1 / 28

Optical Burst Switching for Unbuffered Span-Constrained Networks

Optical Burst Switching for Unbuffered Span-Constrained Networks. R. Rajaduray, D. J. Blumenthal Workshop on Optical Burst Switching Oct 3 2005 BroadNets Conference Optical Communications and Photonic Networks Department of Electrical and Computer Engineering

vail
Télécharger la présentation

Optical Burst Switching for Unbuffered Span-Constrained Networks

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Optical Burst Switching for Unbuffered Span-Constrained Networks R. Rajaduray, D. J. Blumenthal Workshop on Optical Burst Switching Oct 3 2005 BroadNets Conference Optical Communications and Photonic Networks Department of Electrical and Computer Engineering University of California, Santa Barbara This work was funded by an Intel Grant

  2. Acknowledgements • Technical assistance provided by: • Nicholas Burgan-Illig (part of the Research Internship in Science and Engineering program) • Dr Shlomo Ovadia and Dr Mario Paniccia at Intel Corporation

  3. Outline • Introduction • LER Architecture • Predictive Bandwidth Reservation

  4. Why consider all-optical networks • Graph shows traffic on all US backbones • Growth of 100% per year • Possible solution: All-optical networks • Likely that all-optical networks will be unbuffered • Cost, complexity • No buffering for congestion control or contention resolution Andrew Odlyzko, Crisis and mythology in the telecom world, CommsDaySummit, Sydney, Australia, Feb. 16, 2004.

  5. Why OBS for Unbuffered Span-Constrained Networks? 1 • OBS: excellent choice for unbuffered all-optical networks • Delayed reservation1 feature improves performance • Lower processing overhead per unit data compared to OPS • More efficient bandwidth utilization compared to OCS • Unbuffered network performance degrades with hop count2 • Potential area of application: span-constrained networks • Span-constrained networks have rapidly grown in importance3 • Examples: Local Area Networks (LANs) and Storage Area Networks (SANs) • Hop count ≤ 4, Span distance ≤ 10 km 1Yoo et al SPIE Vol. 3230, pp. 79-90, Nov. 1997 2 M. Yoo et al OFC 1999 p. 177-79 3 S. Ovadia et al IEEE Communications Magazine vol 41, issue 11, Nov. 2003 pp. S24 - S32

  6. OBS Unbuffered Span-Constrained Network Reference Model • 3 level hierarchy • Electronic terminals generate packets at lower bit rate • Concentrator multiplexes and aggregates electronic packets into higher bit rate optical bursts • Optical bursts switched through central subnetwork (core network) • Developed reference model based on hierarchy • Label Edge Router (LER) acts as concentrator and aggregator • Core network of unbuffered Photonic Burst Switches (PBSs) • PBSs forward bursts from ingress to egress edge router • Just Enough Time (JET)1 signaling used to set up path for bursts • Schedules using Latest Available Used Channel with Void Filling (LAUC-VF)2 1 Qiao et al JHSN vol 8 pp. 69-84 1999 2 Xiong et al IEEE JSAC vol 18 no 10 pp. 1838–51 Oct 2000

  7. Outline of Presentation • Demonstrate LER architecture • Chose design parameters from representative set which led to lowest latency • Show that these parameters do not limit throughput • Compare this LER performance to similar for OPS system • Demonstrate contention avoidance scheme for core called Predictive Bandwidth Reservation (PBR) • PBR is designed to operate in parallel with LER • Show thatit is able to reduce loss in network

  8. Outline • Introduction • LER Architecture • Predictive Bandwidth Reservation

  9. Label Edge Router (LER) in OBS Unbuffered Network • Label Edge Router (LER) acts as aggregator and concentrator • Latency can be significant1 • Throughput may be limited due to LER • Designed LER architecture: • Chose a combination of parameters from a set of combinations which led to lowest latency • Ensured that choice did not limit throughput • Compared performance to LER for Optical Packet Switched (OPS) system • LER for OPS system had throughput limitation due to delayed reservation 1Rajaduray et al IEEE J Lightwave Tech pp. 2693-2705 vol 22 no 11

  10. OBS LER Architecture Stage 2 Stage 3 Stage 4 Stage 1 (1 Gb/s) (10 Gb/s) 1 × 10 Gb/s Optical Output 1 1 7 10 Output TX Queue 7 Burst Queues 7 Burst Assembly Units • Stage 1 (Sorting Router): Packets from 10 input channels (1 Gb/s) sorted and sent to 1 of 7 output channels (10 Gb/s) feeding burst assembly units (one per destination) • iSLIP1 protocol used • Internal speed-up to 10 Gb/s from 1 Gb/s • Stage 2 (Burst Assembly): Packets are assembled into bursts within each assembly unit, and sent out • Stage 3 (Transmission Scheduling): Fully assembled bursts queued in burst queues and transmitted to output queue over switch fabric • Stage 4 (Output Transmission): Control label transmitted followed by burst an offset time after according to JET protocol. Assume processing time per-hop of 1 µs 1 N. McKeown, ”The iSLIP scheduling algorithm for input-queued switches”, IEEE/ACM Trans Networking April 1999 pp. 188–201

