Trends in Optical Burst Switching

INSTITUT FÜRNACHRICHTENVERMITTLUNG

UND DATENVERARBEITUNGProf. Dr.-Ing. Dr. h. c. mult. P. J. Kühn

Universität StuttgartINSTITUT FÜR

KOMMUNIKATIONSNETZEUND RECHNERSYSTEME

Prof. Dr.-Ing. Dr. h. c. mult. P. J. Kühn

Universität Stuttgart

Trends in Optical Burst Switching

A Survey of OBS Research

Christoph [email protected]

• Motivation and Introduction of OBS• Key Building Blocks and Selected Results• Trends and Viability

E1 Workshop on OBS/OPS, Stuttgart, June 2004

Institute of Communication Networks and Computer Engineering University of Stuttgart

Internet emerged as the global platform for communication

• Exploding traffic demand

• Highly dynamic and asymmetric traffic profiles

➔ flexible and efficient transport network

• QoS demanding applications

➔ transport network should offer QoS

WDM transport introduced as cost-efficient transport layer

• Increasing discrepancy between optical transmission and electronic switching speed

➔ keep data in optical layer

• Flexible optical buffers difficult to realize

➔ avoid store-and-forward switching

• No complex optical processing

➔ process control information electronically

OBS Design Rationale


• OBS between packet and circuit switching

- support IP traffic dynamics better than OCS

- less complex optical layer than OPS

OBS Design Rationale

optical circuit switching

optical burst switching

granularity

required overprovisioning for given IP traffic pattern

switching complexity

optical packet switching


OBS Scenario

...

OBS network

core node

control-channel

data-channelsOBS link

edge node

......

• Burst assembly in edge node, mostly variable length

• WDM-based transmission

• Fast optical switch

• Separation of control and data


OBS Building Blocks

Definitions

Burst assembly assembly of client layer data into bursts

Burst reservation end-to-end burst transmission scheme

Burst scheduling assignment of resources in individual nodes

Contention resolution reaction in case of burst scheduling conflict

burst reservation

QoS / service differentiation

......

burst assembly

burst schedulingcontentionresolution


• Burst assembly triggered by time or size or both

- burst arrival process

- burst length distribution

➔ Strong impact on performance

Burst Assembly

......client layer data,

e.g., IP packets today 40…1500 Byte 32 ns…1.2 µs at 10 Gbps

optical bursts 10 kByte…10 MByte8 µs…8 ms at 10 Gbps

control unit

data

control


Burst reservation

• small vs. large pretransmission delay

• blocking in core vs. at edge

➔ Determined by application scenario: network size and burst length

Burst scheduling

• Huge amount of proposals for optimized resource utilization

• Void-filling

- Offsets produce voids ➔ void-filling algorithms

➔ complexity of void-filling is not prohibitive (2 implementations reported)

➔ performance benefit sometimes overestimated

Resource Allocation

Pretrans-mission delay

Data

Control∆2

∆1

∆3

Offset Control

end-to-end setupone-pass reservation

Data ACK


• Burst loss possible due to bufferless statistical multiplexing

• Application of OBS in high-speed metro/core networks

➔ lost data has to be retransmitted on end-to-end basis

➔ very low burst loss probability required (e.g., 10-6)

➔ Need for highly effective contention resolution

• Wavelength domain wavelength conversion

- very effective as all WDM channels shared among all bursts

- but: low burst loss probabilities only for 100≥ λs

➔ additional schemes necessary

• Time domain fiber delay lines (FDLs)

• Space domain deflection/alternative routing

• Segmentation only conflicting part of burst to be dropped

➔ Optimized combination of these schemes

Contention Resolution



• FDL buffer reservation in OBS and OPS

- Different reservation strategies, early reservation with OBS

- Joint work with Walter Cerroni in COST 266

• FDLs like offsets lead to reservations spread over time ➔ voids

- void filling can reduce this negative effects

- No improvement by void filling for offset == 0 or constant

0.0 1.0 2.0 3.0 4.0 5.0

mean basic offset / mean burst transmission time

10-3

10-2

10-1

100

bu

rst

loss

pro

ba

bili

ty

0 1 2 3 4normalized basic buffer delay

10-7

10-6

10-5

10-4

10-3

10-2

10-1

100

burs

t/pac

ket l

oss

prob

abili

ty

OBS, 4 FDLsOBS, 6 FDLsOBS, 8 FDLs


10-7

10-6

10-5

10-4

10-3

10-2

10-1

100

burs

t/pac

ket l

oss

prob

abili

tyOPS, RNFOPS, MINLOPS, MING

8 FDLs

void filling, variable offset

no void filling, variable offset


Quality of Service

QoS DifferentiationMechanisms

Additional QoS Offset

Preemption(Segmentation)

