1

Optical Burst Switching

Outline

• Introduction to Optical Transport Paradigms

• Optical Burst Switching– Architecture– Burst Assembly– Routing– Signaling– Channel Scheduling– Contention Resolution

• Prioritized Burst Segmentation and Composite Burst Assembly for QoS

• Absolute QoS Differentiation

• TCP over OBS with Burst Retransmission

Optical layer

IP Router

Optical Switch

WDM

Layered Network Model

IP layer

Wavelengths (channels)

(Electronic)

Applications Demands

Voice Over IP

Streaming Video

Grid Computing

Storage Area Networks

Multimedia

Data

Applications Service Requirement

Optical Transport Paradigm

High Bandwidth

Dynamic Provisioning

Service Differentiation

Optical Circuit Switching

Optical Packet Switching

Optical Burst Switching

Optical Circuit Switching

• For each request, set-up a circuit• Static allocation of bandwidth for the entire duration

of the connection

• Pros: – Suitable for smooth traffic (Voice)

• Cons: – Long circuit set-up latency– Inefficient for bursty traffic (Data)

Optical Circuit Switching (cont.)

• Circuit switched networks optimized for Voice

• Data: Accounts of 90% of traffic by 2005

• Data tends to be bursty• Static bandwidth allocation is not efficient

Optical Circuit Switching

2

Optical Packet Switching

• A photonic packet contains a header and the payload

• Packet header is processed all-optically at each node and switched to the next hop

• Pros:– Statistical multiplexing of data– Suitable for bursty traffic

• Cons: – Very fast switching speeds (nanoseconds)

Optical Burst Switching

• An optical burst contains multiple IP packets

• Out-of-band control header transmitted ahead of data burst transfer

• Electronic control plane and optical data plane

• Pros:– Practical alternative to optical packet switching– Statistical multiplexing of data– Suitable for bursty traffic

Motivation for OBS

HighLowMediumLowHighOptical Burst

Switching

HighHighFastLowHighOptical Packet

Switching

LowLowSlowHighLowOptical Circuit

Switching

TrafficAdaptively

Proc. / Sync. Overhead

Switching Speed Req.

SetupLatency

Bandwidth Utilization

Optical Switching Paradigm

OBS combines the best of the two while avoiding their shortcomings

Outline

• Introduction to Optical Transport Paradigms

• Optical Burst Switching– Architecture– Burst Assembly– Routing– Signaling– Channel Scheduling– Contention Resolution

OBS Network Architecture

• Core– Signaling– Scheduling– Contention Resolution

• Edge– Burst Assembly– Routing

Burst Header

Offset Time

Core Node

Edge Node

WDM Link

IngressNode

EgressNode

Input Traffic

Output Traffic

IP

OBS Node Architecture

Switch

2

1 1

2

O/E/O2 2

1 1Control Packets

Data Bursts

Offset Time

ControlWavelengths

DataWavelengths

Scheduler

Control Packet Processing

[U-Buffalo]

3

• Core– Signaling– Scheduling– Contention Resolution

• Edge– Burst Assembly– Routing

Core Node

Edge Node

WDM Link

IngressNode

EgressNode

Input Traffic

Output Traffic

Optical Burst-Switched Network Burst Assembly

• Aggregate multiple (IP) packets going to the same destination into a single burst

• Assembly Mechanisms: Timer-based and Threshold-based

• Timer-based assembly:

– For each fixed timer interval, all the packets in the queue are framed into a single burst

• Threshold-based assembly:

– After a fixed length threshold is reached, all the packets in the queue are framed into a single burst.

Burst Assembly (cont.)

