Service differentiation for variable length packets in OPS with - - PowerPoint PPT Presentation

UNIVERSITEIT GENT

Service differentiation for variable length packets in OPS with recirculating FDLs

Chris Develder, Jan Cheyns Mario Pickavet, Piet Demeester

• Dept. of Information Technology (INTEC)

Ghent University - IMEC, Belgium

PS.Mo.C8, 29 Sep. 2003

• C. Develder, et al., "Service differentiation for var. length packets in OPS with recirculating FDLs"

2

Outline

• Context: the DAVID-project
• Switch architecture
• QoS approaches
• Performance criteria
• Simulation set-up
• Results
• influence of nr. of buffer ports
• influence of “class offset”
• influence of buffer delay
• influence of load
• Conclusions

PS.Mo.C8, 29 Sep. 2003

• C. Develder, et al., "Service differentiation for var. length packets in OPS with recirculating FDLs"

3

Context

• Optical packet switching
• Network
• MAN: metro rings, MAC protocol
• WAN: full-mesh, (G)MPLS-based control
• Key components
• Ring node (OPADM), Hub, Gateway, OPR

http://david.com.dtu.dk

WAN MAN

ring node IP router hub gateway packet router ring node IP router hub gateway

PS.Mo.C8, 29 Sep. 2003

• C. Develder, et al., "Service differentiation for var. length packets in OPS with recirculating FDLs"

4

Switch Architecture

• Node in core OPS network (backbone)
• Switch functionality:
• fully non-blocking switching matrix (SOA-based)
• wavelength conversion to solve contention
• FDLs to provide buffering

FDL delay = D all-optical space switch F fibers W wavelengths B buffer ports

PS.Mo.C8, 29 Sep. 2003

• C. Develder, et al., "Service differentiation for var. length packets in OPS with recirculating FDLs"

5

QoS approaches (1)

• Known approaches to provide QoS in OPS/OBS:
• resource reservation (dedicated wavelength converters, buffers);

static or dynamically

• OBS-JET with differentiated offsets
• burst segmentation: tails have larger survival chances (implicit QoS)
• intentional drops: drop low priority traffic to free resources…
• This study:
• no resource reservation, no intentional drops, no segmentation

OBS-JET: higher offset = higher priority segmentation: drop head of overlapping packet tail tail

PS.Mo.C8, 29 Sep. 2003

• C. Develder, et al., "Service differentiation for var. length packets in OPS with recirculating FDLs"

6

QoS approaches (2)

• Header offset differentiation: offset O
• well-known OBS-JET; high priority = larger offset
• high priority bursts are known to the switch longer in advance
• Look-ahead: look-ahead delay H
• no different offsets; but look-ahead delay at input: time H to “change our mind”
• Slotted control: slot time T
• headers are delayed electronically, and handled in batches each timeslot

PS.Mo.C8, 29 Sep. 2003

• C. Develder, et al., "Service differentiation for var. length packets in OPS with recirculating FDLs"

7

Scheduling algorithm

• for each arriving packet, do:
• check if there is a free wavelength on the output port it’s destined

for, using LAUC-VF

• if no free wavelength: find free buffer wavelength, using LAUC
• if packet is buffered: do not reserve output wavelength yet, but

repeat scheduling upon re-entrance of switch (PostRes)

• look-ahead:
• high priority packets may preempt low priority ones
• preempted packets are re-scheduled using same algorithm upon

time of preemption

• slotted control:
• packets scheduled at same time are sorted: first the high priority

packets

PS.Mo.C8, 29 Sep. 2003

• C. Develder, et al., "Service differentiation for var. length packets in OPS with recirculating FDLs"

8

Performance criteria

• loss rate:
• amount of data lost / amount of data sent
• fairness:
• are longer packets strongly discriminated against?

