Title: ISP Backbone Traffic Inference Methods to Support Traffic Engineering
1ISP Backbone Traffic Inference Methods toSupport
Traffic Engineering
Olivier Goldschmidt Senior Network Consultant
2Outline
1. Problem Description 2. Inputs to the Models
3. Constraints of the Models 4. Inference
Methods Pseudo-Inverse Method Linear
Programming 5. Test Results 6. Conclusion and
Open Issues
3RATIONALE
A major headache for Internet Service Providers
is to estimate the end-to-end traffic volumes on
their backbone network.
Reliable traffic estimates between ingress and
egress points are essential to traffic
engineering purposes such as ATM PVC or LSP
layout and sizing.
4Problem Description
An "easy" solution is to turn on NetFlow or
IP-Accounting on all ingress and egress
interfaces.
But such solution is - Costly - Impractical
5Problem Description
Objective of traffic inference is to "guess" end
to end aggregate traffic using limited
information.
6Inputs to the model
Deterministic Information
Measured Information
Usage Information
7DETERMINISTIC INFORMATION
Network Topology
Types of routers and links
Routing paths between end points
8MEASURED INFORMATION
Baselining Information on network interfaces
using SNMP
Partial RMON/RMON2 information using
selective probes (NetFlow or IP account.)
9USAGE INFORMATION
Data that can be correlated with the traffic
on the network
Allows to derive additional constraints on
the network traffic.
10WAN Link
Ingress-Egress points
Internal routers
113
3
Assume that reading are symmetric.
3
3
Interface flow reading
12(No Transcript)
13(No Transcript)
14CONSTRAINTS
153
16PSEUDO-INVERSE METHOD
17LINEAR PROGRAMMING METHOD
18OBJECTIVE FUNCTION COEFFICIENTS
Choice of coefficients for the objective function
will determine the precision of the end to end
traffic estimates.
Obvious choice is to set all coefficients to 1
and to maximize or to minimize the objective
function
But this choice is not neutral
19EXAMPLE
Assume these are the true traffic demands
Notice that all interface flows are equal to 20
20If all objective coefficients are equal to 1
If objective function is maximized
If objective function is minimized
21But if coefficient are equal to the number of
hops of demand route
Is a solution whether objective function is
maximized or minimized
22Another advantage of the LP method
Allows to add constraints that represent usage
information.
For instance constraint the very unlikely
end-to-end traffic to be close to zero.
Also known traffic from NetFlow or IP accounting
readings can be included as constraints in the
linear program.
23Test Results
NETWORK
- 60 Routers
- 114 WAN Links
- 529 Traffic demands
- Bandwidth from 0 to 256 Kbps
24Test Results
1. Route the demands 2. Compute the resulting
interface flows 3. Apply the Linear Programming
method to estimate the end-to-end traffic
demands 4. Compare those estimates with the
original traffic demands in of absolute
difference estimate-true value/true value The
following charts show of demands with given
relative error
25Objective coefficients number of hops
26Objective coefficients number of hops
27All objective coefficient 1
28Netflow turned on on five random routers
29Netflow turned on on five most used routers
30Netflow turned on on ten random routers
31Comparison of different results
32Conclusions
Objective coefficients in LP need to be scaled
Turning NetFlow on a few selected interfaces can
greatly improve the traffic estimates.