Transcript intro-scalability

Scalability is King
1
Internet: Scalability Rules
Scalability is : a critical factor in every decision
Ease of deployment and interconnection
The intelligence is in the outskirts of the network
2
Internet: Design Goals
The goal: to interconnect multiple existing
networks and technologies
• Packet switching over circuit switching
• Reasons:
 the applications were suited for this: rlogin
 Packet switching was well understood
 More flexible, less committal
“The design philosophy of the DARPA Internet
protocol”, SIGCOMM 88
3
Internet Design: Secondary Level Goals
Fault-tolerance (DARPA is the army)
Support multiple communication services
Accommodate various technologies & networks
Allow distributed management
Enable cost effective resource
Ease of interconnection: adding a host
Resources used in the Internet must be
accountable (less attention)
4
Fault Tolerance: The Internet Premise:
One Robust Connected Network
Size of
Largest
Connected
Component
Random
Highest Degree first
order
Highest Significance
first order
4000
3000
2000
1000
2988
2656
2324
1992
1660
1328
996
664
332
0
0
#Deleted nodes
I t e r a t i ons
Robust to random, sensitive to focused failures
The network tends to stay as one connected component
5
I. State: The Enemy of Scalability
Routing state: the information that a router
needs to keep in order to route
Large routing state is bad
• Consumes memory
• Makes look-up slow
State = information = intelligence
The trade-off: intelligent vs scalable behavior
6
I. An Interesting Antithesis
Telephone network:
• End devices are stupid
• Core is intelligent
Internet:
• End devices are intelligent
• Core is stupid (forward
packets, no guarantees)
Telephone network:
• Circuit switching
• Ultra reliable, guarantee
quality(real-time)
• Centralized control
• Difficult to add new users
• Cumbersome to new apps
Internet:
•
•
•
•
•
Packet switching
Unreliable: no guarantees
Decentralized
Easy to add new users
Easy to add new applications
7
I. Internet: Routing State
What does a router know?
Routing table:
• IP prefix -> outgoing link
What is an IP prefix?
Scalability:
• IP prefixes aggregate many IPs to one entry
• Routers do not* keep per connection information
* Some exceptions exist.
8
I. State Per Flow: pros and cons
What would state per flow give us?
Differentiate our behavior per flow
• Provide QoS
Be fair: share resources accordingly
Monitor and measure: accountability - costing
Ensure that packets follow the same path
9
II. Centralized vs Decentralized Routing
Centralized
•
•
•
•
All information exists in one place
Decisions are taken there
The decision involves the whole path
OSPF: all routes have all the info:
Decentralized
• Each router makes a decision in isolation
• Each router knows partial/aggregated information
• RIP, BGP are some protocols
10
III. Dynamic vs. Static Routing
Routing is an optimization problem
What are we trying to optimize?
Static metrics (non time varying)
• Number of hops (hopcount)
• Link capacity
• Buffer space
Dynamic metrics (time varying)
•
•
•
•
End2end (e2e) Delay
Link delay
Link utilization
Available buffer space
11
III. Dynamic or Static Routing?
Static routing
• Stable
• Arguably: utilizes resources well in the long run
Dynamic routing
• Flexible - adaptable
• Better performance (short term at least)
12
IV. QoS Aware Routing
Quality of Service is tightly coupled with
Reservations
• I need 10Mb/s with at most 100msec e2e delay
Problems:
• Find a suitable path
 Centrally easier,
 Distributed: start reserving, then backtrack
• Reserving resources along the path
• Protecting resources from future flows
13
For next time
Refresh your memory: Kurose - Ross
• Textbook: chapter 1: intro
• Textbook: chapter 4: Network Layer and routing
Read: Clark’s sigcomm ‘88 paper
Case study: How stable is Internet routing:
• Run traceroute (linux) and see what happens
14