Data Channel

Control Channel

IP Packet

Time or length threshold is reached

A header is generated and sent out

Burst Assembly

Node

IP Packet Queues

DST

0

1

N

A unique packet queue for every destination egress node

Burst transmitted after offset time

Burst Aggregation Delay in ms

[U-Buffalo]

Assembly Framework• Parameters

– Burst Length: Minimum – Maximum– Threshold: fixed-sized burst in the network– Timer: constant burst arrivals into the network– Burst Priority: mapping IP classes to burst priorities

Pros/Cons• Burst Length

– Long Bursts – High aggregation delay; low control overhead– Short Bursts – Low aggregation delay; high control overhead

Guidelines• Select assembly mechanism based on higher-layer application

requirement• Select burst length based on optical-layer requirements

• Core– Signaling– Scheduling– Contention Resolution

• Edge– Burst Assembly– Routing

Core Node

Edge Node

WDM Link

IngressNode

EgressNode

Input Traffic

Output Traffic

Optical Burst-Switched Network Routing

• Source Routing: – All the edge nodes know the complete topology of the network– Source chooses the fixed shortest path to destination

• Fixed Shortest Path Routing– Minimum number of hops (lower loss)– Minimum total physical distance (lower delay)

• Extensions/Research Issues:– Fixed alternate path routing– Dynamic source routing (load balancing)– Intra- domain and Inter-domain routing

4

• Core– Signaling– Scheduling– Contention Resolution

• Edge– Burst Assembly– Routing

Core Node

Edge Node

WDM Link

IngressNode

EgressNode

Input Traffic

Output Traffic

Optical Burst-Switched Network Classification of Signaling Techniques

• Direction– One-way (unacknowledged)– Two-way (acknowledged)

• Initiation– Source– Destination– Intermediate

• Resource– Persistent– Non-persistent

• Reservation – Immediate– Delayed

• Release– Explicit– Implicit

• Computation– Centralized– Distributed

Signaling Technique

• Distributed and Non-Persistent techniques

• Reservation Mechanism: Based on the start of the reservation– Immediate Reservation: Immediately after the control heater – Delayed Reservation: At the start of the burst

• Release Mechanism: Based on the release of the reservation– Implicit Release: based on burst length information – Explicit Release: explicit release control packet used

HeaderBurst

Offset

Immediate Reservation

DelayedReservation

ImplicitRelease

Explicit Release

REL

Tell-and-wait (TAW) Signaling

Direction: Two-way Initiation: Source (SIR) / Destination (DIR) Reservation: ImmediateRelease: Explicit

• Header contains source, destination

• Pros:

– Lower loss: burst sent only after bandwidth reservation in the core

• Cons:

– High path setup latency

REL

REL

REL

REL

Just-In-Time (JIT) Signaling

Direction: One-way Initiation: SourceReservation: ImmediateRelease: Explicit

• Header contains Src and Dst• Header sent out while the burst is

being assembled• Burst sent after assembly (ensure

that min offset has reached)• Explicit release message send after

data burst transmission or an In-Band-Terminator (IBT) sent with the data burst

Src DstHeader

δ

δ

δBurst

Time

ST

REL

REL

REL

Just-Enough-Time (JET) Signaling

Direction: One-way Initiation: SourceReservation: Delayed Release: Implicit

• Header contains burst length, offset time, source, destination

• Pros:– Low path setup delay– High bandwidth utilization

• Cons:– Higher data loss due to contention

for resources in the core

Src DstHeader

δ

δ

δBurst

Offset

Time

ST

Offset time = δ * (No. of Hops) + ST

Note: Burst Aggregation Delay in ms; Offset time in order of μs

5

Summary: Signaling Techniques

ExplicitImmediateSourceOne-wayJust-In-Time (JIT)

ImplicitDelayedSourceOne-wayJust-Enough-Time (JET)

ExplicitImmediateSrc (SIR)/ Dst (DIR)

Two-wayTell-And-Wait (TAW)

ReleaseReservationInitiationDirectionSignaling Techniques

Tradeoff: Efficiency vs. Simplicity

Just-In-Time (JIT)

0

0Header

1

1

Control Channel

Data Channel

2

2

0REL HeaderHeader

0

0Header

1

1

2

2

OT0

OT1OT1

Control Channel

Data Channel

Just-Enough-Time (JET)

• Core– Signaling– Scheduling– Contention Resolution

• Edge– Burst Assembly– Routing

Core Node

Edge Node

WDM Link

IngressNode

EgressNode

Input Traffic

Output Traffic

Optical Burst-Switched Network

Channel Scheduling

• Which output channel to use?– If none is available, use contention resolution

• Two categories of scheduling algorithms– Without void filling

• Maintain the latest available time on every channel– Latest Available Unscheduled Channel (LAUC)– First Fit (FF)

– With void filling• Maintain the starting and ending time for every scheduled burst

on every channel– Latest Available Unscheduled Channel with Void Filling (LAUC-VF)– First Fit with Void Filling (FF-VF)

Channel Scheduling (cont.)

