Upload
independent
View
0
Download
0
Embed Size (px)
Citation preview
End-to-End Performance of HeterogeneousMulti-EPON/OBS Networks
Maurizio CasoniMaurizio Casoni(1) and Walter Cerroniand Walter Cerroni(2)
(1) Department of Information EngineeringUniversity of Modena and Reggio Emilia
Italy
(2) D.E.I.S. – University of Bologna, Italy
Maurizio Casoni ICCCN 2011, Maui, August 1, 2011 1
Outline
IntroductionOptical Burst Switching
Ethernet Passive Optical Network
EPON-OBS inter-working
Investigated scenario
Numerical resultsTCP performance
Simulation through ns2
Conclusions
Maurizio Casoni ICCCN 2011, Maui, August 1, 2011 2
Optical Networks: Evolution
DWDM techniqueTransmission rate in the range of Tbit/s
Architectural semplificationFrom IP over ATM over SONET over WDM to IP over WDM
Need to exploit in an effective way the huge transmission bandwidthwith IP traffic
Wavelength Routingall-optical data networkLow flexibility for IP traffic
Optical Packet SwitchingIdeal transfer mode for IP trafficSevere technological constraints → not feasible in the short/middle term
Optical components immature
Maurizio Casoni ICCCN 2011, Maui, August 1, 2011 3
Optical Burst Switching
Goal: better sinergy between the mature electronic technologies and the new optical tecnologies (mid-term solutions)
Switching granularity between WR and OPSBurst concept: aggregation of IP packets with common features(e.g. destination and QoS), considered as the basic optical unit
Time and space separation of data and control (header) fieldsControl packet employs dedicated channel and precedes therelative data burst
All-optical network, buffer-less and data trasparentHybrid opto-electronic network for control signals(out-of-band signaling)
Simplification of the electronic processing of the control packetsat intermediate nodesReduction of the opto-electronic functionalities required to router
Maurizio Casoni ICCCN 2011, Maui, August 1, 2011 4
Optical Burst Switching
Core nodeHeader processing Burst forwarding
BurstVariable lengthAll-optical domain for data
Edge nodeBurst AssemblyHeader generation
HeaderOut-of-band transmissionO/E/O in core nodes
Maurizio Casoni ICCCN 2011, Maui, August 1, 2011 5
Ethernet Passive Optical Network
Downstream transmissionPoint-to-multipointEach ONU selects its data
Upstream transmissionPoint-to-pointMultiple Access TDMDynamic bandwidth assignement IPACTUse of grant - request packetsLimited disciplineCycle time: Tc
Ethernet Passive Optical NetworkOptical fiberPassive componentsEthernet protocol
Maurizio Casoni ICCCN 2011, Maui, August 1, 2011 6
EPON-OBS inter-working
IP look-up Assembly algorithmTimer basedLength based....
Queuesdestinationclass-of-service
mixed flowper -flow
Maurizio Casoni ICCCN 2011, Maui, August 1, 2011 7
Investigated Scenario
ONU 1
OLT1:NSplitter/Combiner
IWUEdge OBS Node
OBS networkIWU
Sink 1 Sink 2N…
1 Gbit/s10 Gbit/s
ONU 2
ONU i
ONU N
EPON 1
ONU 1 ONU 2 ONU i ONU N
EPON 2
1 Gbit/s OLT
SourcesTCP Reno with selective ackSegment size: 512 byteCBR at 100 Mbit/s
EPON10 ONUsShared access bw: 1 Gbit/sIPACT with limited discipline
OBS networkLink with error-model: burst loss Pb
Bandwidth: 10 Gbit/sBurst assembly: timer based Tmax
Queueing: mixed flow
Sinksack transmissions
Maurizio Casoni ICCCN 2011, Maui, August 1, 2011 8
Performance Evaluation: Metrics
ThroughputMeasure of the variability of the bandwidth usage over a giventime-scaleAverage throughput: amount of successfully transmitted bytes in agiven time interval (e.g. (0,t] )
FairnessChiu/Jain’s Fairness IndexIntra-fairness indexBest intra-fairness => F=1
BBF
P
P
max,
min,intra =
Maurizio Casoni ICCCN 2011, Maui, August 1, 2011 9
Average throughput per EPON vs. Time
Cycle time: Tc1 = Tc2 =2 msTa = 0.5 - 2 msTCP segment = 512 bytesAWND = 512 segmentsBurst loss = 10-3
1. EPON show same performance2. Fairness close to 0.93. Increasing Ta, throughput increases: correlation benefit
Maurizio Casoni ICCCN 2011, Maui, August 1, 2011 10
Average EPON1-2 throughput vs. Ta
1. AW = 512 provides best results2. Correlation benefit is reduced by increasing Ta , in particular for low AW values, for longer RTTs
Cycle time: Tc1 = Tc2 =2 msTa = 0.5 - 2 msTCP segment = 512 bytesAWND = 32-1024 segmentsBWxRTT gives 400 segmentsas ideal TCP tx windowBurst loss = 10-3
Maurizio Casoni ICCCN 2011, Maui, August 1, 2011 11
Fairness vs. Ta and AW
BehaviourGenerally high valuesImprovements for higher Ta
(correlation benefit)Worse for larger Wmax
BecauseDifferent burst composition
High Ta and small AW many TCP flows
Short Ta and large AW very few TCP flows
Maurizio Casoni ICCCN 2011, Maui, August 1, 2011 12
Average EPON2 throughput vs. time
EPON1 Tc1 = 2 msTa = 2 msTCP segment = 512 bytesAWND = 512 segments
1. Best performance for Tc2 = Ta = 2 ms
Maurizio Casoni ICCCN 2011, Maui, August 1, 2011 13
Average EPON1-2 throughput vs. Tc2/Ta
Ta = 0.5 ms Ta = 2 ms
EPON1 Tc1 = 2 msAWND = 512 segments
Tc2 = Tc1Better performance for higher Tc values
EPON 1 performs better for lower Tc2
Maurizio Casoni ICCCN 2011, Maui, August 1, 2011 14
Conclusions
Performance of TCP in a hybrid multi-EPON\OBS optical network
Focus on the IWU or edge router between EPONs and OBS network
Numerical investigation has revealed:
1. A remarkable role for TCP performance by cycle time and burstassembly time
2. In particular, Ta properly set to maximize performance
3. Then, Tc has to be set accordingly for the best EPON\OBSinternetworking (e.g. Tc = Ta) for end-to-end performance
4. With a proper parameter setting, fairness is good as well
Maurizio Casoni ICCCN 2011, Maui, August 1, 2011 15
Current and Future Work
Heterogeneous traffic (UDP and TCP sources)
More than two EPONs with larger number of ONUs
Service differentiation schemes in IPACT and/or assembly
Impact on performance of long-range PONs
Maurizio Casoni ICCCN 2011, Maui, August 1, 2011 16
THANK YOU FOR YOUR ATTENTION
[email protected]@ieee.org
http://www.dii.unimo.it/casoni
… suggestions are very very welcome
Department of Information EngineeringUniversity of Modena and Reggio Emilia