Traffic and λ Grooming

Transcription

1 Traffic and λ Grooming Tibor Cinkler, Budapest University of Technology and Economics Abstract The article gives an overview of traffic grooming and lambda grooming in multilayer networks, and discusses the advantages and drawbacks of these methods as well as applications and future alternative directions. T here is no doubt that in the near future data communications will be based on optical networking. However, as the amount of traffic grows, the technology enabling optical networking develops as well. Therefore, evolution is expected in this field, not revolution. In the early phases of wavelength-division multiplexing (WDM) employment the capacity of point-to-point links was extended by carrying multiple signals over a single fiber on different wavelengths (Fig. 1). The next step was adding optical add-drop multiplexers (OADMs) to those links to be able to add and drop signals of certain wavelengths. This was followed by forming rings of OADMs, and interconnecting these rings by optical cross-connects. In these rings, the problem of protection was trivial; however, when traffic had to be routed over multiple rings, the problem of routing was not trivial, and resource usage, particularly for protection, was not optimal. Furthermore, there was a large amount of pass-through traffic at inter-ring junctions. The next step was to build mesh networks, networks of general topology with arbitrary density and arbitrary constraints on maximal shortest cycles. Dynamic or Static Network: Switched or Dedicated (Leased-Line-Like) Connections? The author is supported by Ericsson, János Bólyai Foundation, OTKA grant D42211, COST 266, and ETIK. There are networks of both types. Synchronous digital hierarchy/synchronous optical network (SDH/SONET) is a typical leased network with static dedicated connections, while asynchronous transfer mode (ATM) can be not only leased but a switched network as well. While in networks with leased connections the network operator sets up and tears down the connections in days or weeks, in switched networks the connections are set up in a few seconds through user signaling, and the resources are released immediately when not needed anymore. This leads to more efficient resource usage; however, it needs a complex signaling system. Furthermore, routing must be fast enough even in large networks; therefore, the state information of network elements has to be spread (flooded, advertised) throughout the network, and source routing should be used. There are two problems with switched networks. First, if, instead of dedicated, shared protection is chosen to save a significant amount of resources, source routing becomes very complex. Namely, the source must know not only the network topology and link states, but also the pairs of working and protection paths for all other demands! This calls for extending existing link state advertisement or flooding protocols, which results in very heavy signaling load. The other drawback of switched networks is that there is no traffic matrix given in advance; instead, the traffic pattern changes continuously. Some sequences in the traffic pattern might appear that lead to suboptimal network state, significantly reducing throughput. Typically, for steadier traffic (e.g., in long-haul transport networks that carry a huge amount of aggregated traffic), the static configuration seems to be advantageous, while in the access and metro parts of the network the per-connectionbased dynamically switched approach is more promising. Single- or Multi-Layer Networks? In single-layer switched wavelength routing dense WDM (WR-DWDM) networks we assume that as the demand arrives, the source destination pair is connected by two unidirectional lightpaths (typically along the same path) either without or with wavelength conversion, but without any interaction with the upper layers. As examples four specifications can be mentioned: ODSI UNI: Optical Domain Service Interconnect user network interface ( OIF UNI: Optical Interworking Forum ( IETF MPLambdaS: Multiprotocol lambda switching (MPλS) by Internet Engineering Task Force ( ITU-T ASON: Automatic Switched Optical Network by International Telecommunication Union Telecommunication Standardization Sector (ITU-T) ( Examples of multilayer network architectures are: Generalized multiprotocol label switching (GMPLS) proposed by IETF Automatic switched transport network (ASTN) proposed by ITU-T The idea of these activities is not only to build multiple layers one over the other, but also to define standard interfaces, protocols, and signaling systems for building up connections on demand with required traffic and quality parameters as well as for tearing them down. For example, in case of GMPLS five networking layers are identified [1]: A packet switching capable (PSC) layer that can handle packets or cells (e.g., IP, ATM, MPLS) Time-division multiplexing (TDM) capable layer that handles larger TDM frames (e.g., SDH STM-1,4,16), λ-switching capable (LSC) layer, either WR-DWDM or MPλS /03/$ IEEE IEEE Network March/April 2003