PS.Mo.C8, 29 Sep. 2003

• C. Develder, et al., "Service differentiation for var. length packets in OPS with recirculating FDLs"

9

Simulation set-up

• Parameters:
• F=6 input/output fibres
• W=8 wavelengths per i/o fibre
• B=0..64 recirculating buffer ports
• D= delay in buffer
• L= average packet length
• 40% high priority; 60% low priority
• Traffic model:
• train length: minimal L/2, average L
• train length distribution: (length - L/2) follows neg. expo distribution
• train inter-arrival: Poisson process
• uniform distribution over output fibres

W wavelengths FDL delay = D all-optical space switch F fibers B buffer ports

PS.Mo.C8, 29 Sep. 2003

• C. Develder, et al., "Service differentiation for var. length packets in OPS with recirculating FDLs"

10

Influence of nr. of buffer ports (1)

• settings:
• O=H=T=2L; buffer D=2L
• load 0.8
• loss rates:
• fairly strong class separation
• slotted control higher avg

loss and less differentiated

• no significant difference

between look-ahead and

• diff. offsets
• delays:
• slotted worse than others
• more buff ports = delay

instead of loss

1.E-06 1.E-05 1.E-04 1.E-03 1.E-02 1.E-01 1.E+00 16 32 48 64

• nr. buffer ports B

loss rate

• bs, total
• bs, high
• bs, low

lah, total lah, high lah, low slot, total slot, high slot, low

0.5 1 1.5 2 16 32 48 64

• nr. buffer ports B
• avg. nr. recircs recv
• bs, total
• bs, high
• bs, low

lah, total lah, high lah, low slot, total slot, high slot, low

low priority high priority

• verall avg

PS.Mo.C8, 29 Sep. 2003

• C. Develder, et al., "Service differentiation for var. length packets in OPS with recirculating FDLs"

11

Influence of nr. of buffer ports (2)

• unfairness:
• more pronounced for look-ahead; which
• stems from preemption: re-scheduling packets leads to less optimal

wavelength allocation (ie. worse config of “gaps” left for other packets)

0.06 0.08 0.10 0.12 0.14 0.5 1 1.5 2 2.5 3 3.5 burst size (unit L) loss rate

• ff, B=8, total

lah, B=8, total slot, B=8, total

packet size (unit L)

PS.Mo.C8, 29 Sep. 2003

• C. Develder, et al., "Service differentiation for var. length packets in OPS with recirculating FDLs"

12

Influence of class offset

• settings:
• buffer B=8, D=4L;
• load 0.8
• varying “class offset” O=H=T
• overall loss rate:
• stronger class separation: high

priority forces more low priority losses

• boundary (reached for smaller

“class offsets” with look-ahead)

• unfairness:
• significantly stronger for look-

ahead

0.06 0.07 0.08 0.09 0.10 0.11 0.12 1 2 3 4 class-offset (unit L) loss rate

• ff

lah slot

0.95 1.00 1.05 1.10 1.15 1 2 3 4 class-offset (unit L)

• avg. pktsize drop
• ff

lah slot

PS.Mo.C8, 29 Sep. 2003

• C. Develder, et al., "Service differentiation for var. length packets in OPS with recirculating FDLs"

13

Influence of load

• settings:
• load =0.1…0.9; buffer: B=0 or 8, D=2L; “class offset”: O=H=T=2L
• loss rates:
• class separation slightly diminishes for increasing load
• slotted control achieves much weaker separation; esp. if there is buffer (B=8)
• buffer very much helps to increase sustainable load

1.E-08 1.E-06 1.E-04 1.E-02 1.E+00 0.2 0.4 0.6 0.8 1 load loss rate

• bs, pri.1
• bs, pri.0

lookahead, pri.1 lookahead, pri.0 slottedctl, pri.1 slottedctl, pri.0

• bs, pri.1
• bs, pri.0

lookahead, pri.1 lookahead, pri.0 slottedctl, pri.1 slottedctl, pri.0 B=8 B=0

PS.Mo.C8, 29 Sep. 2003

• C. Develder, et al., "Service differentiation for var. length packets in OPS with recirculating FDLs"

14

Conclusions

• compared various QoS approaches
• slotted control achieves less strong separation, but
• simpler scheduling algorithm
• may be suitable for low to medium loads (cf. for load=0.5, high

priority loss ~1E-6)

• look-ahead
• achieves equally good (or slightly better) loss rates and delays as

OBS-JET with differentiated offsets

• separation limit is reached for shorter “class offsets” than OBS-JET
• induces more intra-class unfairness (i.e. stronger discrimination of

longer bursts)

UNIVERSITEIT GENT

That’s all, folks!

… thanks for your attention … any questions?