QoS routing in DWDMOptical Packet Networks

W. CerroniCNIT – Bologna Research Unit, ITALY

F. Callegati, G. Muretto, C. Raffaelli, P. ZaffoniDEIS – University of Bologna, ITALY

2

Optical Packet Switching (OPS)

Legacy Networks

Optical Packet RoutersLong-term technical solution

Edge Systems

Statistical multiplexing of different flows on each wavelengthRouting-based signalling

WDM Links

3

Store & forward with optical buffers

• Realized with B Fiber Delay Lines (FDL):– the delay must be chosen at packet arrival– packets are delayed until the output wavelength is available– available delays are consecutive multiples of the delay unit D

(different choices are also possible)– packets are lost when the buffer is full, i.e. the required delay is

larger than the maximum delay achievable DM = (B -1)D

0

D

(B -1)Dt0 t0+D t0+2D t0+(B -1)D

t0

…

4

Wavelength and Delay Selection Problem

• The forwarding algorithm determines: – the output fiber (from the routing table) and the output wavelength– if all wavelengths are busy:

• packet delayed in FDL buffer or • packet dropped, because the required delay is not available

• Wavelength and Delay Selection (WDS)– choose the wavelength according to availability in time– choose the delay in order to minimize the gaps between buffered

packets and maximize the wavelength utilization

D GAPt0 t0+D t0+2D t0+4Dt0+3D

λ1λ2λ3λ4

D

Pac

ket L

oss

Pro

babi

lity

5

Algorithms for adaptive routing in OPS

• Routing algorithms can be– static: routing tables change only when the topology

changes– adaptive: routing tables include alternatives to

shortest path depending on the congestion state of the network

• DWDM OPS network must– combine the flexibility of adaptive routing with the

resources made available by WDM– design routing procedures outperforming the

conventional shortest path routing– provide QoS differentiation at the routing level

6

Algorithms for adaptive routing in OPS

• Solutions for undifferentiated traffic presented at ONDM ‘04

• The routing algorithm provides:– a default path: shortest path used as a first chance– a few alternative paths: used in case the default is congested

• Traffic flows are routed according to different path selection strategies (increasing complexity):– SL (Single Link): only the default path is used (static routing)– SA (Single Alternative): a single alternative path is used

(ineffective: performance close to SL)– MA (Multiple Alternative): more than one alternative paths are

used

7

Multiple alternative routing

WDS

default path

alternativepath 1

SOURCE

DESTINATION

In case the default path is congested, the best wavelength is chosen on one of the alternative path

alternativepath 2

8

Multiple alternative routing

default path

alternativepath 1

SOURCE

DESTINATION

In case the default path is congested, the best wavelength is chosen on one of the alternative path

WDS

congestion

alternativepath 2

9

Multiple alternative routing

default path

alternativepath 1

SOURCE

DESTINATION

In case the default path is congested, the best wavelength is chosen on one of the alternative path

congestion

alternativepath 2

WDS

10

QoS differentiation at the routing level

• Due to FDL buffering constraints, traditional priority queuing and scheduling techniques are not feasible– QoS differentiation at the OPS node level possible through

resource partitioning (cf. JSAC Jan. 2000, Comp. Net. 15/03/2004)

• Integration of QoS management into adaptive routing algorithms– aggregate QoS classes (sort of DiffServ approach)– simple set-up: 2 priority classes– High-Priority (HP) traffic: always routed along the shortest path

(SL) using resource partitioning• limited packet loss• limited delay and packet jitter

– Low-Priority (LP) traffic: two options• always routed along the shortest path (SL) using available resources• overflow traffic re-routed to alternative paths (MA)

11

Resource partitioning: FIX strategy

• H wavelengths out of W are reserved to HP traffic– the remaining W – H wavelengths are shared between HP and

LP traffic

• The reserved wavelengths are fixed– e.g. H=2 à ?1 and ?2 are reserved– when ?1 and ?2 are busy, HP and LP experience the same loss

t0 t0+D t0+2D t0+4Dt0+3D

λ1λ2λ3λ4

t0 t0+D t0+2D t0+4Dt0+3D

λ1λ2λ3λ4

LPHP

12

Resource partitioning: RES strategy

• H wavelengths out of W are reserved to HP traffic– the remaining W – H wavelengths are shared between HP and

LP traffic

• Any H wavelengths are reserved based on the actual occupancy– e.g. H=2 à LP packets are allowed as long as more than 2

wavelengths are available

t0 t0+D t0+2D t0+4Dt0+3D

λ1λ2λ3λ4

t0 t0+D t0+2D t0+4Dt0+3D

λ1λ2λ3λ4

LPHP

13

Network performance evaluation

•– 5 nodes, 12 links– W = 16 wavelengths per link– connectionless transfer mode– Poisson arrivals at each node– exponential packet size (optimal

average value according to node dimensioning)