2 Waveband switching capable (WBSC) layer that handles jointly more λs of a wavelength band Fibre switching capable (FSC) layer that handles fibers with all the traffic flowing over them Some documents even mention an additional layer, the layer 2 switching capable (L2SC) one, to be used between the PSC and TDM layers, while some documents leave out the WBSC layer. There are three models regarding interoperation of these layers [2, 3]: The overlay model, where the layers operate separately. The upper layer can see the underlying network as it is configured, without having any direct influence on it. The peer model, where the layers operate jointly (i.e., any demand can be satisfied by the network layers jointly). When a routing decision has to be made the whole multilayer network is considered as one. The augmented model is something between the two above models. To be able to carry the exponentially increasing amount of traffic all fibers in cables and all wavelengths in fibers are needed. However, we still have to be thrifty, and not assign a wavelength channel with capacity of, say, 10 Gb/s to a mobile phone call with bandwidth requirement of, say, 14 kb/s, but to offer sub-lambda bandwidth granularity. Multiple layers are definitely needed, although the number of these layers should be minimized as much as possible. What Is Grooming? Instead of a definition let us see an example. If you want to fly from your home town to a far minor destination, you will probably not have a direct flight. The reasons are first, that too many planes would be needed to operate all these direct flights regularly; second, these flights would be almost empty. Both the initial investment (CAPEX) and operational expenses (OPEX) would be prohibitively high! Furthermore, if there is a direct flight but it is already fully booked, you can still fly to your destination by changing planes one or more times. This will be probably a longer route, you will need more time, but you will still be able to reach your destination on time. This leads to better flight utilization and cuts ticket costs. The idea of grooming in communication networks is the same. To be able to carry a larger amount of traffic at lower cost by the available network infrastructure, some traffic streams have to be carried jointly (i.e., have to be groomed). Grooming has already been used in frequency-division multiplexed (FDM) and plesiochronous digital hierarchy (PDH) time-division multiplexed (TDM) transmission systems, where the aim was to demultiplex the fewest higher-order signals as possible for adding or dropping channels. Some documents distinguish end-to-end (subrate) and intermediate (core) grooming. Here we will discuss the case of intermediate or core grooming, where not only the edge nodes but the internal (intermediate) nodes of the network are capable of performing grooming. Grooming in SDH/SONET Networks In SDH/SONET networks the traffic granularity is poor. There are some discrete bandwidth values defined by virtual containers (VCs) that can be chosen at certain points. These values often do not match your exact needs; you always have to round up. For example, you can choose among VC-4, VC- 3, and VC-12, where one VC-4 can accommodate 3 VC-3s or 63 VC-12s, while a VC-3 can be filled by 21 VC-12s. This is a very rigid multiplexing structure with poor granularity that leads to lower utilization of network resources. Optical network functionality Ring networks Point-to-point WDM links Figure 1. The optical networking evolution road map. Optical packet switching Optical packet switching Wavelength switching: ASON ASTN Mesh networks: OTN Time Therefore, for carrying an IP traffic stream, it is preferred not to use from end to end a single small VC, but rather to join (groom, bundle) multiple traffic streams to fill a larger VC (or even concatenated VCs) and carry them jointly along multiple hops. Recently, new enhancements have been defined that allow carrying data signals over transport networks more efficiently. These include Generic Framing Procedure (GFP), virtual concatenation, and the Link Capacity Adjustment Scheme (LCAS) [4] often referred to as new-generation SDH/SONET. GFP defines how to map/groom a wide variety (e.g., ATM, Ethernet, PPP) of data signals into the same SONET/SDH frames. Virtual concatenation allows concatenating multiple parallel transport units for transmission of these aggregated (groomed) high-bandwidth data signals, even if they go along different paths. LCAS allows the capacity shares over these links to be increased/decreased dynamically. Grooming in ATM and MPLS Networks In ATM networks the