# Combining multihoming with overlay routing (or, how to be

Document Sample

```					Combining multihoming with
overlay routing
(or, how to be a better ISP
without owning a network)

Yong Zhu, Constantine Dovrolis,
and Mostafa Ammar
Georgia Institute of Technology
Speaker: Chen-Hung Yu
Basic form of Internet

Singlehoming
Multihoming & Overlay routing
OSP – Overlay Service Provider

Operates a Multihomed Overlay Network
(MON)
MON node – a multihomed router
An Internet provider that does not own a
network
MON nodes

If a MON node is multihomed with K ISPs
 each flow has K direct MON paths
With N MON nodes  each flow increases
to K 2  N  1 indirect MON paths
Outline

Introduction
Model and Problem Formulation
MON design heuristics
Evaluation and Discussions
Conclusions
MON design problem

Where to place MON nodes and how to
select upstream ISPs for each nodes.
Objectives
Profitable
Better performance
Less expensive
Revenue, cost, customer subscribe
The problem involves …

ISPs
Performance of the native network
The location and traffic matrix of potential
customers
The OSP routing strategy
Pricing function
Node deployment costs
ISPs and the native network

POP p = (l, i) – the access point to ISP i at
location l, P denotes the set of all POPs
LOC(p) = l; ISP(p) = i
We denote Il as the set of all ISPs the can be
connected from location l
TPP
Native-layer performance – matrix
The entry  p , q represents the propagation RTT
from POP p to q
Estimate the matrix T

Directly measure – e.g., ping
If can’t, try to find the model
Mostly depends on the physical distance
between two POPs
Interdomain – RTT increases with the
number of AS in the route



Driving
distance from
p to q
Interdomain cases   A constant depends on
the # of AS hops h


MON representation

A MON node is present at POP p if the
node is located at LOC(p) and connected
to ISP(p)
POP selection vector

The locations of all MON nodes
Customers and OSP-preferred flows

The workload of customer u is a set of
flows F(u)
A flow f = (sf, df, rf, τf)
OSP-preferred flow and OSP-preferred
path
Subscribe – At least a fraction H of a
customer’s traffic is in OSP-preferred flows
OSP routing strategy

Direct-Routing-First (DRF)




OSP revenues
Let     be the OSP pricing function
The total OSP revenue


Required upstream capacity at POP p

The pricing
function used
Total capacity cost               by the ISP at
POP p

OSP costs

Required upstream capacity at POP p


The pricing
Total capacity cost               function used
by the ISP at
                                   POP p

Cost of deploying a
Total node deployment cost    MON node at location
l

Problem statement

Inputs:
Native network information
OSP information
Customer information
Determine the POP selection vector MON
to maximize the profit:
Problem statement (cont.)

constraints
At most N MON nodes
Maximum multihoming degree
NP-hardness
Reduction from the set covering problem
MON design heuristics

Select up to N locations for placing MON nodes
Select up to K upstream ISPs for each
deployed MON node
Present four heuristics differ in terms of
their inputs
Four Heuristics
 RAND
 CUST –
 places N MON nodes at the locations with the maximum number
of customers
 Each selected l then selects the             locally present
ISPs with the maximum coverage
 TRFC – uses the aggregated traffic rate that originates
from all potential customers
 Places MON nodes at locations where “traffic heavy” customers
are located
 Select ISPs that receive the maximum traffic rate from customers
 PERF
Performance-driven (PERF)

If there are OSP-preferred direct paths,
then CUST and TRFC perform quite well.
However, when many customer flows only
have indirect OSP-preferred paths, they
will fail.
Associate
Evaluation

Compare the MON design heuristics
OSP profitability and performance
Depending on # of MON nodes, degree of
multihoming, node deployment cost, OSP/ISP
pricing ratio
Examine various OSP routing strategies
Simulation setup

Customer
CUST-POPUL, CUST-UNFRM
Flow
RATE-GRVTY, RATE-UNFRM
Pricing

^
P (r )
     Rp 
Pp ( r )
Effect of number of MON nodes
Effect of OSP routing strategy
Effect of pricing ratio
Effect of maximum multihoming degree
conclusions
 Examine multihoming + overlay routing from the
pragmatic perspective of an OSP
 Meet the objectives – profitable、better
performance、less expensive
Use a performance-aware MON design heuristic
Deploy nodes at “key” locations
Connect each MON node to ISPs that can directly reach
traffic-heavy destination POPs
Direct path > indirect path
Charge less than competing native ISPs
Determine good trade-off between the # of MON nodes
and multihoming degree  based on the d(l)

```
DOCUMENT INFO
Shared By:
Categories:
Stats:
 views: 10 posted: 6/12/2009 language: English pages: 30