A Dynamically Adaptive Hybrid Algorithm for Scheduling Lightpaths in Lambda-Grids

Neena R. Kaushik and Silvia M. FigueiraSanta Clara University

Presented by Yang-suk Kee

Outline

Background Lightpath Scheduling Experiments Conclusion

Background

Advance reservation• Guarantees that resources will be available at a pre-determin

ed time to participate in the execution of a Grid application Lambda Grid

• Lightpath A wavelength data channel liking multiple optical segment

s A single wavelength through the path Different wavelength with wavelength conversion

Data-intensive grid application Multiple paths

Problem Definition

Advance reservation of multiple lightpaths • How to schedule multiple lightpaths requested by advance re

servation over optical network with no converter?

B D

C

A

),,( 2es tt

),,( 1es tt

Source Destination

Lightpath Scheduling

Skeleton• Step1: determines edge-disjoint paths • Step2: determines wavelengths for paths

Edge-disjoint path• nth edge-disjoint path

A path that does not share any of its edges with the previous n-1 edge-disjoint paths

1st edge-disjoint path is the shortest path• Algorithm

Repeat Dijkstra’s shortest path algorithm by removing the edges that were part of the shortest paths.

Lightpath Scheduling Schemes Spreading

• Balances the wavelength assignment throughout the available edge-disjoint paths

Packing• Concentrates the wavelength assignment in the shortest

path first followed by the alternate edge-disjoint paths

Spreading Packing

B D

C

A

B D

C

A),( 21 ),( 21

),,,( 4321

A->C (4)

Wavelength-Balancing Algorithm

BeginFor i = 1 to number of wavelengths

For j = 1 to number of edge-disjoint pathsIf wavelength[i] is available for all segments in edge-

disjoint path[j]Allocate wavelength[i] for all segments in edge-

disjoint path[j]If all requests of user are satisfied

Print successElse

Print the number of requests satisfied and number denied

End

Wavelength-Concentrating Algorithm

BeginFor i = 1 to number of edge-disjoint paths

For j = 1 to number of wavelengthsIf wavelength[j] is available for all segments in edge-

disjoint path[i]Allocate wavelength[j] for all segments in edge-

disjoint path[i]If all requests of user are satisfied

Print successElse

Print the number of requests satisfied and number denied

End

Scheduling Comparison

Balancing Concentrating

B D

C

A

A->C (4), B->D (4), A->B (1), B->D (2)Four wavelengths per link

B D

C

A

(x)

(x)),( 21 ),( 21

),( 21

),( 21

)( 3

)( 3

)( 3

),,,( 4321

),,,( 4321

Blocking Probability

Requests for 1-hop lightpaths in a 32-node ring: uniform requests

Requests for 8-hop lightpaths in a 32-node ring with chords: constant requests.

Lessons

From the experiments using simple topologies• Partial mesh/ring/ring with spike/ring with chord

Balancing algorithm has similar or better blocking probability than concentrating one in most cases

Concentrating algorithm is better• With short hop (1-hop)

Shorter time slots provide lower blocking probability

=> Hybrid algorithm of balancing and concentrating

Hybrid Algorithm of Balancing and Concentrating begin firstpass = 1; while (firstpass is not equal to 3)

• for i = 1 to number of wavelengths for j = 1 to number of edge-disjoint paths

if first pass is equal to 1 and edge-disjoint path has more than x-hops

Continue … check if it satisfies balance algorithm

end (for edge-disjoint paths loop)• end (for number of wavelengths loop)• Increment firstpass by 1

end (while loop) return the number of requests satisfied end

Experimental Setup

Metric of quality• Blocking probability (denial rate)

Simulation-based• FONTS (Flexible Optical Network Traffic Simulator)

Generates on-demand and advance reservation requests Uses Stochastic models

• LRSS (Lightpath Request Scheduling Simulator) Simulates scheduling algorithms Takes as input the network topology and a trace of reque

sts for ligthpaths (from FONTS)

request

rejectB N

NP =

FONTS (Flexible Optical Network Traffic Simulator) Request arrival time

• Poisson distribution Source node/Destination node

• Constant/uniform/arbitrary probability Size of data to transfer (Tera bytes)

• Constant/uniform/arbitrary probability/heavy-tailed Number of lightpaths requested

• Constant/uniform/heavy-tailed Advance reservation start time

• Poisson distribution Number of time slots

• Constant/variable

Simple Topology

Requests for 1-hop lightpaths in a 32-node ring: uniform requests

All schemes except balancing are identical

Requests for 8-hop lightpaths in a 32-node ring with chords: constant requests.

16 17 18

20 24

National Lambda Rail

National Lambda Rail: Constant Requests

Conclusion & Discussion

Conclusion• Provides simulation tools for lightpath scheduling• Quality of scheduling algorithms depends on traffic characteri

stics• Hybrid scheduling achieved lower blocking probability

Discussion• Blocking probability represents network utilization properly?• This hybrid algorithm is close to balancing algorithm except t

he case of ring topology.• Need more study about the degree of nodes, length of hops,

and network utilization