SaLbEa

SaLbEa

LAUC

LAUC-VF

D0

D1

D2

D3

Wavelengths

D0

D1

D2

D3

First Fit

FF-VF

LAUC

LAUC

Channel Scheduling Techniques

LAUTi , GAPiO(W)LAUC

S(i,j), E(i,j)O(W log(n))FF-VF

S(i,j), E(i,j) , GAPiO(W log(n))LAUC-VF

LAUTiO(W)FF

State Information

TimeComplexity

SchedulingTechniques

W: # of data channels; n: # of burst currently scheduled

LAUTi : Latest available unscheduled time on channel i

S(i,j) : Starting time of burst j on channel i ; E(i,j) : Ending time of burst j on channel i

GAPi : time difference between the starting time of the arriving burst and LAUTi (or the ending time of the previous burst)

6

Optical Burst-Switched Network

• Core– Signaling– Scheduling– Contention Resolution

• Edge– Burst Assembly– Routing

Core Node

Edge Node

WDM Link

IngressNode

EgressNode

Input Traffic

Output Traffic

Data Burst Contentions

• Contention occurs when more than one burst attempts to go out of same output wavelength at the same time

• Unique to all-optical networks– Traditional networks employ electronic buffering to resolve contentions– Lack of optical buffers (cannot store light)

Core Switch

Drop Entire Burst

• Drop Policy– One of the bursts will be dropped in its entirety– Even though overlap between the bursts may be minimal

Optical Buffering

• Achieved through Fiber Delay Lines (FDLs)

• Several different FDL architectures proposed– Feed-Forward (above) / Feed-Backward (loop)– Input-Buffer / Output-Buffer– Share-per-Port / Share-per-Node

• Issues– Limited buffer capacity (200 km = 1 ms)– Additional hardware cost– Affect signal quality

Burst in

FDLs

Delayed Burst out

Input-Buffer Architecture

Bursts are buffered before actually scheduling them onto a specific wavelength on the intended output fiber

Output-Buffer Architecture

Bursts are buffered after scheduling them onto a specific wavelength on the intended output fiber

Wavelength Conversion

• Converting the wavelength of an incoming channel to another wavelength at the outgoing channel

• All-optical wavelength converters

• Techniques– Four Wave Mixing– Cross Gain Modulation– Cross Phase Modulation

• Issues:– Additional hardware cost

7

Deflection Routing

• Deflect contending bursts to alternate port

• Pros– Save bandwidth: A

dropped burst wastes the bandwidth on the partially established path

– Save time: The delay becomes very large when retransmitting a blocked burst in long-distance links

– No additional hardware cost• Cons

– Out-of-sequence delivery– Potential looping

• Loopless Deflection Routing

F G H I J

A B C D E

Without Deflection

With Deflection

A DB C E A DB C I J E

Tot

al T

rans

. Tim

e

Tot

al T

rans

. T

ime

Deflection Nodes

Tim

eSa

ved

Without Deflection With DeflectionA DB C EA DB C E A DB C I J E

Tot

al T

rans

. Tim

e

Tot

al T

rans

. T

ime

Deflection Nodes

Tim

eSa

ved

Without Deflection With Deflection

[U-Tokyo]

Core Switch

Burst Segmentation

Original burst

Contending burst

Tail Dropping

Head Dropping

Dropped Segments

• When contention occurs, only overlapping segments are dropped

• Two Approaches: Head Dropping and Tail Dropping

V.M. Vokkarane and J. P. Jue, “Burst Segmentation: An Approach for Reducing Packet Loss in Optical Burst Switched Networks,” SPIE/Kluwer Optical Networks Magazine, vol. 4, no. 6, pp. 81-89, Nov.-Dec. 2003.

