an alternative solution to the electro-optic and service bottleneck problems in integrated...

Computer Networks and ISDN Systems 25 (1993) 1089-1105 1089 North-Holland

An alternative solution to the electro-optic and service bottleneck problems in integrated multi-Gbit/s LANs: the SUPERLAN architecture

A d r i a n P o p e s c u a a n d R a g h u v a n s h P r a s a d Singh b

a Department of Telecommunication and Computer Systems, Royal Institute of S 100 44 Stockholm, Sweden b Bell Communications Research, 331 Newman Springs Road, Red Bank, NJ 07701, USA

Technology, P.O. Box 70043,

Abstract

A. Popescu and R.P. Singh, An alternative solution to the electro-optic and service bottleneck problems in integrated multi-Gbit/s LANs: the SUPERLAN architecture, Computer Networks and ISDN Systems 25 (1993) 1089-1105.

This paper examines the two main bottleneck problems in the design of integrated multi-Gbit/s local area networks (LAN), one from the network perspective--the electro-optic bottleneck problem--and other from the user application perspective - - t he service bottleneck problem. The major design challenge is to develop viable network architectures which effectively deal with these problems, and allow sharing of the vast amount of bandwidth available on optical fiber among distributed users offering traffic with different and conflicting performance requirements whose peak rates are constrained by the limited speed of electronic interfaces. A novel architectural concept is proposed here to address these issues. It makes use of the wavelength division multiaccess (WDMA) architecture where time-synchronous channels placed in different wavelengths are dedicated to different applications, e.g., continuous or variable bit-rate video services, high speed computer data transfer, 64 kbi t /s voice application, application-dependent control mechanisms for media access, error supervision/recovery and flow control, network clock distribution, etc. In the architecture proposed in this paper, the total user traffic on the fiber is separated into two classes--isochronous and nonisochronous--each carried on separate wavelengths at multi-Gbit/s rates. The principal features of this architecture, called the SUPERLAN architecture, are: extreme flexibility in providing services to diverse network and traffic conditions, low processing rates that are commensurable with the node electronic speeds, simplicity of protocols, a minimal number of readily available Gbi t / s components at the physical layer (PHY) and network performance restricted only by optics and not by electronics or protocols.

Keywords: Wavelength Division Multiplexing (WDM); synchronization; multiservice; network architecture; multi-Gbit/s LANs; traffic control

1. Introduct ion

With the availability of the fiber optic technology, new network architectures are emerging to effectively utilize the abundant transmission capacity. The choice of WDM technique has fur-

Correspondence to: A. Popescu, Department of Telecommuni- cation and Computer Systems, Royal Institute of Technology, P.O. Box 70043, S 100 44 Stockholm, Sweden.

ther resulted in manifold increase in transmission capacity. However, the performance increase of the supporting nodal electronics needed for switching, buffering and control purposes has not matched this trend, so novel architectures must be devised in order to make use of the huge transmission capacity. This picture is further complicated by the fact that the future networks must provide support to a multitude of broadband traffic classes with conflicting service re-

0169-7552/93/$06.00 © 1993 - Elsevier Science Publishers B.V. All rights reserved

1090

quirements. Performance and cost-effective LAN architectures need to employ efficient methods to share the system resources among the network stations in a manner that circumvents the mismatch between the transmission and processing speeds, and at the same time, provides adequate quality of service to all user applications. Gener- ally, existing integrated LANs employ sophisticated mechanisms to control and coordinate access to the communication media, so that the service performance requirements of each user application are met. But because the needed processing speeds are limited, the network transmission resources are not efficiently utilized. This is the kind of approach in which the performance needs of the user applications are adapted to a given network environment. This means that the network architecture is designed first and then complex functions are incorporated in the media access protocol to satisfy the performance needs of the applications. Our approach to high speed LAN design is just the reverse, i.e., we adapt the underlying network transport mechanisms to the needs of the specific applications. Based on this consideration, we propose a new LAN model, in which we assign one wavelength to each application or traffic class, partitioned on the basis of

some characteristics, such as loss or delay sensi- tivities, holding times, bit rates, etc. Moreover, this is a model in which each traffic class has a separate dedicated access control mechanism, which is used specifically to arbitrate access for resources on the separate wavelength carrying the traffic from the same traffic class. The transport mechanism for each traffic class as well as the associated control channel may be different from each other, and depends on the performance requirements of the signals being trans- ported. Such an approach to network design offers important advantages, such as extreme flexibility in providing services to different network and traffic conditions, low processing rates that are commensurable with the node electronic speeds, simplicity of protocols, a minimal number of readily available high speed components and network performance restricted only by optics and not by electronics or protocols. In this con- text, we present here the simplest model. Based on delay sensitivity, we divide the total user traffic on the network into two classes--isochronous (voice and video) and nonisochronous ( d a t a ) - that are WDM integrated in a multi-Gbit/s LAN at the low speed network interface.

