Fast recovery in IP networks using Multiple Routing Configurations
Amund Kvalbein Simula Research Laboratory
Motivation
Increasing use of the Internet for applications with stringent performance requirements
Telephony, videoconferencing, online games ISPs must adhere to tough SLAs
The recovery mechanisms in the Internet are not designed for these requirements
Many (most) failures are short lived Failures are advertised too widely! This gives slow reaction and fosters instability
Sept 19 2007 GaTech networking seminar 2
Our approach
Failure reaction should be local
To avoid instability and overhead Challenge: avoid loops
Failure reaction should be proactive
To reduce recovery times and packet loss Challenge: minimize overhead
Sept 19 2007
GaTech networking seminar
Outline
Multiple Routing Configurations
The basic idea Generating backup configurations Forwarding
Evaluation Load balancing improvement Implementation issues Wrap up
GaTech networking seminar 4
Sept 19 2007
Multiple Routing Configurations
Guaranteed protection against single link, node or SRLG failures Same mechanism for both link and node failures
Generally difficult to distinguish at neighbor
A configuration is the graph and the weight function
Different weight setting in each configuration
Sept 19 2007 GaTech networking seminar 5
The general observation
An unused link can fail without consequences So can a single-connected node Several links and/or nodes can be protected in one logical topology
All nodes are still reachable
Build topologies so that all elements are protected
Few such topologies are needed to protect all elements!
Sept 19 2007 GaTech networking seminar 6
Isolated links and nodes
An isolated link has infinite weight A restricted link has a high weight wr wr is chosen so that the link is used only as a last resort A node is isolated when all attached links are either isolated or restricted
Traffic never goes through an isolated link or an isolated node!
Sept 19 2007 GaTech networking seminar 7
Building backup configurations
6 1
4 2 5
C0
Sept 19 2007 GaTech networking seminar 8
Building backup configurations
6 1 6 1 6 1
4 2 5
4 2 5
4 2 5
C1
Sept 19 2007
C2
GaTech networking seminar
C3
9
Forwarding
6 1
6 1
4 2 5
4 2 5
Sept 19 2007
GaTech networking seminar
10
How many configurations are needed?
16
32 64 128 512
Sept 19 2007
GaTech networking seminar
11
How long are the backup paths?
Sept 19 2007
GaTech networking seminar
12
What about load distribution?
Sept 19 2007
GaTech networking seminar
13
Why bother to avoid overload?
- its only for short while
Motivation for fast rerouting
Do not loose packets Increase stability
FRR should not make it worse for unaffected traffic
Sept 19 2007
GaTech networking seminar
14
Routing performance during FRR
Given TM estimate: What decides the load distribution?
Link weights in C0 Structure of backup configurations Link weights in backup configurations
Three step approach
Optimize link weights in C0 Build backup configurations Optimize link weights in backup configurations
Sept 19 2007
GaTech networking seminar
15
Building backup configurations
Optimize C0 independently Identify the heaviest nodes (most traffic) Build configs with good connectivity for heavy nodes
Sept 19 2007
GaTech networking seminar
16
Optimizing link weights
Heavy optimization task
Dependencies between configurations
Local weight search heuristic
Based on well known Fortz/Thorup method
Optimize only for most severe link failures Take advantage of configuration structure
A link failure only activates one or two backup configurations
Sept 19 2007
GaTech networking seminar
17
Evaluation Max link load
Real and synthetic network topologies Gravity model traffic demands
Network
Geant
Failure free
0.68
MRC n=5
1.01
MRC n=10
1.08
OSPF
1.20
Cost239
Sprint US (POP)
0.66
0.64
0.99
1.10
0.99
1.10
0.99
1.10
German Telecom
Sept 19 2007
0.66
1.02
1.02
1.17
18
GaTech networking seminar
Evaluation Number of configurations
Sept 19 2007
GaTech networking seminar
19
Implementation issues
Representing backup configurations
IETF: Multi-Topology routing
Can calculate independent shortest path trees in each topology
Need ability to switch configuration in-flight Marking packets
Same problem as in MT-routing Reuse of ToS/DSCP bits has been proposed
Sept 19 2007 GaTech networking seminar 20
Summary
MRC guarantees protection against any single link or node failure Modest state overhead Small path length stretch for recovered traffic Flexibility in how recovered traffic is routed Realistic to implement
Sept 19 2007 GaTech networking seminar 21
Related work
Failure Insensitive Routing (FIR)
Relies on interface-specific routing tables to infer link failures
Not-via addresses
Calculates one configuration for each protected element
Sept 19 2007
GaTech networking seminar
22
MRC extensions
Multi-failure protection
SRLG, uncorrelated failures Can guarantee protection against two independent failures (at a cost)
Improved configuration construction
Eliminate isolated links Use deflection in forwarding procedure
Use in TE context
Spread demands on several topologies
Lab implementation
Using Quagga routing software
Sept 19 2007 GaTech networking seminar 23
