”

Policy

How it works

Header Space Analysis

Header Space Analysis

12

34

1

2

3

4

Header Space Analysis

12

34

1

2

3

4

Port ID

Can A talk to B?

12

34

1

2

3

4

Port ID

All packets from A that can reach B

Header Space Analysis*NSDI ‘12+

Consequences– Abstract forwarding model; protocol independent

– Finds all packets from A that can reach B

– Find loops, regardless of protocol or layer

– Can prove that two groups are isolated

Can verify if network adheres to policy

Software

Hassel tool

– Reads Cisco IOS Configuration

– Checks reachability, loops and isolation

– C: 60ms for Stanford Backbone

– Python: 10 mins for Stanford Backbone

Code– http://bitbucket.org/peymank/hassel-public

Checklist

• Intellectually interesting.

• And improves the practice.

• And industry doesn’t like (yet).

In Summary…

What has worked for me

• There doesn’t need to be conflict between– An intellectually interesting idea, and

– Improving the practice.

• Plenty of research problems can do both.

• Improving the practice means going beyond writing papers. Whatever it takes.

• Listen to industry, but not too much.

• Give ideas away.

Software Defined Networks

• Allows a stronger intellectual foundation to networking

• Allows us to define the right abstractions

• Will allow us to transfer technology much faster, in both directions

• Is already closing the gap with industry

Almost the end…

Something that worries me

Internet: biggest societal change in our time.

• Our field is growing in size and scope.

• More researchers, more topics, more developers.

• Yet the Sigcomm tent remains very small.

Sigcomm

• Too far removed from the practice.

• Too conservative.

• Rule of thumb:

– If the area is new, Sigcomm won’t accept it.

• Bad for students and young professors.

Sigcomm attendees

88% 87% 91% 86%

12% 13% 9% 14%

0%

50%

100%

2008 2009 2010 2011

Industry, not research

Research

( Attendees Industry% Papers )

Sigcomm( 500 15% 35 )

Siggraph( 13,000 60% 115 )

Sigcomm 2015 ? ( 2,000 40% 70 )

Sigcomm

Let’s make the tent bigger

– More topics

– More papers

– More industry participation

The End

Giving the network a wedgie

Nick McKeownStanford University

Specialized

Control

Plane

Specialized

Hardware

Specialized

Features

Apps

Specialized

Control

Plane

Specialized

Hardware

Apps

Network OS

Simpler

Hardware