Physical Layer Performance of Optical Packet Switches: a Practical Approach

Transcription

1 Physical Layer Performance of Optical Packet Switches: a Practical Approach Ivan Aldaya, Gabriel Campuzano, Joaquin Beas, Gerardo Castanon Department of Electrical and Computer Engineering Tecnologico de Monterrey Monterrey, Mexico {campuzano, gerardo.castanon}@itesm.mx Walter Cerroni, Carla Raffaelli, Michele Savi Department of Electronics, Computer Sciences and Systems University of Bologna Bologna, Italy {walter.cerroni, carla.raffaelli, michele.savi}@unibo.it Abstract Simulation results of the OSNR performance of an optical switch/router are presented in this paper. The paper is completed with a cascadability analysis, which takes into account the optical link that joins the nodes, and some results that show the effects of the scheduling algorithm on the optical performance. I. INTRODUCTION The evolution of optical networking experienced a groundbreaking increase in the transmission capacity with the advent of Dense Wavelength Division Multiplexing (DWDM). More recently, advances in photonic technology allowed to exploit the huge bandwidth offered by the optical domain also for switching operations inside the nodes. However, today s optical networks are typically designed to be circuit-oriented and then not very suitable to highly dynamic traffic profiles. A further improvement in the aggregated network capacity is possible by migrating to more flexible paradigms such as optical packet switching (OPS) or optical burst switching (OBS) [1]. Nodes are, therefore, required to be much more dynamic to fast switch or route packets or bursts from different inputs to the required output [2]. Performing the switching in the electrical domain results in costly and power-consuming equipment, as well as not very scalable switches/routers. A proposed alternative is to convert only the header to the electrical domain as the header includes routing information to be processed and can be transmitted at a lower bit rate while keeping the data payload in the optical domain, thus exploiting the high bandwidth offered by the optical devices. On the other hand, basic optical switches are not capable of performing 3R signal regeneration (synchronization, amplification and reshaping) in the same way as their electrical counterparts do. Therefore, signal impairments accumulate from node to node along the routing path. As a consequence, it is extremely important to assess the effects of the optical switching operations on the optical signal quality, with two main objectives: to analyze the impact of the forwarding algorithms on the signal degradation/regeneration and to assess the scalability of the network in terms of number of optical switching nodes to be crossed on a path. Previous studies have been devoted to the impact of optical switching on the signal quality. However, most of them focused on single parts or devices within the switch/router; e.g., [3] deals with passive components but it does not consider any wavelength conversion. Some others made an analysis of the whole system like [4], where the physical performance analysis of a circuit-oriented switch is done. In [5] the optical path of a switch is studied, but input requirements and scalability are not considered. This paper presents the physical layer assessment study of an optical switching node, carried out as part of an experimental testbed aimed at implementing a software router emulator of an optical router as a whole [6]. Differently from other simulation-based studies, an optical router emulator must provide a detailed characterization of both the control and data planes in order to represent a valid substitute for a much more expensive photonic testbed. For this reason, the analysis of the physical layer discussed here was performed assuming commercially available devices, whenever possible. Once all the components of the optical data plane were defined, the impact of the switching algorithm on the signal quality and the node cascadability were analyzed. At this point, only amplitude modulated signal is considered since phase modulted signal are not suitable to wavelength conversion based on cross-phase modulation. Regarding the cost and consumption assessments, they are out of the scope of this work. The paper is organized as follows. Section II introduces the optical switch architecture under investigation. Section III presents the optical link model used to connect different nodes. Then, performance results are presented in Section IV. Finally, Section V concludes the paper. II. OPTICAL SWITCH ARCHITECTURE Figure 1 shows the switching fabric analyzed in this work: an example of Broadcast-&-Select architecture [7] interconnecting N input fibers to N output fibers, each carrying W wavelengths. Optical components include Semiconductor Optical Amplifiers (SOAs), EDFA, MUX/DEMUXes, splitters/couplers etc. SOAs are used as on/off switching gates as well as amplifiers. Input Fiber Delay Lines (FDLs) are used to delay the optical signal for the amount of time required to perform the scheduling functions and to configure the switching devices. This architecture can be used to implement

