1 Ph.D. student funded by the Dutch Technology Foundation STW.

Transcription

1 OVER WAVELENGTH-ROUTED ACCESS NETWORKS Marcos Rogério Salvador 1 and Sonia Heemstra de Groot s: {salvador,heemstra}@csutwentenl Tel: / Fax: Communications Protocols (COM) Group / Telematics Systems and Services (TSS) Centre for Telematics and Information Technology (CTIT) / University of Twente PO Box AE Enschede, the Netherlands Abstract Wavelength-routed networks exploit the tremendous bandwidth of optical fibers and allow for optical routing of packets thus making end-to-end delay much lower In this paper we discuss some of the protocol issues that have to be dealt with such that Internet applications can smoothly run over wavelength-routed access networks Topics addressed include routing and lightpath assignment, latency reduction in multihop path and multicast support Keywords: WDM, Wavelength-routing,, Access Networks 1 INTRODUCTION The coming decade will see a large increase in the demands placed on communication network infrastructures The quantitative growth is already visible today and will become dramatic with the introduction of new users and heavy applications (eg, distributed virtual reality applications) The quality and diversity of the communication services to be delivered by future networks will also evolve dramatically For instance, tele-medicine will require much higher reliability, guaranteed maximum jitter and minimum throughput than those provided by today's networks Although actual electronics technology-based network solutions have so far been able to cope with the demands placed on network infrastructures, it is doubtful whether they will be able to get along with such an unpredictable demand at a reasonable complexity and cost (if able at all) For instance, actual electronics-based network solutions can operate at speeds up to 40Gb/s at the expense of high complexity and cost, while the potential of optical fibers is in the order of one thousand times more This has been the motivation behind the investigation of optical networking and Wavelength Division Multiplexing (WDM), in particular WDM is an optical technology that provides several distinct wavelength channels on the same fiber, enabling up to 25 Tbps bandwidth When WDM networks are all optical, they also provide much lower end-to-end delay and, hence, much higher throughput by avoiding queuing and electronic processing of packets at intermediate nodes of a path WDM networks have been extensively studied over the past few years, particularly in core networks With the expected convergence of not only data networks but also telephony, radio and television networks to the Internet, the use of WDM as its underlying transmission layer has attracted much interest lately In this scenario fits Flamingo, a multidisciplinary project that aims at supporting multimedia broadband services in access networks Flamingo's focus on access networks comes from the tendency of the 80/20 rule, which says that 80% of the traffic is local and 20% is external, to reverse leading access infrastructures to become the major bottleneck in the future s Internet infrastructure if they are not appropriately enhanced In this paper we discuss some of the protocol issues that have to be solved if the Internet Protocol () is to smoothly run over WDM access networks The rest of this paper is organized as follows: Quality of Service (QoS) support is discussed in section 2; section 3 briefly introduces all-optical networks (AONs) and WDM; section 4 gives an overview of Flamingo; the Routing and Lightpath Assignment (RLA) problem is addressed in section 5; latency reduction in multihop paths is approached in section 6; in section 7 multicast is discussed; finally, some concluding remarks are presented in section 8 2 QUALITY OF SERVICE SUPPORT Experiments within the Internet Engineering Task Force (IETF) showed that multimedia applications do not behave well in the Internet because of queuing delays that real-time packets experience, which result in packets dropping and eventually poor QoS In fact, QoS is not supported at all in the current Internet This conclusion led to the elaboration of Integrated Services (Intserv) [2] and more recently Differentiated Services (Diffserv) [1] by the IETF The basic idea behind these architectures is that by reserving resources and applying some traffic control, real-time 1 PhD student funded by the Dutch Technology Foundation STW