– traffic distribution on the network:• balanced (B): each wavelength is

loaded by 0.8 à traffic generated accordingly

• unbalanced (U): each node generates the same traffic à wavelengths on different links carry different loads (max. 0.8)

0

1 4

32

First evaluation on a simple topology

14

Resource partitioning: FIX vs. RES

1e-08

1e-07

1e-06

1e-05

1e-04

1e-03

1e-02

1e-01

1

0 1 2 3 4 5

Pac

ketl

oss

prob

abili

ty

No. of reserved wavelengths (H)

FIX LP

FIX HPRES LP

RES HP

à From now on, always adopt the RES strategy

15

Performance for undifferentiated traffic

B-SL B-MA U-SL U-MA0

2E-04

4E-04

6E-04

8E-04

1E-03

1.2E-03

Pac

ketl

oss

prob

abili

ty

B = balanced trafficU = unbalanced traffic

SL = shortest path onlyMA = multiple paths

Not very high improvement due to the limited number of paths in the test networkà MA proves to be more effective on larger networks (cf. ONDM ’04)

16

The impact of resource partitioning

1 2 3 4No. of reserved wavelengths

B-SL HP

B-MA HP

B-SL LP

B-MA LP

1e-08

1e-07

1e-06

1e-05

1e-04

1e-03

1e-02

1e-01

1

Pac

ket l

oss

prob

abili

ty

1e-08

1e-07

1e-06

1e-05

1e-04

1e-03

1e-02

1e-01

1

Pac

ket l

oss

prob

abili

ty

20 30 40 50Percentage of HP traffic

B-SL HP

B-MA HP

B-SL LP

B-MA LP

10

Percentage of HP traffic = 20% No. of reserved wavelengths = 3

B = balanced trafficSL/MA = routing policy adopted for LP traffic (HP uses always SL)

Accurate HP dimensioning gives a good degree of traffic differentiationà LP routing policy does not affect HPà LP performance slightly affected by HP dimensioning (within the

range considered)

17

Resources needed for guaranteed HP loss

0

10

20

30

40

50

60

70

80

0

Per

cent

age

of H

P c

lass

traf

fic

5 10 15 20 25 30 35Percentage of reserved wavelengths

PLP = 10-6

PLP = 10-5PLP = 10-4

1e-04

1e-03

1e-02

1e-01

1

LP c

lass

pack

et lo

ss p

roba

bilit

y

20 30Percentage of HP class traffic

10

PLP = 10-6PLP = 10-5PLP = 10-4

balanced trafficSL routing policy adopted for LP traffic

PLP = loss probability for HP packets

à HP dimensioning required for a given PLP

18

Network design for unbalanced traffic

• Each node generates the same amount of traffic, uniformly distributed towards the other nodes– each link is subject to a different load, depending on the traffic

matrix (assuming shortest paths only)– no reason to waste resources on underloaded links– provide the links with the resources (i.e. no. of wavelengths)

required to obtain a given average load per wavelength

=

h wavelengtper load requiredlink the on load total

hs wavelengtof no.

à Drawback: resource partitioning is not very effective on links with a small no. of wavelengths

19

Performance of unbalanced network design

10 15 20 25 30 35 40Percentage of reserved wavelenghts

D-SL HP D-MA HP

D-SL LP

D-MA LP

1e-08

1e-07

1e-06

1e-05

1e-04

1e-03

1e-02

1e-01

1

Pac

ketl

oss

prob

abili

ty

Percentage of HP traffic = 20%

D = network design under unbalanced trafficSL/MA = routing policy adopted for LP traffic (HP uses always SL)Network cost (in terms of total no. of wavelengths) as close as possible to the balanced case (i.e. 16*12=192) à amount of traffic generated at each node accordingly

Resulting resource distribution:à from 7 to 28 wavelengths/link

Resulting loss distribution:à from 10-2 to 10-6

à Heavy unfairness

20

Design with constraint on packet loss

• Fix a maximum acceptable value of packet loss probability• Perform the design procedure and evaluate packet loss• Iterate the simulation by increasing the no. of wavelengths on links

with loss exceeding the acceptable value, until the loss constraint is satisfied on all links

0

510

15

2025

3035

4045

50Percentage of additional wavelengths

0 1 2 3 4 5 6 7 8 9 10 11Network links

PLP = 10-5

PLP = 10-3

à still unfair loss distribution among links

à at least the loss constraint is satisfied

à increased overall network costà increased simulation time

21

Conclusions

• QoS differentiation in DWDM OPS networks achieved by exploiting resource partitioning and adaptive routing

• Iterative design procedure to satisfy loss constraints

• Open issues:– need for extensive simulations on large networks to

prove the effectiveness of MA approach– evaluation of the impact of adaptive routing on packet

delay and sequence– extension to a connection-oriented transfer mode and

impact on virtual circuits routing