2 Fig. 1. Optical switching fabric with N = 2 input fibers carrying W =4 wavelengths each. The fabric is equipped with an input conversion stage consisting of a pool of W converters per input fiber. a hybrid packet/circuit switching node, as required by the programmable optical router features [6]. The presence of wavelength converters (WCs) allows packet contentions to be managed in the wavelength domain. The path of an optical signal traversing the switching fabric is detailed in Fig. 2, which includes the numbering of the different stages being crossed. A 0.2 db attenuator representing connection loss has been included in each stage if necessary. Stage 1 includes a multichannel EDFA amplifier, required to satisfy the power requirements of the wavelength conversion stages. Stage 2 is needed to demultiplex the WDM input signal. The input FDL in stage 3 is very short, so a 3 db power penalty is assumed to take dispersion and other effects into account. Stage 4 is a 1:2 splitter, needed to split the signal between the two sub-paths available at stages 5 and 6, depending on whether wavelength conversion is required (5a, 6a) or not (5b, 6b). In case of conversion, stages 5a and 6a include a WC and an SOA amplifier, whereas when conversion is not required stages 5b and 6b consist of an attenuator and an SOA switch. The design of these sub-paths reflects the physical characteristics of the WC assumed here (described in detail in [8]), which has two drawbacks: (i) the output power is very low, so an SOA amplifier is needed at its output; (ii) while such a WC generates the signal on the desired output wavelength, it does not block the (fixed) input wavelength signal, which must be filtered out. The lack of commercially available fast tunable filters (AOTF are only available as laboratory or instrumentation devices) makes necessary the use of a notch filter at the output. The notch filter is not represented in Fig. 2 since it is assumed to introduce negligible effects on the other wavelengths. Since the notch filter always blocks the input wavelength, the alternative sub-path must be used whenever the input signal does not need to be converted. The attenuator at stage 5b is required to reduce the power amplified by the EDFA at stage 1 in order not to damage the SOA switch at stage 6b. When the signal does not need wavelength conversion, this SOA switch is turned to the on state whereas the SOA amplifier at stages 6a is turned off. In case the SOA used as amplifier is to slow to switch off, an additional SOA switch can be include without affecting significantly the OSNR. In the other case, the SOA switch at stage 6b is turned off. The combiner at stage 7 is used to merge the two sub-paths. Another drawback related to the lack of tunable filtering after the WC is caused by the presence, when conversion is performed, of spurious components produced on the other wavelengths. This effect is taken into account at stage 8 by including an additional noise source (Nadd-WCnoise) proportional to the number of wavelength conversions simultaneously performed in other channels within the same WC pool. The combiner at stage 9, the EDFA at stage 10 and the splitter at stage 11 are used to broadcast the signal to each output fiber with enough power. Then the demultiplexer at stage 12 and the SOA switch at stage 13 are used to select the desired output wavelength, which is then multiplexed with the others from the same input at stage 14 and finally combined with the ones from other inputs at stage 15. The interference of in-band crosstalk, generated by other SOA switches due to the finite extinction ratio, is taken into account at the output of the path at stage 16: the worst case is considered here, assuming that there is always interference from the other channels. The WC is the only optical device which is not available on the market yet. Therefore, as a reference model for the physical characterization, the all-active Mach-Zender interferometer WC proposed in [8] was assumed. This technology, based on XPM, was chosen due to its integration capability and its good performance for intensity modulated formats [9]. All the other passive and active optical devices are commercially available and were characterized in terms of power loss and noise by interpolation of values presented in their data sheets, following the approach presented in [3]. This was done with the aim to obtain a realistic representation of the devices. In particular, the power loss introduced by passive devices (splitter, coupler, MUX/DEMUX) was evaluated taking insertion loss, nonuniformity and polarization dependence into account. As for the active devices (SOA, EDFA and WC), the Amplified Spontaneous Emission (ASE) noise was considered, as well as, the the input and output power constraints. III. OPTICAL LINK Since optical fibers introduce attenuation, chromatic dispersion and non-linear effects, links must be designed to overcome these impairments. The most common approach is to divide a link in different spans and to in-line compensate the dispersion and the attenuation of each of them [10]. In addition, it is a common practice to compensate a certain