cells are labeled by two identifiers: virtual path identifier (VPI) and virtual channel identifier (VCI). The VPI defines the group of groomed connections, while the VCI identifies certain connections within these groups. In ATM networks there is the challenge to design the virtual topology (the system of virtual path connections, VPCs) that best accommodates the groomed connections (virtual channel connections, VCCs). Similarly, in MPLS networks multiple upper layer traffic streams (e.g., IP sessions) are groomed into the system of label switched paths (LSPs). Furthermore, LSPs can be embedded (nested) one into the other hierarchically, by label stacking. In contrast to label swapping, where the labels are first removed and then substituted by the new label, in case of label stacking a new lower layer label can be added over the existing one. Always the last added label is the valid label; after removing it the next label of the stack will be the valid one. The label stacking capability and the number of stacking levels depends on the actual technology for MPLS. This is referred to as hierarchical grooming. Furthermore, in case of GMPLS this is done not by a unique network technology, but by a few different technologies layer by layer. As an example Fig. 2 illustrates a transport section, where the switching/crossconnecting can be done at different layers using different technologies with different granularity at the fiber, λ, or VC-4 level. Grooming in Optical Networks We distinguish two kinds of grooming: first, when in WR- DWDM networks the virtual topology of wavelength channels is built and then the traffic streams of upper layers are groomed using electrical TDM and carried by wavelength channels [5]; and second, when the wavelengths are groomed IEEE Network March/April

3 VC-4 λ Fiber Cable Figure 2. Illustration of a multilayer transport section with grooming. TDM λxc Fibers Wavelengths Hybrid XC Figure 3. The architecture for traffic grooming in WR-DWDM networks. Wavelengths λxc FXC Cables Fibers OXC Figure 4. The architecture for λ grooming in multifiber networks. into fibers and handled jointly. These approaches are referred to as traffic grooming and λ (or wavelength) grooming, respectively. Figure 3 shows a traffic grooming switching architecture that first demultiplexes the wavelength channels of a fiber (e.g., by an arrayed waveguide grating, AWG). These channels can be switched/cross-connected by a wavelength crossconnect (λxc). The content of some (or in worst case all) of these wavelength channels is converted to electric signals and then by a TDM capable switch (e.g., an IP switch or router) can be remultiplexed. It is advantageous to remultiplex electronically only a small fraction of all wavelength channels for two reasons: first, to keep the load of the TDM switch low; second, to allow use of TDM switches that have less ports than the total number of incoming wavelength channels. Note that the conversion both from the optical to the electrical domain and vice versa must be supported at any wavelength, so wideband (or tunable) receivers and tunable transmitters are required. Figure 4 shows a λ grooming architecture. Here the fibers of cables can be cross-connected (FXC), while the wavelength channels carried by some of these fibers can be demultiplexed and switched/cross-connected by the λxc. The gain of λ grooming is that only a fraction of fibers have to be demultiplexed, and therefore only a fraction of wavelength channels have to be switched one by one. This leads to a decreased number of switch crosspoints. Similar to traffic grooming, optical traffic grooming can be performed by substituting electronic packet switches by optical packets or burst switches. Note that in the multilayer GMPLS/ ASTN architecture grooming can be performed between any two stacked layers. Is Grooming Needed at All? Although there are numerous aspects, the answer seems to be positive. First we will discuss traffic grooming, then wavelength grooming. References [6 9] give an excellent overview of grooming, including the most important definitions, and they show some applications as well. Traffic Grooming As an example let us consider a four-node network (Fig. 5) with three demands. If no grooming is allowed, for all the demands a single-hop wavelength path is required (i.e., at least three wavelengths per link) (Fig. 5a). Either all-optical or opto-electro-optical wavelength conversion may be allowed; however, it does not change anything. If we assume that node B is grooming capable, and the grooming is limited by either the number of ports of the upper layer switch or the bandwidth of wavelength channels, then two demands can be groomed as shown for demands A-B and A-D over link A-B in Fig. 5b. Here two wavelengths are sufficient. If there are no limits on grooming or limits are loose enough, full grooming is possible (i.e., one wavelength is sufficient). In this case, however, the load of the upper layer can be significant. 18 IEEE Network March/April 2003

