Transporting Packets from End-to
Download
Report
Transcript Transporting Packets from End-to
Seminar on
“Clean Slate Design for the Internet”
Nick McKeown
[email protected]
1
High level
2
“Given what we know today, if we were to
start over with a Clean Slate, how would we
design a global communications network?”
“Ideally, how will the network look in 15-20
years, and how will we get there from here?”
Prelims
3
What’s wrong with the Internet…?
Why is the research and business community
not already solving it?
What are other groups doing?
What we plan to do at Stanford
An example of “Clean Slate” design
Original Architecture
4
A dumb connectionless packet-forwarding packetswitched infrastructure, with high-level functionality
at the edge
Single, simple lowest-common denominator data
delivery service (IP), with reliable stream service built
on top
Fixed-size numerical addresses with {network, host}
hierarchy; one per physical network interface
Later
Separation of IP and TCP (including congestion control
using packet loss as congestion signal)
Subnetting, autonomous systems (EGPs and IGPs), DNS,
CIDR
What is needed
Wouldn’t we like a network that we can trust to be
always there, always on, easy to use, universally
accessible, secure, and economically viable.
David Cheriton’s example: If the FAA carried all of its
traffic over the public Internet, you'd be nuts to fly.
Some obvious desirable characteristics
5
Robustness and Availability
Security
Naming and Addressing: accountability vs anonymity
Predictability
Mobility
Economic Viability
What else?
Prelims
6
What’s wrong with the Internet…?
Why is the research and business community
not already solving it?
What are other groups doing?
What we plan to do at Stanford
An example of “Clean Slate” design
Prelims
7
What’s wrong with the Internet…?
Why is the research and business community
not already solving it?
What are other groups doing?
What we plan to do at Stanford
An example of “Clean Slate” design
What are others doing?
8
Background
Incrementalism and “victim of success” of Internet
New era of more radical and fundamental thinking
about the future of networks and communications
New-arch (MIT)
100x100 (CMU)
Geni (NSF/Gov)
New-arch (2000)
9
Requirements for new network
Mobility: Highly dynamic and efficient
Policy-driven auto-configuration
Highly time-variable resources
Allocation of capacity
http://www.isi.edu/newarch/
100x100 (CMU/Stanford/Rice)
NSF Large ITR (2003-2008)
Questions:
10
Can structure be used to make networks more robust,
predictable and manageable?
What economic principles drive the operation of access and
backbone networks?
What security primitives must be built into the network?
Can/should network and protocol architectures be designed
to take advantage of long-term technology trends?
http://100x100network.org/
NSF Geni Initiative (2005)
11
CISE major effort, seeking congressional funding of
approx $300M starting 2008
Two parts: Research program; Global experimental
facility to explore new architectures
Areas of interest:
Creating new core functionality, including naming,
addressing, identity, management.
Developing enhanced capabilities: building security intot he
architecture; design for high availability;
privacy/accountability; design for regional differences and
local values
Deploying and validating new architectures
Building higher-level service abstractions
Building new services and applications
Developing new network architecture theories
Prelims
12
What’s wrong with the Internet…?
Why is the research and business community
not already solving it?
What are other groups doing?
What we plan to do at Stanford
An example of “Clean Slate” design
Prelims
13
What’s wrong with the Internet…?
Why is the research and business community
not already solving it?
What are other groups doing?
What we plan to do at Stanford
An example of “Clean Slate” design
What we plan to do at Stanford
Weekly Seminar in Fall and Winter
14
Fall: Talk by professor followed by discussion
Goals
To get thinking about the problem
To learn from each other
To identify some collaborative research projects
Prelims
What’s wrong with the Internet…?
Why is the research and business community
not already solving it?
What are other groups doing?
What we plan to do at Stanford
An example of “Clean Slate” design
How to design backbone networks from a clean slate?
15
Backbone Networks: Emerging Structure
10-50 routing centers interconnected by long-haul optical
links
Increasingly rich topology for robustness and loadbalancing
Typical utilization < 25%, because
16
Uncertainty of traffic matrix network is designed for
Headroom for future growth
Headroom to carry traffic when links and routers fail
Minimize congestion and delay variation
Efficiency sacrificed for robustness and low queueing delay
How flexible are networks today?
What fraction of allowable traffic matrices can they support?
Abilene
25% Over Prov: 0.025%
50% Over Prov: 0.66%
AT&T
25% Over Prov: 0.0006%
50% Over Prov: 0.15%
17
Verio
25% Over Prov: 0.0004%
50% Over Prov: 1.15%
Sprint
25% Over Prov: 0.0003%
50% Over Prov: 0.06%
Verio, AT&T and Sprint topologies are from RocketFuel
Desired Characteristics
Robust
Recovers quickly; continues to operate under failure
Flexible
Will support broad class of applications, new customers,
and traffic patterns
Predictable
Can predict how it will perform, with and without failures
Efficient
Does not sacrifice cost for robustness
18
Backbone Design
19
Assume underlying reliable mesh of physical
circuits
1.
Dynamic circuit switching over underlying mesh,
or
2.
Load-balanced logical network.
Describing today
Approach
Assume we know/estimate traffic entering
and leaving each Regional Network
Use Valiant Load Balancing (VLB) over whole
network
20
Requires only local knowledge of users and
market estimates
Enables support of all traffic matrices
Valiant Load-Balancing
2r1r2 /rN
r1
1
rN
r2
2
3
N
…
4
r4
Capacity provisioned over existing
robust mesh of physical circuits
21
r3
A Predictable Backbone Network
22
Performance: 100% throughput for any valid traffic
matrix.
Only need to know aggregate node traffic.
Under low load, no need to spread traffic.
Robustness
Upon failure, spread over working paths
Small cost to recover from k failures: Provision approx 2rirj/r(N-k)
Simple routing algorithm
Efficient
VLB is lowest cost method to support all traffic matrices
Similar cost, while supporting significantly more traffic matrices.
How expensive would VLB be?
Cost normalized to VLB routing.
Cost of switching = cost of transmission for 370miles
Abilene
25% Over Prov: 0.026% Cost: 0.87
50% Over Prov: 0.66% Cost: 1.04
AT&T
25% Over Prov: 0.0004% Cost: 0.94
50% Over Prov: 0.14%
Cost: 1.12
23
Verio
25% Over Prov: 0.0003% Cost: 0.99
50% Over Prov: 1.08%
Cost: 1.19
Sprint
25% Over Prov: 0.0002% Cost: 0.86
50% Over Prov: 0.04%
Cost: 1.04
Open questions
24
Worst case propagation delay doubled
Low variance in delay
There are “express paths”
(How) are multiple VLB networks connected,
and how does performance change?
Economics and policy: how do operators
compete?