3 Fig. 2. Physical paths in the optical data plane of the proposed architecture. Fig. 3. Link scheme with precompensation/postcompensation. TABLE I LINK PARAMETERS General parameters Bit rate 10Gbps Modulation format OOK NRZ Total input power to SMF 0dBm Total input power to DCF 5dBm Number of stages (N) 5 Precompensation percentage 30% Number of channels 4 Channel spacing 100GHz Emission central wavelength 1550nm Amplifier parameters Noise figure 5dB Fiber parameters SMF Span length 100km SMF Attenuation 0.2dB/km SMF Dispersion (D) 16.8ps/(nm km) SMF Dispersion Slope (S) 0.057ps/(nm 2 km) DCF Span length 21km DCF Attenuation 0.45dB/km DCF Dispersion (D) 80ps/(nm km) DCF Dispersion Slope (S) 0.23ps/(nm 2 km) amount, 30% for instance, of the chromatic dispersion at the beginning of the link, since this avoids possible resonances in the link. Figure 3 represents a common link design approach, where a pre-compensation stage includes an optical amplifier (EDFA) followed by a Dispersion Compensating Fiber (DCF) with negative chromatic dispersion obtained from [10]. Then N stages follow, each one including a standard Single Mode Fiber (SMF), a DCF and the required EDFAs to compensate the losses. The link ends with a post-compensation DCF and a final amplifier. Non-linear effects are avoided by limiting the total input power to the SMF and the DCF to 0 dbm and 5 dbm, respectively. Other characteristics of the link are listed in Table I. In [12] an analytical assessment of the optical link is performed paying special attention to the chromatic disper- Fig. 4. Evolution of the OSNR, chromatic dispersion, and power for a 5-span link. Fig. 5. OSNR vs. actual link distance for different span lengths (50 to 100 km). sion. However, a simpler approach can be adopted as a first approximation analysis. As a matter of fact, the equivalent OSNR (Optical Signal to Noise Ratio) is obtained from the effective noise figure [10] which is the result of applying the Friis noise formula. An additional OSNR penalty due to the non-compensated chromatic dispersion, non-linearities and filter and component impairments is included [11]. Fig.4 displays the results of Matlab simulations evaluating the set of equations in [10] for the equivalent OSNR, cumulated dispersion and power level along the actual link distance, taking into account the length of SMF and DCF spans, using the formulas in [10]. Note that DCF spans do not contribute to the actual link distance. In the same way, the OSNR vs. the actual link distance can be seen in Fig. 5 for different SMF span lengths as parameter. The equivalent link OSNR curves shown in Fig. 5 can be approximated as a function of the total link length L T, and the SMF span length L S, resulting in: OSNR link [db] =a(l S ) ln(l T )+b(l S ) (1) After performing the logarithmic approximation for each span length, the parameters of these approximations, a(l S ) and

4 b(l S ), are obtained from the curves in Fig. 5, resulting in: a(l S )= L S L S (2) b(l S )= L S L S (3) In real networks, the link output power is not strictly constant but depends on many factors such as age of components, polarization dependency, gain ripple. We assumed a ±1 db variation. For the sake of simplicity, the optical power level was considered uniformly distributed. The power variation at the output of the link is an important parameter since this may influence the performance of the downstream switch (Sec. IV). IV. RESULTS This section provides simulation results of the proposed switching architecture. The performed simulation is divided in two stages. First, the switch presented in Section II is analyzed in terms of output OSNR and power. Later different switches are interconnected considering the link model presented in Section III to make a cascadability analysis. A. Single switch/router optical signal performance In this section a router with N =2input/output fibers and W =4wavelengths is considered. As explained in Section II, the behavior of the switch/router in terms of total power transfer and OSNR degradation/regeneration highly depends on whether wavelength conversion has been performed for a given channel. In addition, the number of simultaneously conversions carried out in the other channels within the same WC pool affects the OSNR performance. In order to show this phenomenon, Fig. 6 shows the power level (upper subplot) and the OSNR (lower subplot) when traversing the different stages introduced in Fig. 2. In particular, Fig. 6(a) represents the power level and OSNR when no additional wavelength conversion is performed while Figs. 6(b), 6(c) and 6(d) correspond to the cases of a single, tow and three additional conversions, respectively. The input OSNR has been set to 20 db and the power to -6 dbm per channel. Power evolution is quite straight-forward since it is updated in each stage according to the passive component losses or amplifier gain. It can be seen a misalignment between wavelength conversion and no wavelength conversion cases only for the stages 5 to 9 due to the different paths followed. This difference is mitigated by the in-line EDFA in stage 10, which is configured in powerclamped mode and equalizes the power level at its output. An additional point to mention is the high conversion losses the wavelength converter introduces as it is shown in wavelength conversion case (after stage 5 in conversion case). Regarding the OSNR for the path, it is affected only by active optical devices and the different