Sharing Normal Bandwidth during a

Failure

Rajarshi Gupta, Eric Chi, Jean Walrand{guptar, echi, wlr}@eecs.berkeley.edu

University of California, Berkeley

2

Background Backup paths are necessary in many

networking scenarios

We look at : Provisioned paths (normal and backup) with

guaranteed bandwidth Protection against one failure Path Protection (vs. link/span protection)

Examples: Optical, MPLS

3

Outline

Related Work Key Idea

Centralized Algorithm Distributed Algorithm

Analysis Results

4

Shared Protection Path

Disjoint normal paths can share backup bandwidth

C

B

GD

A

E F

Bandwidth on path C-E-F-G is shared by backup paths of both A-B-G and C-D-G

5

Related Work

Kodialam, Lakshman (2000) SCI and SPI

Liu, Tipper, Siripongwotikorn (2001) Matrix formulation to characterize SPP

Qiao, Xu (2002) DPIM-SAM and DPIM-MA

Several other work in the area of Optical, ATM and G/MPLS networks

6

Key Idea Can even share normal bw during failure

Upon failure, all traffic goes away from normal path (C-D-F-G)

The bandwidth on D-F may then be used for backup traffic (C-E-D-F-B)

C GD

A

E

F

B

7

Notation

Flow is characterized by (src, dst, bN, bB) bN = normal bandwidth bB = backup bandwidth

Each link stores quantities BT(j) = total bandwidth on link j BN(j) = normal bandwidth on link j BB(j) = backup bandwidth on link j Available bw BV(j) = BT(j) – BN(j) – BB(j)

8

Enhanced State Variables

Each link also stores two vectors X(i,j) = extra bw on link i when link j fails Y(i,j) = bw that goes away from link i when link j

fails

BB(i) is the most backup traffic that link i needs to support, caused by failure at any of the other links

9

Centralized Algorithm

When a new flow (src, dst, bN, bB) with normal path N and backup path B is accepted, we update

10

Augmented Routing Indifferent towards chosen routing algorithm Provides better accounting of available

resources, so augments performance of any routing algorithm

Example using Dijkstra’s Algorithm Discard all links with inadequate normal bw Compute normal path using source routing Remove links used in normal path Discard links whose BV < BB (instead of BV < bB) Compute backup path on resulting graph Update link states accordingly

11

Distributed Algorithm

If link i receives normal path msg

If link i receives backup path msg

Separate message (including normal path information) sent to normal and backup paths

For a load-insensitive algorithm (e.g. shortest path), distributed algorithm works as well as centralized

12

Analysis of Gains

Define the following

Normal sharing recovers an amount G(i) for link i

We can upper bound G(i)

13

Analysis of Gain Scenario

The upper bound can be achieved in a ring

Gains occur when

Total Gains:

14

USA Topology

OC-192

OC-48

15

USA Topology: Results

Primary: Least LoadedBackup: Least Loaded

16

USA Topology: More results

Routing Algorithms Primary Path: Min Hop Backup Path: Min Hop

Routing Algorithms Primary Path: Min Hop Backup Path: Cumulative

Backup

17

Ring Topology

OC-48

OC-12

Metropolitan Area Network

18

Conclusion Key fact: even normal bandwidth may

be shared during failures

Update link states to account for sharing of both normal and backup resources Centralized Distributed

Simulation results show significant gains in network resource utilization