OPTICAL BURST SWITCHING

1. OPTICAL BURST SWITCHING Carla Raffaelli

2. Main limitations of optical packet switching • Packet by packet header processing • Critical timing during packet forwarding • o/e/o header conversion at each node • Node by node contention resolution • Limited capacity optical buffers • Wavelength converters header payload Tp

3. Main limitations of optical circuit switching • A RTT is needed to set up a circuit • Intermediate switches are busy for longer times with respect to the message duration • It is efficient when message transmission is much longer then the RTT for cirsuit set up. 1 RTT to set up a circuit

4. Burst switching principles • A technique which combines the advantages of both circuit and packet switching • Out-of-band signaling is used to configure switching nodes along the burst forwarding path • No confirmation of path set up is given to the initiating part • Reservation algorithms are needed to set up switch resources • Scheduling algorithms are needed to choose wavelength • Burst storaging is typically made at network edges • Burst are typically asynchronous and variable length

5. JIT:Just-in-time reservation • A set up message is sent to configure the cross-connects along the path lungo il percorso • Cross connect configuration takes place during the propagation of the path set up message through the network • Data transmission begins after a suitable delay with respect to set up emission and in any case no confirmation is given to the sender • Network resources are esplicitly released by means of a release message sent after the end of the burst

6. Time diagram of JIT reservation • Explicit setup and release • Some network resource are unused for long time

7. Just-Enough-Time (JET) reservation • An offset time between CP (control packet) and burst • CP carries the burst length information • Facilitates delayed reservation (DR) for intelligent, efficient allocation of BW, including look-ahead scheduling.

8. Time diagram of JET reservation • Estimated setup and release • Switch resources are reserved for the minimum time needed to perform burst forwarding (JET:Just Enough Time) • The best resource utilization is paid with more complex scheduling algorithms

9. Comparison OBS/OPS/OCS

10. Wavelength scheduling problem • How to assign an output wavelength to a burst? • These algorithm are typically coupled with techniques to support quality of service ? l1 l2 l3 l4

11. Burst scheduling • Two categories of scheduling algorithms • With void filling • Search for intervals that fit burst length • Can minimizes the starting void (Min-SV or LAUC-VF) or the ending void (Min-EV) etc.. [Xu et.al. Infocom’03] • Without void filling • Only use open interval (also called Horizon/LAUC) [Turner’99]

12. Scheduling Algorithms

13. Burst assembly at edge node • Legacy network interfacing • Burst classification (address, QoS, …) • Burst assembly (per flow, mixed flow…) • Burst transmissionon optical Interworking Unit Input Unit Transmission Unit N Dest CoS ass S

14. Per-flow aggregation • Ingress per-flow queuing • Optical packet assembled with segments of the same flow • An assembly time-out for each active flow is needed • High complexity of assembly mechanism Assembly Queues F1 F2 Transmission Queue Fn

15. Mixed-flow aggregation • TCP segments from different flows and with the same optical destination address aggregated in the same optical packet • Only one assembly time-out is needed • Low complexity of the assembly mechanism F1 F2 Assembly Queue Transmission Queue Fn

16. Basic TCP control functions • Flow control • the TCP window size is used to prevent the sender from flooding the receiver • Congestion control • TCP window is dynamically updated in relation to the network state as perceived by the sender

17. TCP Congestion Control • Congestion window adaptation • Additive increase/multiplicative decrease After timeout 25 cwnd = 20 20 15 Congestion window (segments) 10 ssthresh = 10 ssthresh = 8 5 0 0 3 6 9 12 15 20 22 25 Time (round trips)

18. TCP loss detection • Retransmission Time Out (RTO) • Triple duplicate ACKs (TCP Reno) • Fast retransmit/Fast recovery fast retransmit triple dup acks Congestion window (segments) Time (s)

19. Impact of OBS network on TCP • Edge node • Assembly algorithms • Mixed flow/ per flow • Time out, threshold-based • Core node • Scheduling algorithms • Contention resolution schemes • Wavelength domain • Time domain • Network • Routing algorithms • Deflection routing • QoS routing TCP performance (throughput, fairness) are influenced by OBS networks

20. Classes of TCP sources • Fast source • All segments of the maximum window are emitted in Tb • Slow sources • At most one segment emitted in Tb • Medium source Ba(bit/s) access network rate, L (bit) segment length, Wm (bit) maximum window size, Tb (s) burstification time out Optical bursts

21. Burst loss • Multiple segment losses • Depend on the level of aggregation of segments in a burst • Retransmission time out is the main indication of loss for fast sources • Congestion window shrinks to 1 MSS when a burst is lost • Slow sources recover mainly by means of fast recovery/fast restransmit Burst loss is a consequence of contention in core nodes

22. Correlation benefit • Effect related to correlated segment delivery • Fast/medium source • Fast window reopening is due to concentrated acks • Congestion window quickly reaches its maximum value

23. TCP Source TCP Receiver Optical Core Switch ER IR Ba Bo Access link Optical link Optical Core Network IWU IWU TCP send rate for different sources • Bo = 2.5 Gb/sec • RTT =600 ms • Max window size Wmax=128 MSS • MSS = 512 bytes • Tb= 3 ms 1.0e+06 Ba=200 Mb/s (fast) Ba=100 Mb/s (medium) 8.0e+05 Ba=1 Mb/s (slow) 6.0e+05 TCP Send rate (bit/sec) 4.0e+05 More segments are in a burst, the higher the TCP performance 2.0e+05 0 0.0001 0.001 0.01 0.1 Burst Loss Probability

24. Variable delay • Delay due to burst assembly task • Edge architecture • Algorithm employed (Time out, threshold-based,…) • Delay due to the presence of FDLs • Core architecture • Delay due to the scheduling algorithm

25. Modeling TCP throughput • A simple model is able to calculate throughput as a function of burst loss probability p • p is a Bernoulli r.v. • Aggregation is accounted through the average number of segment in a burst E[N] • The average TCP throuhput is calculated starting from the formula • where E[Y] is the average number of segments transmitted in bursts during during the interval I between two time out periods and E[R] is the average number of segments transmitted during the time out period ZTO • The result is p is due to losses arising in the network and in the core nodes

26. Core switch architecture

27. 1e+006 800000 600000 TCP Send Rate (bit/sec) 400000 N=8 N=16 N=32 200000 N=64 N=96 N=128 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Offered Load, Aw Effects of core node design TCP send rate as a function of offered load per wavelength • N = wavelengths per fibre • WC = 64 • Tb = 3 ms

28. Core node design optimization 0.6 wc=8 wc=32 0.5 • Maximum load per wavelength at 95% of maximum TCP send rate • WC = no. of wavelength converters; • N = wavelengths per fibre; • Tb = 3 ms; wc=64 wc=512 0.4 Offered Load, A0 0.3 0.2 0.1 0 0 20 40 60 80 100 120 # of Wavelenghts per Fiber, N

29. References • A.Detti, M. Listanti, “Impact of Segment Aggregation on TCP Reno Flows in Optical Burst Switching Networks”, Proc. INFOCOM 2002, Vol.3 pp1803-1812. • J. Padhye, V. Firoiu, D.F. Towsley, J. F. Kurose, “Modeling TCP Reno Performance: a simple model and its empirical validation”, IEEE/ACM Transaction on Networking, Vol. 8, No.2, pp.133-144, April 2000. • J. He, S.-H. G. Chan, “TCP and UDP Performance for Internet over Optical Packet-Switched Networks”, Proc. of ICC 2003, pp.1350-1354.