2 flows will experience lower delays, higher throughputs and, consequently, get the QoS they need Traffic control requires packets to be classified, which demands much more processing than required by the actual best-effort service With the expected (and unpredictable) increase in the network demands, particularly for multimedia traffic, classification will doubtless constitute a bottleneck, particularly in Intservenabled networks, which look at several header's fields Besides, low priority packets as best-effort ones are likely to experience even more delays than nowadays and best-effort traffic will likely continue to represent the majority of the total traffic in the Internet The answers to these problems are more bandwidth and lower queuing delays, and WDM AONs seem to be the long-term solution that makes it possible 3 ALL-OPTICAL NETWORKS AND WDM AONs [7, 8] form a class of networks where optical technology plays the major role in network functionality In such networks data is kept in the optical domain from the source to the destination node Electro-optical conversions and electronic processing occur only at the end nodes The most prominent and mature, WDM is one of the optical multiplexing technologies that make AONs possible WDM effectively uses the enormous optical fiber bandwidth, which is potentially of a few tens of Tbps, by exploiting the colors of the light The total optical fiber bandwidth is divided into a number of distinct, high capacity wavelengths channels (colors) that are multiplexed altogether in the same fiber as parallel links (see Figure 1) STM-64 λ mux 6 x 10 Gb/s fiber λ demux Figure 1 Wavelength division multiplexing STM-64 A connection is setup by establishing a lightpath, ie, an all-optical channel, from the source node to the destination node A lightpath that spans multiple fiber links to provide a circuit-switched connection between two end nodes is called single-hop path Electro-optical conversions, queuing and electronic processing occur only at the end nodes of a single-hop path, resulting in a much lower end-to-end delay In addition, a single-hop path provides transparency in the sense that diverse protocols and bit coding structures can be carried and intermediate nodes do not need to be aware of structure of data being transferred These features are very interesting for Virtual Private Networks (VPNs), for instance The establishment of a lightpath requires the end nodes to tune the same wavelength Different possibilities exist depending on whether transmitters, receivers or both are tunable Lightpaths on a fiber link must be on different wavelengths to prevent the interference of the optical signals Unless wavelength conversion is used that allows signals on a given wavelength to be changed onto another one, a single-hop path can only be established if the same wavelength is available on all links along the route This is called wavelength continuity constraint Node pairs that are not directly connected via single-hop paths must then use a sequence of lightpaths through intermediate nodes to communicate, configuring a multihop path In a multihop path, at each hop, packets coming in on a lightpath must be converted to electronic form, switched or routed electronically and then converted back to optical form and sent out on a different lightpath Figure 2 illustrates a network where node A communicates with node D through a single-hop path, bypassing all the nodes in between, and D communicates with A through a multihop path, where C is the intermediate hop classes: A B C D Figure 2 - Lightpaths WDM networks can be classified in two broadcast-and-select: all input signals are combined and broadcast to all outputs Each destination then selects the signals intended for it Examples of this class are star, bus and ring topology networks Wavelength routing: at each intermediate node light coming in on one port at a given wavelength gets routed out of one port at light speed, possibly at different wavelength An example of this class is a mesh topology network The main advantages of wavelength routing upon broadcast-and-select networks are: no power λ1 λ

3 splitting loss, wavelength reuse and, hence, scalability 2, and dynamic reconfiguration to cope with unbalanced traffic and failures 3 The main advantages of broadcastand-select upon wavelength routing networks are: simplicity, low cost, and natural support for broadcast and multicast The decision for one or another architecture must be based on the network needs and technological limitations Broadcast-and-select architecture has typically been deployed in Local Area Networks (LANs) and small Metropolitan Area Networks (MANs), while wavelength routing architecture has typically been deployed in medium-large MANs and Wide Area Networks (WANs) [7] Nevertheless, due to its features wavelength routing is likely to be deployed in smaller networks too as its cost goes down 4 FLAMINGO: OVER WDM ACCESS NETWORKS over WDM has been investigated in Flamingo 4 Flexible multiwavelength optical local access network supporting multimedia broadband services [6], a multidisciplinary project involving Dutch research institutes and led by CTIT University of Twente The scope of Flamingo includes the construction of optical devices, the design of the physical layer and the conception of novel communications protocol concepts and mechanisms to exploit the potential of WDM and support multimedia communications over Internet Router/ switch Host TCP/ over X over WDM Campus backbone X Departmental LAN Campus network Figure 3 Campus Network Focus is on a