Transcript HuiRetreatFall04

Overview and Status Update of
Hui Zhang
Carnegie Mellon
1
Slides Used In First Retreat

Vision:

100 Mbps to 100 million households

1Gbps to 1 million business

Approach: clean slate, design from first principles

Holistic and inter-disciplinary design

Consider the network as a whole
– Access, metro, core regions
– Service and physical transport layers
2

Consider technology trends for scaling, cost, future-safeness

Archiect with explicit considerations of economic, dependability,
security, mult-service

Design with explict goals of enabling tractable analysis and
modeling
What is the 100x100 Project?

Elevator Speech

What is 100x100 project up to?

What is the “meat”?
– What are the key technical challenges?
– What are the key insights and expected contributions?

Is 100x100 project just about speed?

Why 100 Mbps?
– It is too slow, it is too fast

3
Why cannot IP do 100x100?
Clean Slate Design

Why is Clean State Design necessary?


Is Clean State Design useful?


4
Have you established that incremental evolution is
impossible?
Any hope of introducing a “revolutionary” solution?
Mainstream views

incremental fixing of Internet is sufficient

radical new solutions have little chance of being deployed
Context


6
IP is a great success because its architects set the
right goal: global best-effort reachability

Global addressing scheme

Internetworking architecture

Simple best-effort service
Success is a double-edged sword  the world
demands more from IP and the Internet

Convergence telecommunication infrastructure that
provides 7x24x365 service

Dependability, scale, security, economic sustainability
Convergence Vision
VoIP + Data + Video over IP
POP Long-- haul DWDM backbone
(OC48/192)
7
Have We Already Achieved Convergence?
Subscriber/
Metro Access
Metro Core
Long Haul
Service Node/ASP
Enterprise
Server
End Office/
Collocation
Server
Router
ISP
Voice
Switch
Backbone
Router
Metro Hub
Office
Router
Voice
Switch
Voice
Switch
ATM
LAN
RF
Cable
ACCESS
INTEROFFICE
G(SONET)
Wireless
Copper
Fiber
HAN
8
OXC
l
l
Services
Transport
INTERCITY
G(l)
l
l
IP Robustness?
UUNet 10/03/02 Outage
as seen from AT&T
9
State of the Art is lacking in robustness on every time scale: days, hours,
minutes, seconds
Molasses
Shell scripts
Management Plane
Tomography
•
Planning tools
Databases
•
Figure out what is happening in
network
Decide how to change it
Configs SNMP
rancid modems
OSPF
Control Plane
• Multiple routing processes on each
Link
Packet
router
OSPF
metrics
filters
• Each router with different
BGP
•
OSPF
BGP
FIB
OSPF
BGP
configuration program
Huge number of control knobs:
metrics, ACLs, policy
Data Plane
FIB
10
FIB

Distributed routers forwarding
packets

Based on FIB or labels
A Study of Operational Production Networks
(Joint Between ATT and CMU)
11

Obtained anonymized configuration files for 31 active
networks (>8,000 configuration files)

6 Tier-1 and Tier-2 Internet backbone networks

25 enterprise networks

Sizes between 10 and 1,200 routers
Configuration State for One Network
12
Example Router Configuration File
13
Complex Interaction of States
Management
Plane
Data Plane
Packet
Filters
FIBs
Control
Plane
Legends
Hardwired State
Configuration State
Dynamic State
State Dependency
14
Molasses
Shell scripts
Tomography
Planning tools
Link
metrics
OSPF
BGP
15
OSPF
BGP
Databases
Packet
filters
OSPF
BGP
Management Plane
• Figure out what is
happening in network
• Decide how to change it
Control Plane
• Multiple routing processes
on each router
• Each router with different
configuration program
• Huge number of control
knobs: metrics, ACLs, policy

Data Plane

Distributed routers forwarding
packets

Based on FIB or labels
Systems of Systems

Systems are designed as components to be used in larger
systems in different contexts, for different purposes,
interacting with different components



16
Example: OSPF and BGP are complex systems in its own right,
they are components in a routing system of a network,
interacting with each other and packet filters, interacting with
management tools …
Complex configuration to enable flexibility

The glue has tremendous impact on network performance

State of art: multiple interactive distributed programs written in
assembly language
Lack of intellectual framework to understand global
behavior
Development of the Elevator Pitch

Best-effort service model is a critical reason for
Internet’s success

Success is a double-edged sword

17
Expectation: with true two-way broadband connectivity to every
household (100x100), IP be convergence telecommunication
infrastructure that provides 7x24x365 service

Far from being dependable, scaleable, secure, economic
sustainable

Even when providing best-effort service, it is already
extremely complex
What is the Solution?
18

Simplify,

Simplify,

simplify!