  11. 0 OBS LER Burst Assembly Algorithm Incoming Packets BURST ASSEMBLY Outgoing Bursts T Fmax Fmin • Assume burst assembly starts at time 0 • Burst size, minimum; and maximum threshold burst sizes are normalized to max data receivable in time T • If before T, normalized burst size ≥ normalized max threshold, send out burst • Demonstrated in animation • Soft threshold: Keep last packet which pushed burstsize over thresholdx • Hard threshold: Start a new burst with last packet • If at T: • Normalized burst size≥normalized minimum threshold, send out burst • Elsecontinue aggregating until normalized burst size≥normalized minimum threshold • Choice of burst assembly parameters impacts LER latency

  12. Choice of OBS LER Burst Assembly Parameters • Norm min threshold (Fmin) = 0.025, norm max threshold(Fmax) = 0.085 leads to lowest latency among set of design parameters (shown in graph) • Reduces burst assembly and transmission time between stages • Choice of T determines impact of self-similarity • Can affect variance of burst assembly or variance of burst size • T = 64 μs to negate impact of self-similarity

  13. LER Design Guidelines • Design of Stage 2: • Burst assembly parameters have been chosen (Fmin = 0.025, Fmax = 0.085, T = 64) • Hard threshold preferred over soft threshold algorithm • Lower mean and standard deviation of inter-stage transmission time • Design of Stage 3: • Choice made from Longest Queue First (LQF), Oldest Burst First (OBF) and Random selection • OBF leads to lowest Stage 3 waiting time variance amongst chosen techniques

  14. Comparison to Optical Packet Switch Label Edge Router (OPS LER) • Setting Fmin = 0.025; Fmax = 0.085; T = 64: no limit on OBS LER throughput • Compared OBS LER to 2-stage OPS LER • Latency of OPS LER lower (85 to 93%) than Optical Burst Switched (OBS) LER • Throughput of OPS LER limited by JET control channel overhead • For typical processing time of 1 μs; hop count and packet length distribution considered; throughput of OPS LER limited to 0.7

  15. Conclusions for Label Edge Router (LER) • Designed an LER architecture for an OBS unbuffered span-constrained network • Made choices from set of parameters • Choices made led to lowest mean or variance of latency • Choices did not limit OBS LER throughput • Compared to Optical Packet Switched LER • Latency of OPS LER is lower than minimum latency of Optical Burst Switched LER • OPS LER throughput limited due to control signaling overhead

  16. Outline • Introduction • LER Architecture • Predictive Bandwidth Reservation

  17. LER2 LER3 4 LER1 PBS2 PBS1 LER4 4 4 4 PBS3 4 4 4 4 4 4 4 LER8 LER5 PBS4 PBS5 4 LER7 LER6 Motivation for Predictive Bandwidth Reservation (PBR) • Desire high bandwidth efficiency, low loss; and low latency in network core • Statistical multiplexing using BW reservation/queueing over upstream links of routes - can increase BW efficiency (orange links in route) • Dedicated wavelengths for downstream links to reduce loss (black links in route) • Contention over upstream links may lead to loss • Contention resolution – reactive solution to contention • Contention avoidance through bandwidth reservation - proactive solution • Introduced Predictive Bandwidth Reservation (PBR) • Bandwidth reservation in parallel with edge router operation • Uses predictions made based on past statistics after monitoring LER channels

  18. DP Reserved ΔAC ΔLP 2ΔLP ΔAT ΔAT - Δoffset Basic PBR LER PBS Stage 1 Stage 2 Stage 3 Stage 4 T = 0 PBS PBR Processor Prediction using LER statistics • PBR request sent to PBS either when burst assembly starts (Stage 2) or ends (Stage 3) • Choice depends on which one minimizes latency (Stage 3 is shown) • Duration of reservation set to Fmax× T (Stage 2-initiated), or burst duration (Stage 3-initiated) • Requested start time of reservation set using predictions from past statistics after monitoring LER channels and can depend on propagation delay between LER and PBS • PBS PBR processor makes reservation in desired outgoing PBS channel at first avail time • Send confirmation to the LER • LER sets transmission deadline for burst transmission to make desired time • Control label precedes burst transmission by JET offset

  19. A Need for improved PBR techniques Burst B misses reservation t5 - ΔLP t2 - ΔLP PBS Switch Fabric LER Output 1 A B t1 t2 • Bursts only allowed to make 1 reservation • If burst misses reservation, transmitted at next available time • May collide with other burst or reservation and be dropped • Major cause of burst drop is reservation overlap • Burst A makes a reservation on output channel 1 of PBS • Burst B makes reservation on output channel 2 which overlaps burst A reservation • Burst B is delayed, misses reservation and may be dropped • Basic PBR must be improved to deal with overlapping reservation problem Output 2 B B t3 t4 t5 PBS PBR Processor