crosstalks. It is important to note that the noise introduced by an optical amplifier increases when the input optical power to this device decreases (high gain needed). This together with the fact that high OSNR degrades more for the same added noise power, explains the obtained results: when the OSNR is not considerable high and the power level is not low (stage 1) the OSNR degradation due to EDFA1 is not significant. In contrast, when wavelength converted channel is simulated, the power level at the output of the WC (stage 5a) is extremely low (as previously discussed) while the OSNR is high due to the OSNR regeneration capability of the WC. Under these conditions, the OSNR degradation due to the following SOA amplifier is considerable. Additional simultaneously performed wavelength conversions further reduce the OSNR (crosstalk degrades the signal in the last stage). Next, the effect of variable input power and OSNR is studied. Taking into account that the output optical power does not vary due to the switch design (the power after the final stage in always the same, see Fig. 6), the only figure of merit to display is the output OSNR. Fig. 7 shows the output OSNR in terms of the input OSNR in the worst WC induced crosstalk case, maximum (equal to 3) number of simultaneously performed wavelength conversions. Different curves correspond to different input powers per link, from dbm to dbm, that cover the whole input dynamic range of the switch. The upper subplot represents the system output OSNR when no wavelength conversion is carried out in the studied channel while the lower subplot correspond to the wavelength conversion case. As can be seen in the figure, the behavior of the switch strongly depends on whether wavelength conversion is performed or not. In case wavelength conversion is not performed (upper subplot), the system is limited by input OSNR for values up to 30 db. Here the degradation of the OSNR is not significant because the input signal is already corrupted. After this value (30 db), the output OSNR is limited to 23 db. This can be explained taking into account that in this case the output OSNR is limited by the noise added by the other wavelength conversions (crosstalk). Furthermore, the figure shows how Incrementing the input power does not improve the OSNR. Indeed, the noise introduced by amplifiers is here shadowed by the noise added by the simultaneous wavelength conversions. Instead, when wavelength conversion is performed, the output OSNR increases as the input power increases. Indeed, in this case also the noise introduced by the amplifier that follows the WC (stage 6a) contributes to the output OSNR. This noise is significant because the input power of this amplifier is very low (due to the WC output). The input power to this amplifier is linearly related to the input power to the switch, so the noise introduced by the amplifier and the output OSNR are highly influenced by the input power. Fig. 8 generalizes the results shown in Fig. 7 to different numbers of simultaneous wavelength conversions. The response to the lowest and highest input powers are represented as limiting values. The figure makes clear the effect of the number of additional wavelength conversions on the overall signal performance, the higher the number of additional wavelength conversions, the worse the obtained output OSNR. The number of wavelength conversions simultaneously carried out depends on the decisions of the applied scheduling algorithm, so it has an impact on the optical signal performance and it should be designed taking these physical effects into account.

5 Fig. 6. Power and OSNR evolution along the stages of the optical switch/router for waveltngth conversion and non-wavelength conversion. (a) No additional wavelength conversion, (b) One additional wavelength conversion, (c) two additional wavelength conversions, and (d) three additional wavelength conversions. Fig. 7. Output OSNR in terms of the input OSNR for different input power levels in the worst WC induced crosstalk case. Fig. 8. Output OSNR in terms of the input OSNR for different input power levels and different numbers of additional wavelength conversions. B. Multiple switches/routers optical signal performance This subsection brings together the results of the Subsection A of the present section and the link model proposed in Section III to analyze a simplified multiple hop link. In a first step, multiple optical switches or routers are simulated together without considering the effects of optical links. Further simplification and assumptions are made. On the one hand, it is assumed the scheduling algorithm does not implement any decision based on optimizing output OSNR or taking into account the input power or OSNR. On the other hand, the probability a channel is wavelength converted is independent of the probabilities of conversion of other channels and it is the same for all of them. Under these assumptions, the scheduling algorithm effects can be abstracted and characterized by the wavelength conversion probability, P, and the number of simultaneous wavelength conversions follows a binomial distribution, which success probability equals to the wavelength conversion probability. Fig. 9 shows the rate (over