4 C As mentioned in the preceding section the lack of wavelength channels calls for traffic grooming. However, then the network looses its transparency. The most significant drawback is that the whole network has to use a unique upper electrical layer that will perform traffic grooming (i.e., remultiplexing in time and space). If two traffic streams have different formats (e.g., Ethernet and PDH frames), they cannot be groomed. This unique upper layer must support all services with their various traffic and quality requirements, while if we use a transparent network without grooming, we could dedicate specific end-to-end wavelength channels to different services. Then the protocol established over a channel will take care of special requirements for that service. Furthermore, in case of grooming, the bit rates, encoding, and protocols should all be well defined and interoperable (if not unique) for the whole network. The advantage of traffic grooming is that it allows better resource utilization through arbitrary (sub-lambda) bandwidth granularity of the upper (electrical) layers. The physical length of paths will be shorter on average as well, but due to buffering where remultiplexing (grooming) is performed, the total latency will still be larger than in the transparent solution (with no grooming). Notice that grooming will not decrease the path length if there are so many wavelengths per fiber that all demands could be routed over the shortest path by a single-hop lightpath. When grooming is performed, a significant number of wavelengths per fiber can be saved if the wavelength channel capacities are significantly larger than the bandwidth requirements of end-to-end demands. This means, for example, that instead of 128 wavelengths, a 32-wavelength system is sufficient, or the network size and throughput can be increased without having to extend the system for more wavelengths. In this case, when the number of wavelengths is reduced, the average length of the signal path will increase, and therefore the latences will further increase! Traffic grooming with electronic TDM also regenerates the signal and allows wavelength conversion at no extra cost. From the network design and configuration point of view, although both are algorithmically very complex (NP-hard, e.g., the running time increases exponentially as the problem size grows), transparent networks are easier to handle, since one layer has to be optimized only once instead of at two or more layers, as in grooming. If the dynamic case is considered instead of the static one, routing becomes more complex. For optimal routing (and optimal grooming) the more complex peer model is preferred. As an example let us consider routing with shared protection in the 25-node COST 266 European reference network. Figure 6 presents the relative wavelength-link utilization, as the number of wavelengths per link grows (an increment of five has been used). Traffic grooming is supported in all nodes for all ports; however, the limiting factor is that the capacity of a single wavelength link can accommodate 10 units of traffic uniformly for the whole network. All traffic has bandwidth demand of one unit, which means that for the curve marked as 10, up to 10 demands can be groomed into a single wavelength link, and for the curve marked as 2, only up to two demands can be groomed into a single wavelength. If no grooming is allowed (i.e., the wavelength link capacity is equal to the bandwidth demands), the minimal number of wavelengths per link required to satisfy all demands was between 156 and 160 (not included in the figure). When allowing grooming of up to two demands per wavelength-link the required number of wavelengths was between 66 and 70 (marked by a red oval in Fig. 6 at the end of curve 2). It has A A (a) (b) B B A B D (c) Figure 5. Illustration of grooming: a) no grooming; b) limited grooming; c) full grooming. dropped by 57 percent, to less than half! When allowing grooming of up to three demands, the required number of wavelengths drops further to a value between 46 and 50 (i.e., by 70 percent from the case with no grooming). If up to 10 demands can be groomed, only 11 percent of the wavelengths are required as with no grooming. We can see that traffic grooming is very advantageous whenever the bandwidths of demands are lower than half the bandwidth of wavelength channels. If these values are close it is preferable to use λ grooming. λ Grooming The considerations are analogous for the lambda grooming case as well. With no grooming all the wavelength channels in all fibers are demultiplexed. Let W be the number of wavelengths per fiber, and F the number of fibers per cable. Then at a node with four neighbors (four cables connected), a switch with 4WF ports is needed, i.e., a 4WF 4WF switch. Then the number of switching elements required will be (4WF) 2. Such a switch is typically either not available or extremely expensive. However, the routing protocol through such a switch will be simpler and the wavelength blocking probability will typically be lower. As examples, 2D micro-electromechanical systems (MEMS), bubble switches, and switches based on liquid crystals or holograms should be mentioned [10]. In 3D MEMS only 8WF switching elements are needed. In these switches a switching element can switch either a single wavelength C C D D IEEE Network March/April