But based on what?
Observations that Lead to Simplifications


Tremendous focus on protocols/protocol architecture,
should also focus on

networks and network architecture

Interplay between network architecture and protocol architecture
Architecture agnostic to technology trend


Box/Switch/Router centric distributed control has
fundamental limitations, should also explore

19
Should leverage technology trends that enable simple network
architecture
Control architecture that can implement a wide range of networkwide policies
Protocol Architecture vs. Network Architecture

Protocols: TCP, IP, BGP, OSPF

They work in arbitrary network, but do not work as
well in any network

Dependable network requires
– Structured network, and
– protocols that can take advantage of network architecture
20
Structured Access/Metro Networks
21
Structured
Backbone Networks
22
Structured Network Interconnection
Backbone Network
Access Routers
CPE Routers
23
Observation One

24
Structured network + protocols taking
advantage of network structures achieve

Dependable, simple, and understandable network

Dependable, simple, and understandable
protocols
Observations that Lead to Simplifications


Tremendous focus on protocols/protocol architecture,
should also focus on

networks and network architecture

Interplay between network architecture and protocol architecture
Architecture agnostic to technology trend


Box/Switch/Router centric distributed control has
fundamental limitations, should also explore

25
Should leverage technology trends that enable simple network
architecture
Control architecture that can implement a wide range of networkwide policies
Key Technology Trends

26
Trend 1: Deep fiber deployment key to achieving
ubiquitous, high capacity connectivity to home

Wireless: important complementary technology

Copper & cable can also be used for last 1000 feet

Fiber enables not only scalable high speed, but also
longer distance

Trend 2: low cost, low energy, high capacity, autoconfigured, environmentally hardened access packet
switches

Trend 3: ultra high capacity (perabit) backbone switches
Implications for Network Architecture

Deep fiber + access switch  large scale
packet access networks


27
In contrast, traditional access network size limited
by copper transmission distance
Large scale access network + ultra high
speed backbone switch  architecture with

regional node that terminates access network and,

backbone network with a smaller number of richly
connected switches
An Example Structured 100x100 Network
28
Observations that Lead to Simplifications


Tremendous focus on protocols/protocol architecture,
should also focus on

networks and network architecture

Interplay between network architecture and protocol architecture
Architecture agnostic to technology trend


Box/Switch/Router centric distributed control has
fundamental limitations, should also explore

29
Should leverage technology trends that enable simple network
architecture
Control architecture that can implement a wide range of networkwide policies
Robust, Simple Control and Management

Network is about coordination of switches

Distributed state management

Multiple goals
– Reachability
– Policy control
– Security
– Resiliency
– Traffic Engineering, load balancing
– VPN

30

Multiple layers (switching + optical)

Diverse switching primitives (OXC, label switches, IP switches, vLAN
switches)
Status quo of control and management: extreme
complex, non-linear and fragile
Good Abstractions Reduce Complexity
Management
Plane
Control
Plane
Data Plane
31
Configs
FIBs
Decision
Plane
FIBs
Dissemination
Data Plane

All decision making logic lifted out of control plane

Eliminates duplicate logic in management plane

Dissemination plane provides a control channel to/from
data plane
Development of the Elevator Pitch

Internet technology is far from being adequate to support 100x100 vision

Goal: dependable, scaleable, secure, and economic sustainable
telecommunication infrastructure

It is already extremely complex

Solution: simplify, simply, simply!


32
Simplicity should also mean that the design is understandable
Three key ideas that may lead to simplification

Take advantage of structured network

Leverage and develop technologies that enable structured networks

Design simple and powerful network-wide control
abstractions/mechanisms
Why 100x100? Why not 1000x100?

Key is to focus us on an end goal of a network

33
Principles derived can be general
Why Clean State Design?

A powerful research methodology that helps to crystallize the issues


Many good examples in systems research that take one idea to extreme:
RISC, SmallTalk, NFS, IP
A mind set that may result in different research, e.g.

Incremental approach to security
– How to detect and stop Blaster, Code Red?

Clean state design approach
– What would be the fundamental capability of a strategic adversary?
– What are the fundamental limitations/possibilities of any network-based or hostbased security mechanism?
– What should be the minimal & necessary set of layer 3 security mechanism?
34

A concrete and complete different design point highlights possibilities

Understanding the target first helps to plan the trajectory of evolution
Big Bets Research

We make big bets to get big breakthroughs

Visionary ideas carrying intellectual risk

Can’t predict outcomes in advance

The Christopher Columbus Effect
Randy Bryant: Dean of SCS, CMU
“Strategic Vision for CS in CMU”
35
Can We Make a Difference?


36
Monopoly positions in all technology areas

Microsoft in OS

Cisco in router

Intel in processor

Oracle in database
People are usually

too optimistic in prediction of two years out, but

too pessimistic in prediction of five or ten years out
Can We Change IP and the Internet?

37
What is IP anyway?

Service interface: services seen by VoIP, Web

Data plane (IPv4, IPv6 packet formats)

Control plane (OSPF, ISIS, BGP, LDP)

Management plane
Learning from Ethernet Evolution Experience
Ethernet or 802.3
Early Implementations
•Bus-based Local Area Network
WAN
•Collision Domain, CSMA/CD
B/R
•Bridges and Repeaters for distance/capacity extension
•1-10Mbps: coax, twisted pair (10BaseT)
Current Implementations:
LAN
Everything Changed Except Name and Framing
Router
Ethernet
Conc..
HUB
•Switched solution
•Little use for collision domains
•80% of traffic leaves the LAN
Switch
•Servers, routers 10 x station speed
Server
38
WAN
•10/100/1000 Mbps, 10gig coming: Copper, Fiber
How To Evolve?

One possible path to future

Fix packet format: IP or IPv6 or Ethernet or MPLS frame
formats

Evolve switch-switch protocols (NNI)
– control/management plane, where the intelligence is
39
ATM Net
Telephone
Network
40
X.25 Net
ATM Net
Telephone
Network
41
X.25 Net
IP
ATM Net
Telephone
Network
42
X.25 Net
IP
IP
ATM
Telep
X.25
43
ATM
Telep
44
X.25
IP