• Basic idea: to provide packet-level loss granularity in OBS• Burst is divided into segments• Each segment consist of single or multiple IP packets/ATM cells

Burst Segmentation

Signaling Issues

2

1

ba aST

Does tail-dropping have signaling overhead at downstream nodes?

a

OBS CORE

Node 1

Node N

ac

c

b2

1

2

1

2

1

ST

ST

NO

Segmentation in Void Filling

Non-Preemptive [No Signaling Overhead]

Preemptive [QoS Reasons; Additional Signaling Overhead]

Simulation Network

14-node NSFNET

Assumptions

NOWavelength ConversionNOOptical Buffering10 μsSwitching Time1500 bytesPacket Length10 Gb/sLink Transmission Rate 100 μs (exponentially distributed)Average Burst LengthPoissonBurst Arrivals

8

Packet Loss Performance

Segmentation

Drop

Log Scale

Presentation Outline

• Introduction to Optical Transport Paradigms

• Optical Burst Switching

• Prioritized Burst Segmentation Framework– Burst Segmentation – Prioritized Scheduling – Composite Burst Assembly

Differentiated Services

• Problem: To differentiate bursts based on application specific requirements in an all-optical core network

• Standard IP-based techniques are not applicable– IP differentiation primarily based on electronic buffering

• Design criteria – Provide 100% class isolation– Scalable: support multiple priorities in the network

Additional Offset-based QoS

• Concept: Earlier the control packet reaches the node, higher the probability of a successful reservation for the data burst

• Higher priority bursts : Longer offset time

0

0Header

Burst

OT0

1

1

OT1

Control Channel

Data Channel

• Issue:

– Not scalable: since the offset increases with number of priorities

Prioritized Burst Segmentation Framework

• Efficient utilization of network resources

• Support for different classes of traffic

Composite Burst Assembly PrioritizedScheduling

BurstSegmentation

IEEE JSAC 2003, SPIE/Kluwer Optical Networks 2003, IEEE [ICC, OFC, GLOBECOM] 2002

Prioritized Burst Segmentation Framework

• Burst-level differentiation– Burst priorities– Differentiated scheduling and contention resolution

• Original burst – burst that is scheduled earlier• Contending burst – burst that arrives later

• Packet-level differentiation– Packet classes– Composite burst assembly – differentiated location of

packets within a burst

9

Core Switch

Burst Segmentation

Original burst

Contending burst

Tail Dropping

Head Dropping

Dropped Segments

• When contention occurs, only overlapping segments are dropped

• Two Approaches: Head Dropping and Tail Dropping

Contention Resolution Policies

• Drop Policy: – Drop the entire contending burst

• Deflect Policy: – Deflect the contending burst to the alternate port– If the alternate port is busy, drop the burst

• Segment Policy: – Segment the original burst– Drop the tail segments – Transmit the contending burst

Core Switch

Segmentation with Deflection

a ab

a`

a`

b

• Segment-First Policy:– Segment the original burst

– Deflect/Drop the tail segments

– Transmit the contending burst

Core Switch

Segmentation with Deflection (cont.)

a a

bb

• Deflect-First Policy:– If alternate port available

• Deflect the contending burst– If alternate port busy

• Segment original burst • Drop the tail segments• Transmit the contending burst

– 100% class isolation– Scalable to support multiple priorities in the network

Core Switch

Core Switch

Low Priority

High Priority

High Priority

Low Priority

Prioritized Scheduling

Case 1

Case 2

Contention Resolution Schemes

• DP: Drop Policy– Drop entire contending burst

• SDP: Segment-Drop Policy– Segment original burst and drop segments

• DDP: Deflect-Drop Policy– Attempt to deflect contending burst, otherwise drop it

• SFDP: Segment-First and Deflect Policy– Segment original burst– Attempt to deflect segments, otherwise drop

• DFSDP: Deflect-First, Segment, and Drop Policy– Attempt to deflect contending burst, otherwise– Segment original burst and drop segments

10

Loss Probability Delay Composite Burst Assembly

Core Switch

High Priority

High Priority

Case 1

A A A A A

B B A A A

B B A A A

High Priority

High PriorityCore Switch

Case 2 Composite Class Bursts

Single Class Bursts

Prob. Loss

A A A A A

Composite Burst Assembly (cont.)