6 Fig. 9. Rate of packets below minimim OSNR without considering optical links between the nodes. Fig. 10. Rate of packets below minimim OSNR considering optical links between the nodes. the unity scale) of packets below the required optical quality needed to ensure the bit error rate in the communication is below 10 9 ; which is 17 db according to [3]. Simulation was carried out for 5000 packets. For the lowest considered P, 0.1, the number of low quality packets remained low up to 4 consecutive hops (5 switches), then increases significantly. For other values of P, the number of packets below minimum OSNR rises earlier, four switches/routers for 0.25 and three for 0.5 and 0.75, respectively. This behavior can be explained based on the results shown in Fig. 8. If the wavelength conversion probability is low, most of the packets cross the optical system without being wavelength converted, when nowavelength conversion is done it does not matter the input OSNR, some signal degradation occurs. If the signal quality is high, the rate of packets under the limiting condition is low but as the number of crossed switches/routers increases, the quality is progressive deteriorated. With a higher conversion probability the behavior is different. For high input OSNR the quality degradation if the packet is converted is significant and the number of packets with quality above threshold decreases quickly. However, when the packets with low quality are converted to another wavelength, signal regeneration occurs. These two mechanisms compensate each other and quite stable state is achieved. The higher the value of P the faster the steady state is reached and the lower its value. In a second stage, the impairments of the optical links are considered. Fig. 10 shows the obtained results of the simulation with the same parameter as used when no fiber was included. The general tendency is very similar to that previously explained with the difference that the reached values for the non-suitable packets are significantly higher; as a result of the added link impairments. It is also important to mention that the increment is much abrupter than in the case without fiber. V. CONCLUSIONS This paper presented an analysis of the physical performance of an optical switch/router. In particular the effects of wavelength conversion have been considered, showing how they impact on the OSNR of the switch and node cascadability. In the future, the node scalability in terms of number of wavelengths per link and number of links per node should also be carefully evaluated, to understand the real throughput per node achievable without degrading the signal excessively. The paper also highlights that the decisions taken by the forwarding and control algorithms may influence the physical performance of the data plane. Therefore, a scheduling algorithm should be designed taking into account the optical layer architecture. ACKNOWLEDGMENT This work was carried out with the support of the BONE-project ( Building the Future Optical Network in Europe ), funded by the European Commission through the 7th ICT-Framework Programme. REFERENCES [1] S. J. Ben Yoo, Optical Packet and Burst Switching Technologies for the Future Photonic Internet, IEEE/OSA Journal of Lightwave Technology, Vol. 24, No. 12, pp , December [2] A. Pattavina, Architectures and performance of optical packet switching nodes for IP networks, IEEE/OSA Journal of Lightwave Technology, Vol. 23, No. 3, pp , March [3] R. Gaudino, G. A. Gavilanes Castillo, F. Neri, J. M. Finochietto, Simple Optical Fabrics for Scalable Terabit Packet Switches, Proc. of IEEE ICC 2008, pp , Beijing, China, May [4] B. Ramamurthy, D. Datta, H. Feng, J. P. Heritage, B. Mukherjee, Impact of transmission impairments on the teletraffic performance of wavelength-routed optical networks, IEEE/OSA Journal of Lightwave Technology, Vol. 17, No. 10, pp , October [5] C. Raffaelli, M. Savi, G. Tartarini, D. Visani, Physical path analysis in photonic switches with shared wavelength converters, Proc. of ICTON 2010, Munich, Germany, June [6] W. Cerroni, C. Raffaelli, M. Savi, Software Emulation of Programmable Optical Routers, Proc. of HPSR 2010, Dallas, TX, June [7] A. Stavdas C. Matrakidis, C. Politi, Migration of broadcast-and-select optical crossconnects from semi-static to dynamic reconfiguration and their physical layer modelling, Elsevier Optics Communications, Vol. 280, No. 1, pp , 1 December [8] D. Wolfson, A. Kloch, T. Fjelde, C. Janz, B. Dagens, M. Renaud, 40-Gb/s all-optical wavelength conversion, regeneration, and demultiplexing in an SOA-based all-active Mach-Zehnder interferometer, IEEE Photonics Technology Letters, Vol. 12, No. 3, pp , March [9] S. J. Ben Yoo, Wavelength conversion technologies for WDM network applications, IEEE/OSA Journal of Lightwave Technology, Vol. 14, No. 6, pp , June [10] G. P. Agrawal, Fiber-Optic Communication Systems, John Wiley, Rochester, NY, [11] 0. Vassilieva, et al., Numerical comparison of NRZ, CS-RZ and IM- DPSK formats in 43 Gbit/s WDM transmission, Proc. of LEOS th Annual Meeting of the IEEE, pp vol.2, [12] S. Pachnicke, J. Reichert, S. Spalter, E. Voges, Fast analytical assessment of the signal quality in transparent optical networks, IEEE/OSA Journal of Lightwave Technology, Vol. 24, No. 2, pp , February 2006.