5 Wavelength-link utilization (%) Figure 6. Ultilization of wavelength links for different wavelength channel capacities as a function of the wavelengths per link. channel or all the wavelengths of a fiber. The cost of these switches grows as the number of switching elements they contain grows. In contrast to the above described case without λ grooming, λ grooming can switch all the traffic carried by a fiber with all the wavelengths it carries, saving W1 switching elements. Assuming that only some (e.g., a quarter) of the fibers should be demultiplexed into individual wavelengths, a (WF + 3F) (WF + 3F) switching matrix is sufficient. It will be roughly 1/16 of the size (for large W), and will have a price an order of magnitude lower than without grooming. Lambda grooming will result in a more complex routing protocol, particularly if shared protection is needed. Resource usage will be slightly worse; however, as shown above, a considerably smaller switch will suffice. In contrast to traffic grooming, λ grooming does not improve granularity and does not reduce the number of wavelength channels. It makes the paths typically longer but does not add remultiplexing delay. The routing becomes more complex. The advantage is a significant reduction of crosspoints in switches. Hierarchical Grooming Traffic grooming can be performed over λ grooming as well. However, then the routing and configuration are extremely complex. If protection is needed for failures, the problem scales further. As examples, MPLS with label stacking or the GMPLS/ASTN architectures can be mentioned. How to Build ovpns Number of wavelengths per link The idea of traffic grooming and wavelength (λ) grooming can be applied to set up optical virtual private networks (ovpns). Typically an ovpn will consist of as few wavelength paths as possible; numerous traffic streams belonging to that ovpn will share these wavelength paths, and a wavelength link will be used by one ovpn only [3]. In other words, the ovpn will be a virtual topology built of wavelength channels that will be used for routing the traffic of the VPN. There are two other possibilities as well. First, two or more VPNs may share a single wavelength channel. Then conventional VPNs are established over a wavelength routed WDM network. Second, a wavelength path is used exclusively to satisfy a single demand; this wavelength path will be dedicated to a single specific connection. Here, traffic grooming is avoided, but λ grooming can be deployed. λ-link capacity The Future: Optical Traffic Grooming? WDM (or DWDM) is analogous to circuit switching with a very large unit (e.g., 10 Gb/s) of capacity. Optical TDM (OTDM) is analogous to TDM, particularly to asynchronous TDM (i.e., packet switching). Therefore, the advantage of WDM is that a connection with dedicated resources is allocated guaranteeing quality of service (QoS). For switched networks this connection is torn down when no longer needed and reestablished if needed again. This is useful for connections of constant bit rates close to the optical channel capacity. However, the traffic characteristic is increasingly bursty, and therefore it might not be justified to build up a high bandwidth connection for sending only a few bytes in a longer interval. Therefore, variants of OTDM are promising where the labeled information is sent and forwarded to the destination hop by hop. There are three approaches to OTDM according to the timescale considered [11, 12]: Optical flow switching (OFS) Optical burst switching (OBS) Optical packet switching (OPS) Compared to the bit duration of data streams considered, the switching speed of optical switching elements is relatively low; thus, a significant guard time has to be left between certain information units, and between the header and payload of each information unit that may lead to large overheads in case of very short packets. Therefore, OPS is of limited practical interest nowadays, but bursts of packets or whole flows can be switched together, saving all the guard times between certain packets. OBS is of particular interest in the short term, while OPS will enable optical traffic grooming in the longer term. Optical burst switching is a