• Probability of loss of a packet in a burst increases from head to tail of the burst

• System Parameters– Number of input packet classes (edge)– Number of burst priorities (core)

• Illustration: – Number of input packet classes = 4– Number of burst priorities = 2

Packet Class: A B C D

Burst Priority: 0 1(Highest)

AAAAAAAA

DDDDDDDD

Assembled Burst Contents

BBBBBBBB

CCCCCCCC

Burst Priority

0

0

1

1

4 Packet Class 2 Burst Priorities

AAAABBBB

Assembled Burst Contents

CCCCDDDD

Burst Priority

0

1

Composite Class Burst

Single Class Burst

AAAAAAAA

DDDDDDDD

Assembled Burst Contents

Timer & Threshold

BBBBBBBB

CCCCCCCC

Burst Priority

0

1

2

3

Threshold

Threshold

Threshold

Aggregation

AAAAABBB

DDDDDDDD

Assembled Burst

Timer & Threshold

BBBBBCCC

CCCCCDDD

Burst Priority

0

1

2

3

Threshold

Threshold

Threshold

Aggregation

Single Class Burst

Composite Class Burst

4 Packet Class 4 Burst Priorities

11

Generalized Burst Assembly Framework

• Parameters– Burst Destination– Burst Priority– Burst Length: Minimum/Maximum total number of packets in a

burst of a given priority– Packet Class Threshold: Minimum/Maximum number of packets of

a given class in a burst of given priority– Packet Class Location: Location of various packet classes within a

burst– Assembly trigger:

• Timer-based• Threshold based

– Total number of packets reaches threshold– Number of packets of given class(es) reach threshold

• Traffic grooming: packet destination may not necessarily correspond to burst destination

Prioritized Burst Segmentation Summary

• Objective: Minimize packet loss and provide application specific services in an OBS network– Burst Segmentation:

• reduced packet loss due to contentions– Prioritized Scheduling:

• differentiates services (100% class isolation)• Scalable solution to multiple priorities

– Composite Burst Assembly: • Further enhance differentiation between traffic

4:2 Mapping under Single Burst Assembly

One output port of an ingress node

Output link

4:2 Mapping under Composite Burst Assembly

Output link

One output port of an ingress node

Burst Priority

Packet Class

NOWavelength ConversionNOOptical Buffering10 μsSwitching Time1250 bytesPacket Length10 Gb/sLink Transmission Rate 100 packetsFixed Burst LengthPoissonBurst Arrivals

Simulation Assumptions

• 14-node NSFNET

• Input traffic ratios for packet classes:

A - 10%, B - 20%, C - 30%, and D - 40% (A: Highest)

• Class A Timeout - 50 ms

Prioritized Scheduling:Packet Loss Performance

• Contention resolution Scheme 3– Segmentation without deflection

High Priority

Low Priority

Traffic: 20% High – 80% Low

12

Composite Burst AssemblyPacket Loss Performance

4 Packet Classes: 2 Burst Priorities

Absolute QoS in OBS Networks

QoS Models

• Relative QoS Model– Provides relative service differentiation– High-priority class has better performance than low-priority

class

• Absolute QoS Model– Provides worst-case service guarantee– Necessary to support QoS sensitive applications

Absolute QoS Model in OBS

Input Traffic

Core Node

Edge Node DWDM Link

Ingress Node Egress Node

Input Traffic

Output Traffic

AdmissionControl

Absolute QoSDifferentiation atEach Core Node

• Core– Channel Scheduling: Guarantee Maximum Per-node Loss

• Edge– Admission Control– Resource Provisioning

Channel Scheduling Mechanisms

• Early drop – Intentionally drops the bursts of lower-priority class

• Wavelength grouping – Schedules the bursts on the provisioned wavelengths

Guarantee End-to-End Loss Requirement

• : MAX end-to-end loss probability of Class i• : MAX per-node loss probability of Class i • : Network diameter• To compute based on

13

Early Drop Mechanism

Standard Drop Mechanism Early Drop Mechanism

• Two Early Drop schemes to decide– Early Drop by Threshold: drops lower-priority bursts when loss of

higher-priority bursts exceeds the Max. loss threshold– Early Drop by Span: drops lower-priority bursts in order to guarantee loss

of higher-priority bursts in a certain range

High priorityLow priority

Early Drop Mechanism (cont.)

