CSE 561: Computer Communication and Networks

Class Notes: Oct 16, 2002

 

Administration

Tom: Sure, today after class, or after 4:45pm, not tomorrow.

  Monday all afternoon after class.

                       Janet: email for appointment.

Yes, by all means.  "The key is not to prevent people from stealing  your ideas; it is to get them to steal your ideas"

 

How to run an ISP, part II -- Interdomain Routing

 

evaluation metrics for routing:

- reliability :  make sure that everybody find the same route.

- scalability : no centralized point of control.

- cost

- performance (related to revenue)

Baseline approach -- SPF:

- concentrates on reliability

- ignores performance

 

 

BGP started in the SPF world:

-          In BGP an entire organization == one node; shortest path in terms of the number of ASs crossed

-          Initial design ignored cost and performance

 

 

So how do we add performance to the Interdomain Routing?

-          Use link weights to approximate performance

-          For the sake of scalability, link weights are adjusted locally

-          System is dynamically unstable

 

How is it done in Intradomain Routing?

-          The link weights are usually set centrally by humans on basis of trafic flows and link capacities;

-          The problem is NP hard.

-          It does not allow for splitting traffic among different routes unless paths have equal weights (but if the weights are equal, both paths get equal shares of the traffic). 

-          To address the problem: using circuits/ fractional routing. Make the network more regular (like tree).

 

 

 

Shortest AS routing:

-          Might be reliable

-          Might not be optimal

 

From the point of view of a single ISP, it has:

-          Transit relationships -- buy a way to reach the rest of the Internet

-          Peering -- free exchange traffic

-          Customers -- people who buy transit from it

 

-          Who to buy transit from

-          Which ASs to use for outgoing traffic

§         This is done in a straightforward way based on the IP prefix; regional carriers may price quite low for transit to their customers but quite a lot for the rest of the Internet.

§         General rule: first try to send to customers, then to peers, then on transit links; 

§         You have to worry a bit about performance, too... (some cheap links are not as good)

 

-          Which links to use to send traffic to a given AS

§         Choose the closest (early exit)

-          How to influence incoming traffic?

§         Pricing;

§         Prefix filtering (choose which routes to advertise; e.g. do not advertise yourself to your peers as a connection to locations that you yourself have to pay for)

§         AS link padding (e.g. I can advertise different entry points with different links weights). Reliability is a problem.

§         MEDs (Multi Exit Discriminators) can tell the other ASs which particular links to use to enter your net.

§         "Community Attrbutes", e.g. send traffic for ATT customers directly to ATT; the community attributes tell us who its customers are; (the problem is that community attributes can lie...)

 

Final words of wisdom

During the Internet Bubble "The market value of a BGP engineer was a million dollars"