Backup Path Allocation Based
on A Link Failure Probability
Model in Overlay Networks
Weidong Cui, Ion Stoica, and Randy H. Katz
EECS, UC Berkeley
{wdc, istoica, randy}@EECS.Berkeley.EDU
The Overlay Network
Motivation
The Lower Level Network
The Internet is actually a BIG overlay network.
Link failures in an overlay network usually are not
independent because they may share same
links/nodes in the lower level network.
Having two orthogonal service/communication paths
(a primary path and a backup path) can improve
service availability.
This project studies the problem of how to find and
set up two orthogonal paths based on a correlated
link failure model in overlay networks.
A Link Failure Probability
Model in Overlay Networks
The link failure probability model has



Single link failure probabilities Pr( Li )
Double link failure probabilities Pr( Li , L j )
Assumption: probabilities are SMALL!
Path failure probabilities



Single path failure probability Pr( Pa )  L P Pr( Li )
i
a
Double path failure probability Pr( Pa , Pb )  
(
Pr( Li , L j ))
L P L P
j
How to obtain the link failure probabilities?
Li
b
i
a
The Overlay Network
Pr( Li )  Pr(lu )  Pr(lw )
Lj
lu
lw
lv
The Lower Level Network
Pr( Li , L j )  Pr(lu )  Pr(lv , lw )
Backup Path Routing
Algorithms
Probability Cost Routing Algorithm

Goal: find a backup path such that it minimizes
the joint path failure probability given a shortest
path as the primary path
Pr( Pa , Pb )  L P (L P Pr( Li , L j ))
j

b
i
a
Probability cost for link L j used in the backup path
is:

Li Pa
Pr( Li , L j )
Second Shortest Routing Algorithm

Find a second shortest path as the backup path
which is disjoint from the primary path
Backup Path Bandwidth
Allocation Algorithms
Duplicate Bandwidth Allocation

Same bandwidth on backup paths as on
primary paths
Single-Link-Failure Bandwidth Allocation

Allocate bandwidth on backup paths such
that it can recover from a single link failure
Zero Bandwidth Allocation

A baseline for comparison.
Simulation Experiments
Use GT-ITM to generate random network
topology.
Assume the link failures in the lower level
network are independent.
Calculate the link failure probabilities in the
overlay network based on the link failure
probabilities in the lower level network.
Simulate the link failures in the lower level
network
link failure time
l1
l2
...
......
lM
T
Results: Robustness
Results: Efficiency
Results: Fault Tolerance
Conclusion and Future Work
Probability-cost routing algorithm and singlelink-failure bandwidth allocation algorithm
have good performance on robustness and
efficiency.
Probability-cost routing algorithm is fault
tolerant to noise of overlay link failure
probabilities.
Future Work


Test the algorithms both in NISTNet emulation
testbed and in a real overlay network
Study how to obtain overlay link failure
probabilities in real networks.