trade-off between OPS and WDM. Conclusion In the short-term future dynamically reconfigurable WDM networks will serve metropolitan areas, while the transport networks are expected to be statically configured at first and automatically switched later. Future networks will likely consist of three layers under the IP layer: packet switching capable, wavelength switching capable (WSC), and fiber switching capable. In the near future the PSC layer will be only electrical, since the functionality of the electrical and optical layers differs significantly. Later it can be OPS-based, with electronic control, and the all-optical alternative may be viable as well, all supporting grooming. For multilayer networks with grooming, in the beginning only the overlay model will be applicable, while the peer model, due to its complexity, seems more promising later. Probably the transport part of the network will remain of the overlay type, while the metropolitan part will start to evolve to be based on the peer model. However, the network will have to be partitioned into smaller peer domains, since this model is orders of magnitude more complex. Aggregation of the topology and link state information is needed as well. The peer model is preferred for grooming, particularly for the dynamic case. From the MPLS/MPλS/GMPLS point of view, the labels can be stacked instead of swapped. Label stacking refers to forming hierarchical LSPs. This assumes hierarchical grooming as well. For example, MPLS units are groomed into wave- 20 IEEE Network March/April 2003

6 lengths, then wavelengths groomed into fibers, and fibers groomed into cables. The electronics shall be replaced by optics slowly, from bottom up as the technology allows. A slow evolution from circuit switching to packet switching is expected as well, particularly in the access and metropolitan parts of the networks, while the transport part will likely remain circuit-switched (i.e., based on WR-DWDM). The main advantage of λ grooming is that it saves a significant number of crosspoints in optical switches, while its drawback is increased routing complexity that affects protection as well. Traffic grooming has additional advantages over λ grooming: the signal paths become shorter, or fewer wavelength channels are needed. Traffic grooming is of great interest while the bandwidth of demands is less (or significantly less) than half the wavelength channel capacity. When the bandwidth of demands approaches the wavelength channel capacity, λ grooming is of more interest. References [1] A. Banerjee et al., Generalized Multiprotocol Label Switching: An Overview of Signaling Enhancements and Recovery Techniques, IEEE Commun. Mag., vol. 39, no. 7, July 2001, pp [2] B. Rajagopalan et al., IP over Optical Networks: Architectural Aspects, IEEE Commun. Mag., vol. 38, no. 9, Sept. 2000, pp [3] B. Rajagopalan et al., IP over Optical Networks: A Framework, draft-ietfipo-framework-02.txt, IETF Internet draft, work in progress, expired Dec. 10, [4] T. Armstrong and S. S. Gorshe, Eds., Generic Framing Procedure (GFP) and Data over SONET/SDH and OTN, Feature Topic, IEEE Commun. Mag., vol. 40, no. 5, May 2002, pp [5] E. Modiano and P. J. Lin, Traffic Grooming in WDM Networks, IEEE Commun. Mag., vol. 39, no. 7, July 2001, pp [6] R. Barr and R. A. Patterson, Grooming Telecommunications Network, Opt. Net. Mag., vol. 2, no. 3, May/June 2001, pp [7] S. Thiagarajan and A. K. Somani, Capacity Fairness of WDM Networks with Grooming Capabilities, Opt. Net. Mag., vol. 2, no. 3, May/June 2001, pp [8] L. A. Cox and J. R. Sanchez, Cost Savings from Optimized Packing and Grooming in Optical Circuits: Mesh versus Ring Comparisons, Opt. Net. Mag., vol. 2, no. 3, May/June 2001, pp [9] A. Lardies, R. Gupta, and R. A. Patterson, Traffic Grooming in a Multi-Layer Network, Opt. Net. Mag., vol. 2, no. 3, May/June 2001, pp [10] A. S. Morris, In Search of Transparent Networks, IEEE Spectrum, Oct. 2001, pp [11] S. Verma, H. Chaskar, and R. Ravikanth, Optical Burst Switching: A Viable Solution for Terabit IP Backbone, IEEE Network, Nov./Dec. 2000, pp [12] M. J. O Mahony et al., The Application of Optical Packet Switching in Future Communication Networks, IEEE Commun. Mag., vol. 39, no. 3, Mar. 2001, pp Biography TIBOR CINKLER [M 96] (cinkler@ttt-atm.ttt.bme.hu) has received M.Sc. (1994) and Ph.D. (1999) degrees from the Budapest University of Technology and Economics, Hungary, where he is currently an associate professor. His research interests focus on routing, design, configuration, dimensioning, resilience of IP, MPLS, ATM, SDH and particularly of WR-DWDM-based multilayer networks. He has been involved in a few related projects (ACTS, COST, ETIK) and is a member of IFIP TC6.10, ONDM, and DRCN Scientific Committees. He is the author of over 50 referred scientific publications and three patents. IEEE Network March/April