particular case of access networks: campus backbones interconnecting 2 In the sense of supporting addition of nodes without the need to introducing more wavelengths into the network 3 Broadcast-and-select networks can at some extent be dynamically reconfigured too 4 Work funded by STW; URL: departmental LANs and connected to the core network (see Figure 3) The architecture under consideration is wavelength routing due to traffic engineering and fault tolerance needs of network backbones Physical topology is arbitrary Network nodes consist of 2x2 non-blocking Cross-connect (OXC) devices capable to switch up to eight wavelengths (four per fiber), each one carrying up to 1Gbps Attached to each OXC is an Add-Drop (OAD) multiplexer that allows departmental nodes to inject data (ie, to add lightpaths) into the OXC or get data (ie, to drop lightpaths) from it Figure 4 depicts the OXC Wavelength Demux λ1 λ2 λ3 λ4 Add-Drop Multiplexer Wavelength Mux Figure 4 2x2 OXC for 4 wavelengths Lightpaths can only be dropped from (or added to) an OXC if they are carried on different wavelengths For instance, lightpaths carried on λ1 and λ3 can be dropped from one port while lightpaths carried on λ2 and λ4 can be dropped from the other port These characteristics define the node functionality A lightpath is unidirectional and it is setup by configuring the OXCs and OADs along the path Setup of lightpaths takes place either at initialization time or when the network is reconfigured Lightpath setup at connection basis is expected to become impractical due to the long time needed to coordinate this action The set of all lightpaths between end nodes on top of a given physical topology defines the virtual topology of the network Given the physical topology of the network, the nodes functionality, the number of wavelengths and the traffic matrix representing the long-term average flows between end nodes, the goal is to minimize both the network congestion and the average packet delay, which is due to the queuing

4 delays at intermediate nodes 5 Further information can be found in [7] Wavelength conversion has been considered as a mean to minimize the number of hops in the network Mainly due to its still prohibitive cost investigations have focused on the placement of a limited number of wavelength converter units on a limited number of nodes The benefits of limited wavelength conversion are discussed in [8] In principle wavelength conversion can be performed at the boundary between the campus network and the core network as well as at the boundaries between the campus backbone and departmental nodes However, there still are security and traffic engineering issues to be considered WDM control and management information (eg connection setup) will be transmitted over the network through out-of-band signaling for reliability and security reasons Preference is given for using a transmission medium (eg, coax cable) other than a specific wavelength because network nodes might face problems trying to tune the signaling wavelength and without signaling the network does not work Because TCP/ signaling is mixed with user data it will be inband as usual Concerning QoS support, Intserv will be deployed in the backbone Mapping between Intserv and Diffserv will be supported to allow end-to-end QoS when end users are located outside campus 5 ROUTING AND LIGHTPATH ASSIGNMENT Since lightpaths are setup when the network is initialized, the routing problem is that of finding a lightpath that leads to the (next or destination) hop calculated by This process is referred to as Routing and Lightpath Assignment (RLA) in this paper The RLA problem comes from the fact that sees links by nature while WDM sees lightpaths One link may carry several lightpaths, whereas one lightpath may span many links That means that lightpaths may lead to different nodes other than those calculated by For instance, the selection of the lightpath carried on λ1 at A, in Figure 2, will lead directly to D independent on the next hop calculated by Another problem is that routing calculation takes into account link metric values In WDM, however, metric values are inherent to lightpaths, not 5 The propagation delays are neglected in small high-speed networks as campus ones links Each lightpath has its own metric values In order to meet user requirements and deal with congestion it is very important that metric values are accurate, which is definitely not the case if they are associated to links Therefore, particularly if QoS-based routing [4] is to be supported, accurate metric values must be obtained somehow during routing calculation In this sense lightpaths must somehow be taken into account during the routing calculation process The only restriction on how this can be done is that must continue to be unaware of the particularities of the lower layers, which is one of the main reasons of 's success The easiest way to overcome these problems is to make aware of lightpaths transparently by storing lightpaths rather than links