  20. Bandwidth Resource Allocation allowance Reserved T = 0 ΔLP T = 2ΔLP PBR-BRA LER PBS Stage 1 Stage 2 Stage 3 Stage 4 PBS PBR Processor ΔAC ΔAT Prediction using LER statistics ΔAT - Δoffset • Make Bandwidth Reservation Allocation (BRA) allowance for reservation overlap • Define normalized BRA duration as ΔBRN • Reservation duration after BRA allowance is (1 + ΔBRN) × (reservation duration) • Basic PBR is special case with ΔBRN = 0 • Burst loss reduced with ΔBRN, but reservation system becomes unstable • Arrival rate to reservation system exceeds service rate of reservation system • Can limit throughput of network • Need to maintain stability while reducing burst loss

  21. A PBR-BRA with Locally Aware Reservation (PBR-BRA-LAR) LER B B • Set BRA allowance accordingly to not cause instability • Assume that start time of requested reservation = start time of potential transmission • Ensure potential transmission times do not overlap • Change if necessary • Example: • Potential transmission time of burst A at LER is recorded • Burst B intends to make reservation • Potential transmission overlaps potential transmission of burst A • Reservation time is changed • Can add extra backoff duration to new requested start time • Marginally improves loss, but can lead to reservation system instability t1 - ΔLP t3 - ΔLP t2 - ΔLP t4 - ΔLP t5 - ΔLP Initiated at LER

  22. t3 t4 PBR-BRA with Core Allocation after Monitoring (PBR-BRA-CAM) PBS Switch Fabric LER Output 1 A t1 t2 • Core Allocates bandwidth after Monitoring (CAM) • PBS scans previous reservations for bursts from same LER • Ensures that requestedreservation time does not overlap previous reservation • Changes requested start of reservation time if necessary • Example: • Burst A makes a reservation for time [t1, t2] on output channel 1 • Burst B intends to make reservation • Reservation time checked • Time changed if overlap with burst A reservation time • Lowest loss performance, but limited by reservation system instability • Increased BRA allowance leads to instability at lower network utilization Output 2 B t2 t5 PBS PBR Processor Initiated at PBS

  23. LER2 LER3 4 LER1 PBS2 PBS1 LER4 4 4 4 PBS3 4 4 4 4 4 4 4 LER8 LER5 PBS4 PBS5 4 LER7 LER6 Assumptions • Use burst assembly parameters from previous section: Fmin = 0.025; Fmax = 0.085; T = 64 • Reservation initiated from Stage 3 over 2nd link of route (Example of red link of route, route links ≥ 3) • Bursts can only make 1 reservation • If burst is late, transmit at next available time • If burst is delayed and collides with another reservation slot on same link, the burst is dropped

  24. Comparison of Loss across Techniques • Graph shows comparison of no PBR, basic PBR and improved PBR techniques • Increasing BRA allowance improves performance (compare ΔBRN = 0 vs ΔBRN = 2) • Improvement limited by instability • Using LAR improves performance • Marginal improvement for backoff factor ΔGN > 0 • PBR-BRA-CAM offers burst loss ≤ 0.1% • Increased BRA allowance improves performance, but instability sets in earlier • For BRA allowance = 0, reservation system stable until region of ICU = 0.8 to 0.9 • PBR-BRA-CAM allows network to meet loss requirements for VoIP (req loss = 1%) and multimedia (req loss = 10%) at 80% utilization • Without PBR-BRA-CAM, utilization is limited to 4% for VoIP and 40% for multimedia

  25. Comparison of Loss and Latency • For ICU = 0.4, possible to trade off latency for loss using PBR • For ICU = 0.8, trade-off is only possible for PBR-BRA-CAM • Overlapping reservation problem insufficiently handled by basic PBR • At ICU = 0.9, OBS without PBR has best throughput and latency performance

  26. Conclusions from Predictive Bandwidth Reservation (PBR) Work • Demonstrated contention avoidance using bandwidth reservation (PBR) • Demonstrated improvements to overcome overlapping reservation problem • PBR-BRA-CAM can result in loss ≤ 0.1% up to 80% utilization • Enables network to meet loss requirements for VoIP and multimedia at 80% utilization • Increased loss comes at expense of latency • At low utilization, possible to trade of loss with latency using Basic PBR • At high utilization, tradeoff only possible with PBR-BRA-CAM

  27. Overall Conclusions • Designed an LER architecture for an unbuffered network • Chose parameters from set which led to lowest latency • No throughput limitation for parameters • LER for OPS system had throughput limitation • Demonstrated contention avoidance using Predictive Bandwidth Reservation • Designed improvements to PBR as well • Using PBR-BRA-CAM, can result in low loss • Network can meet loss req for VoIP and multimedia at 80% utilization • Increased loss comes at expense of latency

  28. THANK YOU

More Related