Intentional Dropping

QoS Scheduling of Ctrl. Packets

Resource Reservation

Requirements beyond differentiation capability

• Robustness wrt/ network scenario and traffic characteristics

• Low management complexity

• Minimal processing and signalling effort


Node Design

End-to-end Signaling

SwitchingTechnology

1 100 1 10 100 1 10 100nano sec micro sec milli sec

10

SOAsMEMS

TWCs

second1 10 100

Granularity burstpacket

nation

metro

campus

world

dynamic circuit

Burst Assembly edge delay

joint work with HHI

assumption:core rate approx. 10*access rate

• Granularity determines switching technology and vice versa - switching time << mean burst duration

• Delay of 80km fiber• End-to-end delay constraint: few 100ms


…more like OPS

short: 10…100 µs, some aggregation

one-pass only

λ conv., FDL, deflection routing

Trends in OBS

Future direction of OBS

typical burst length

burst reservation

contention resolution

…

… more like OCS

long: > 1 ms, extensive aggregation

one-pass or end-to-end

λ conv., defl./alternative routing

➔ Viable if solutions are consistent: architecture + technology + control


• OBS networks

- benefit from aggregation and assembly

- several architectural options available

- can offer service differentiation to client layers

• Technology

- cost and availability of switching components still unsuitable

- burst mode transmission requires changes in deployed infrastructure

• Beneficial application scenario still open

- core/transport vs. metro networks

- intensive traffic grooming towards core ➔ benefit of dynamic network?

• OBS has to fit into carriers’ world

➔ Interworking with circuit-switched photonic layer

➔ More effort towards efficiency, robustness, reduced complexity

➔ Application scenario and business model

Viability – Realization

INSTITUT FÜRNACHRICHTENVERMITTLUNG

UND DATENVERARBEITUNGProf. Dr.-Ing. Dr. h. c. mult. P. J. Kühn

Universität StuttgartINSTITUT FÜR

KOMMUNIKATIONSNETZEUND RECHNERSYSTEME

Prof. Dr.-Ing. Dr. h. c. mult. P. J. Kühn

Universität Stuttgart

Trends in Optical Burst Switching

A Survey of OBS Research

Christoph [email protected]

E1 Workshop on OBS/OPS, Stuttgart, June 2004


Provisioning Scenarios

t setup t service=

BurstPacket t service

t setup

week

hour

second

millisec.

week monthsecond hourmillisec. nanosec. microsec.

Flow

t setup t service«t setup t service≤min. t setup

λ-channel

year


Burst Assembly

Can it reduce the detrimental impact of IP traffic characteristics?

• Self-similarity on large time scales

- early work suggested YES

- recent publications prove NO for data plane

• Smoothing on smaller time scales

- consistent results show YES

➔ assembly really yields better performance

Impact on TCP performance

• In general positive due to smoothing

• Assembly timer should be adapted to TCP congestion control


Burst Reservation

• Burst loss at edge

• Dominated by propagation delay

- long-haul networks (tens of ms)