in the routes tables The problem with this approach is the number of addresses that might be required depending on the number of wavelengths in the network On the other hand, this might not be an issue with the deployment of Classless Inter-Domain Routing (CIDR) [9] and v6 [5] 128-bit address Another approach is to employ some sort of routing at the link layer This routing could be as simple as looking up an entry in a lightpaths table (when only single-hop paths are in place) or as complex as routing performed in (if multihop paths exist) The RLA process would then be decomposed in the following steps: 1) finding the shortest link-based path; 2) selecting all lightpaths that match the calculated path (next hop); and then 3) matching the user's requirements against the metric values of these selected lightpaths If no suitable lightpath is found then the process restarts using an alternate route or eventually the connection request is refused Policies based on user requirements and network parameters must also be defined in order to make a good use of lightpaths For instance, real-time traffic should be assigned single-hop paths while besteffort traffic should be assigned multihop paths Note that finding a feasible path with independent constraints, as the case with QoS-based RLA, is NP-complete Hop-by-hop routing as typically used in the Internet makes QoS-based RLA even more difficult to solve Discussions on the benefits of using strict source-based routing, an optional strategy that calculates the entire route from the source to the destination at the source, have gained more audience lately (eg, [3]) with the introduction of QoS in the

5 Internet It is argued that strict source-based routing is simple, easy to implement and upgrade Further it makes NP-complete problems much easier to solve than in distributed ways [4] Strict source-based routing has several problems (eg, computation overhead at the source, the global state maintaining at every node) that prevents it of being used in large networks, though Since Flamingo is meant for access networks, which are typically limited to a few hundreds of nodes, this strategy may prove appropriate 6 LATENCY REDUCTION: THE MULTIHOP CASE Because lightpaths are setup when the network is initialized, packets forwarding in single-hop paths is very simple and fast since it involves the end nodes only In multihop paths the typical store-and-forward strategy takes place, making communication latency higher as the number of hops increases Unlike hop-by-hop routing in the Internet, in which incoming packets are routed at every hop 6 based on a longest match algorithm applied to a (long) address prefix, in our network packets are switched at layer 2 based on short-fixed labels, as usual in circuit switching RLA is performed only for the first packet of a stream (and at the ingress node only if strict sourcebased routing is used) The result is cached in order to reduce latency and prevent packets of a given stream of being routed to different paths All incoming packets are matched against this cache Failing in finding an appropriate entry results in the performing of RLA A single and unique label is assigned to a packet stream and distributed to all the nodes along the path At the ingress node, the label is attached to all the packets of that stream At each hop, incoming packets have their labels matched against the Forwarding Information Table (FIT) s entries A successful matching results in the packet being queued for transmission on the appropriate output interface and wavelength Otherwise RLA takes place FITs are maintained at every node Each entry consists basically of a label, an output interface and an output wavelength One entry can also include the stream s profile or QoS in order to facilitate the classification of incoming packets to the appropriate class of service 7 Figure 5 depicts label switching Dept LAN A Ingress /Router Router Label Out Out ITFC W 1 2 Label switching Intermediate /Router switching Router Label Out Out ITFC W X 1 4 Egress /Router Figure 5 Label switching Router Label Out Out ITFC W Y 1 Dept LAN B The characteristics of our network allow some further improvements As mentioned earlier, lightpaths setup takes place only when the network is initialized or reconfigured Signaling is therefore required only to distribute labels along multihop paths and to make reservations (for real-time traffic) along these paths Thus, if labels are assigned to lightpaths (and distributed throughout the network) when the network is initialized (or reconfigured) then label setup is not necessary for best-effort traffic This is called topologybased label assignment strategy It is expected that the combination of topologybased and request-based label assignment strategies will provide aggregation and no label setup latency for besteffort traffic, and individualized treatment of real-time flows Clearly, some integration between routing and layer 2 switching is needed MultiProtocol Label ing (MPLS) [3] does