• Traffic statistics for class i traffic– : Burst arrival counter– : Burst drop counter– : Online burst loss probability

Early Drop by Span (EDS)

• : span between MIN and MAX loss threshold

• When each Class 1 (low) burst arrives, – If less than MIN loss threshold, no intentional dropping

– If between MIN and MAX loss threshold, drops Class 1 burst with a probability of

– If greater than MAX loss threshold, drops Class 1 burst

Wavelength Grouping Mechanism

• Two wavelength grouping schemes– Static wavelength grouping (SWG)– Dynamic wavelength group (DWG)

• : MIN # of wavelengths provisioned for Class 0 traffic– Computed by Erlang B Formula based on

• : # of wavelengths provisioned for Class 1 traffic– Obtained by

Static wavelength grouping Dynamic wavelength grouping

Wavelength Wavelength

Integrated EDS and WG

• If burst “should” be dropped according to EDS, it is given Label L1• If burst “should not” be dropped according to EDS, it is given Label L0• Do not drop bursts until WG stage• In WG stage, assign minimum number of wavelengths for bursts with

Label L0, and remaining wavelengths for bursts with Label L1• Allows Class 1 bursts to sometimes use Class 0 wavelengths if they are

not being used

Markov Chain Model

14

EDS/EDT Loss Probability Loss Probability ReferencesArchitecture and Signaling• C. Qiao and M. Yoo, "Choices, Features and Issues in Optical Burst Switching," Optical Networks Magazine, Vol.

1, No. 2, pp. 36-44, 2000.• C. Qiao and M. Yoo, "Optical Burst Switching - A New Paradigm for an Optical Internet," Journal of High Speed

Networks, Vol. 8, No. 1, pp.69-84, 1999 • L. Xu, H.G. Perros, and G. Rouskas, "Techniques for optical packet switching and optical burst switching," IEEE

Communications Magazine, Vol. 39, No. 1 , pp. 136-142, Jan. 2001. • J.Y. Wei, and R.I. McFarland Jr., "Just-in-time signaling for WDM optical burst switching networks," IEEE Journal

of Lightwave Technology, vol. 18, issue 12, pg. 2019-2037, Dec. 2000. Burst Assembly• K. Dolzer and C. Gauger, "On burst assembly in optical burst switching networks - a performance evaluation of Just-

Enough-Time," Proceedings 17th International Teletraffic Congress (ITC 17), Salvador, September 2001. • V.M. Vokkarane, K. Haridoss, and J.P. Jue, "Threshold-Based Burst Assembly Policies for QoS Support in Optical

Burst-Switched Networks," Proceedings, SPIE OptiComm 2002, Boston, MA, vol. 4874, pp. 125-136, July 2002. Channel Scheduling and Contention Resolution• Y. Xiong, M. Vanderhoute, and H.C. Cankaya, "Control architecture in optical burst-switched WDM networks,"

IEEE JSAC, vol. 18, no. 10, pp. 1838-1854, Oct. 2000. • V.M. Vokkarane and J.P. Jue, "Burst Segmentation: an Approach for Reducing Packet Loss in Optical Burst-

Switched Networks," SPIE/Kluwer, Optical Networks Magazine, vol. 4, no. 6, pp. 81-89, Nov./Dec., 2003. • V. M. Vokkarane et al, “Channel Scheduling Algorithms using Burst Segmentation and FDLs for Optical Burst-

Switched Networks,” Proceedings, IEEE ICC 2003, Anchorage, AK, May 2003Quality-of-Service:

• M. Yoo, C. Qiao and S. Dixit, "QoS Performance in IP over WDM Networks," IEEE JSAC, Vol. 18, No. 10, pp. 2062-2071, Oct. 2000.

• V.M. Vokkarane and J.P. Jue, "Prioritized Burst Segmentation and Composite Burst Assembly Techniques for QoS Support in Optical Burst-Switched Networks," IEEE JSAC, vol. 21, no. 7, pp. 1198-1209, Sept. 2003.

• Q. Zhang, V.M. Vokkarane, J.P. Jue, and B. Chen, "Absolute QoS Differentiation in Optical Burst-Switched Networks," UTD Technical Report UTDCS-45-03, Oct. 2003.