➔ Only acceptable for large bursts

• Burst loss in network

• Offset compensates processing

- Alternative: FDL in each node

➔ Mostly independent of network size and burst length

∆2

∆1

∆3

Offset

source dest

t

ControlData

one-pass reservation

source dest

t

Request

Data

Pretrans-mission delay Tp

∆2

∆1

∆3

end-to-end setup

ACK


Burst Scheduling

t

reservation horizon

t

individual reservations

Reserve a Limited Duration no void filling, e.g. LAUC, Horizon

Reserve a Fixed Durationvoid filling, e.g. LAUC-VF, JET

• Huge amount of proposals for optimization

- rearrangement of bursts, but: additional signalling needed

- gap minimization

- window-based algorithms for blocking switching matrices

• Two implementations reported for ms and s bursts

➔ complexity of JET is not prohibitive


Burst Scheduling

0.0 1.0 2.0 3.0 4.0 5.0

mean basic offset / mean burst transmission time

10-3

10-2

10-1

100

bu

rst

loss

pro

ba

bili

ty


10-7

10-6

10-5

10-4

10-3

10-2

10-1

100

burs

t/pac

ket l

oss

prob

abili

ty

OBS, 4 FDLsOBS, 6 FDLsOBS, 8 FDLs


10-7

10-6

10-5

10-4

10-3

10-2

10-1

100

burs

t/pac

ket l

oss

prob

abili

ty

OPS, RNFOPS, MINLOPS, MING

8 FDLs

void filling, variable offset

no void filling, variable offset

joint work with Walter Cerroni, University of Bologna

• Offsets lead to reservations spread over time ➔ voids

➔ void filling can reduce this negative effects

• No improvement by void filling for offset == 0 or constant

• Significant improvement only for large offset scenarios

➔ offset-based QoS scheme

➔ FDL buffer reservation


• Basic FDL delay in the order of few mean burst durations

• Combination of FDL buffers and shared converter pools

- Small conversion ratio: prefer FDL better

- Large conversion ratio: prefer conversion better


0 0.25 0.5 0.75 1conversion ratio

10-8

10-7

10-6

10-5

10-4

10-3

10-2

10-1

100

burs

t los

s pr

obab

ility

prefer FDL

0 0.25 0.5 0.75 1conversion ratio

10-8

10-7

10-6

10-5

10-4

10-3

10-2

10-1

100

burs

t los

s pr

obab

ility

prefer conversion

4 FDL’s

no FDL

2 FDL’s

0 1 2 3 4basic buffer delay / mean burst transmission time

10-5

10-4

10-3

10-2

10-1

100

burs

t los

s pr

obab

ility

1 FDL2 FDLs3 FDLs4 FDLs

0 1 2 3 4basic buffer delay / mean burst transmission time

10-5

10-4

10-3

10-2

10-1

100

burs

t los

s pr

obab

ility

8 λs, 8 output fibers, load 0.416 λs, 4 output fibers, load 0.8


Offset-based QoS

HP burst

LP burst

additional QoS-offset δQoS basic-offset

higher loss

lower loss

t


Offset-based QoS

HP burst

LP burst higher loss

lower loss

t

0 2 4 6 8 10

QoS offset / mean transmission time

10-4

10-3

10-2

10-1

burs

t los

s pr

obab

ility

of h

igh

prio

rity

clas

s

simulationanalysis

neg.-exp.

lower boundary

hyperexp. CoV 2

hyperexp. CoV 4

no differentiation

uniform [0, 2]

Pareto CoV 2

8 wavelengths, load 0.6


• High priority class depends on low priority traffic characteristics

➔ severe restrictions on burst assembly strategies

• Offset reduction due to processing leads to unintended differentiation

➔ offset-based QoS not robust in network environment


Offset-based QoS

high priority burst

low priority burst


more lost bursts

fewer lost bursts

t


• QoS offset in the order of few mean burst durations

• high priority class depends on low priority traffic characteristics

• offset reduction due to processing leads to unintended differentiation

Offset-based QoS

0 2 4 6 8 10


10-4

10-3

10-2

10-1

burs

t los

s pr

obab

ility

of h

igh

prio

rity

clas

s

simulationanalysis

neg.-exp.

lower boundary

hyperexp. CoV 2

hyperexp. CoV 4

no differentiation

uniform [0, 2]

Pareto CoV 2

0 1 2 3 4


10-6

10-5

10-4

10-3

10-2

10-1

100

burs

t los

s pr

obab

ility

base ratio 0.01base ratio 0.1base ratio 0.5

hp last hophp first

lp last hop

lp first of 2 hops

of 2 hops

8 wavelengths, load 0.6

Trends in Optical Burst Switching

Documents