this and a specialized version of it is a candidate solution 7 MULTICAST In the current Internet multicast transmissions are achieved by sending one copy of each packet to every receiver, consuming unnecessary bandwidth and processing 8 True multicast, or simply multicast, is the term used to describe the ability of sending only one copy of the information to all the receivers In order to get as low latency as for unicast transmissions and fully exploit the benefits of multicast, the ability to provide multicast optically is desirable However, wavelength routing makes multicast somewhat difficult Specifically, providing multicast sessions with single-hop paths would require OXCs to be able to replicate or split light signals somehow to the leaf nodes 6 This is not the case when strict-based source routing is used 7 Classification is claimed the bottleneck of QoS-enabled networks 8 "During testing, Toys R Us found that it took 6 hours to transfer a 1Mb file to 250 clients using the current Internet, while the same transfer using multicast file transfer took 4 minutes" [10]

6 Due to the wavelength continuity constraint though, setup of a multicast single-hop path is very difficult to be achieved Wavelength conversion may be deployed to alleviate this constraint but the cost of wavelength converter units prohibits full wavelength conversion throughout the network Another challenge to optical multicast is the support of leaf-initiated join and leave, which requires OXCs to be able to be dynamically reconfigured without considerably impacting the actual transmissions In addition, lightpaths are unidirectional, which makes applications like teleconferencing even more difficult to be supported OXCs added with multicast capability are being studied in Flamingo Further studies on multicast issues and physical layer aspects will tell whether or not optical multicast is possible One solution that has been considered and seems to be reasonable is the reservation of a number of multihop paths, linking all the nodes, to deal specifically with multicast traffic In such a configuration, at each hop, incoming packets would be replicated and switched to the leaf nodes through label switching at layer 2, instead of optically This can easily be done by associating more than one output interface/wavelength pair per label at the FITs 8 CONCLUDING REMARKS The Internet is getting more complex and processing demanding as QoS features are introduced With plenty of bandwidth enabled by WDM, the main focus of the project is therefore on latency minimization Due to the yet excessive latency of optical packet switching, the network will be circuit-switched and lightpaths will be setup only when the network is initialized or reconfigured The goal is to minimize the number of hops between any pair of end nodes in the network as well as congestion such that network reconfiguration is kept minimal Routing is calculated for the first packet of a stream only Routing is likely to be performed only at the ingress node of the network as well, as a mean to facilitate the process of finding lightpaths that provide users with the QoS they need This will depend on some performance aspects that are unclear at this point Label-based switching is performed to speedup packet forwarding in multihop paths In addition to the usual request-based label assignment strategy, required by packet streams that need special treatment, labels are also assigned to lightpaths at initialization time, resulting in no label setup latency for best-effort traffic Single-hop multicast paths are also a concern as a mean to reduce latency The challenge is to support leaf-initiated join and leave These ideas are under development in Flamingo A more in depth analysis will determine how appropriate these ideas are Further aspects to be addressed include network reconfiguration triggering, mobility issues and their impact on QoS REFERENCES [1] Blake S et al An Architecture for Differentiated Services IETF RFC 2475, December, 1998 [2] Braden R et al Integrated Services in the Internet Architecture: an Overview IETF RFC 1633, June 1994 [3] Rosen E C et al Multiprotocol Label ing Architecture IETF Internet Draft <draftietf-mpls-arch-05txt>, April, 1999 [4] Chen S and Nahrstedt K An overview of Quality of Service Routing for Next-Generation High-Speed Networks: Problems and Solutions IEEE Network, pp 64-79, November/December, 1998 [5] Deering S and Hinden R Internet Protocol, Version 6 (v6) Specification IETF Internet Draft <Draft-ietf-ipngwg-ipv6-spec-v2-02txt>, August 1998 [6] Niemegeers I Flamingo Flexible Multiwavelength Local Access Network Supporting Multimedia Broadband Services Project Description, 1998 [7] Ramaswami R and Sivarajan K N Networks: A Practical Perspective Morgan Kaufmann Publishers, 1998 [8] Mukherjee B Communication Networks McGraw-Hill, 1997 [9] Rekhter Y and Li T An Architecture for Address Allocation with CIDR IETF RFC 1518, 1993 [10] Stardust Technologies Multicast Backgrounder White paper, 1997 (