In this paper we concentrate first on studying

Adrian Popescu received the Ph.D. degree in electrical engineering from the Polytechnic Institute of Bucharest, Romania in 1985. From 1972 to 1986 he was with the Technological Institute for Telecommuni- cations, Bucharest, where he worked primarily in the field of digital transmission systems. He joined the Royal Institute of Technology, Stockholm, Sweden in 1986, where he is currently working on architectures and performance evaluation of very high speed LANs. His research interests include design, modeling and performance analysis of communication protocols, as well as parallel discrete event simulation.

Raghuvansh Prasad Siagh received the Ph.D. degree in electrical engineering from Yale University, New Haven, CT in 1985. Since 1984 he has been with the Bell Communications Research, Red Bank, NJ, where he has worked on various problems in integrated broadband networks. He spent the academic year 1988-1989 as a guest researcher at the Royal Institute of Technology, Stockholm, Sweden. There he started the work on the design and performance analysis of very high speed LANs. His current research interests include systems engineering, traffic modeling, congestion control and performance evaluation in high speed networks.

Dr. Singh is a senior member of the IEEE. He is on the editorial board of the J. of High Speed Networks, and was a program committee member for the 3rd IFIP Workshop on Protocols for High Speed Networks held in Stockholm, Sweden.

the two basic issues to be resolved in the design of integrated multi-Gbit/s LANs--the electro- optic bottleneck and service bottleneck--and show how their resolution motivates the new architectural solution proposed here. We illustrate how diverse traffic requirements for new (broadband) services demand not only appropriate tun- ing of the parameters of media access mechanisms but also require optimized (virtual) topologies and system architectures as well as transport protocols. Based on this analysis, we report a novel LAN architecture, called SUPERLAN, ex- emplified by a ring topology. Particular emphasis is placed on the fundamental issue of time synchronization between channels placed in different wavelengths of the same physical facility. The synchronization performance is shown to be independent of the number of stations around the ring. The proposed model is intended mainly as a foundation for further functional and analytical studies in the area of high and very high speed networking.

This paper is organized as follows. Section 2 presents some of the critical issues in the design of a multi-Gbit/s network designated for broadband integrated services. In Section 3, we discuss the issue of electronic bottleneck and motivate our proposed solutions. In Section 4, we discuss the service bottleneck, and show how the SU- PERLAN architecture obviates the need for complex protocols for resource allocation, error recovery, etc. Section 5 is concerned with a descrip- tion of the SUPERLAN architecture, emphasiz- ing its distinctive features and advantages. In Section 6 we present the synchronization solution and report its performance. We conclude the paper in Section 7, where we outline our ongoing work.

2. Background

A multi-Gbit/s LAN must be designed to support a wide range of applications generating isochronous and nonisochronous traffic with arbi- trary bit rates, both narrowband and broadband. The increasing need for high bandwidth networking, under increasingly strict performance constraints, to provide integrated communication services ranging from the ubiquitous 64 kbit/s voice up to multi-Gbit/s video and distributed,

1091

parallel computing environments has posed fundamental challenges to LAN design. With the available fiber optic technology, the potentially usable bandwidth is enormous. This means that the performance bottleneck of the network is moved from the transmission channel to network nodes. This bottleneck has two components in terms of the increased ratios of nodal processing time to cell/packet transmission time and of propagation delay to cell/packet transmission time. The main question is what techniques should be employed for the network protocols (media access and transport, i.e., error recovery, flow control, etc.) and how they should be imple- mented in order to divide the available raw bandwidth to support services with widely varying performance demands. Typically, sophisticated access procedures and data buffering must be exe- cuted very quickly to match the service requirements and transmission speeds on optical fiber. But the difficulty in designing and implementing electronic processors for handling media access and transport protocols operating beyond Mbit/s rates and the lack of practical optical components with memory functions have resulted in a clear disparity between the rate at which data may be transmitted and the rate at which data may be processed and switched. Such implementation-re- lated problems, created by the fundamental speed mismatch between the electronic and optical components of a very high speed LAN, have resulted in electro-optical interfaces being the real performance bottleneck. So the design of new architectures and new access protocols tar- geted to remove the electro-optic bottleneck from network nodes represents the first issue to be resolved in designing multi-Gbit/s networks.

The second bottleneck that prevents the high bandwidth of the optical fiber media from being available to hosts/users is the so-called service bottleneck between the media access layer and the higher layer (Fig. 1). The incoming traffic is highly heterogeneous in its characteristics and performance requirements, with different and contradictory bandwidth demands, holding times and call arrival rates. Moreover, each traffic dif- fers in its admissible access and transit delays, access throughput, etc. This traffic can be divided into two general classes, isochronous and nonisochronous, with different performance demands [13]. Isochronous traffic generally includes Con-

1092

tinuous Bit Rate (CBR) traffic and Variable Bit Rate (VBR) traffic of continuous type, e.g., VBR video traffic. Nonisochronous traffic sources generate intermittent traffic and usually alternate between active and inactive periods. The lengths of these two periods may randomly vary, and during active periods bit rates may be constant or may vary randomly depending on the application.

As the term indicates, for the first class information is generated either in a steady time-synchronous mode in the case of synchronous traffic, or in a nearly synchronous mode in the case of isochronous traffic. This traffic, characterized by long holding times and modest setup times, may accept high access delay but puts rather stringent requirements to network in terms of transit delay (up to tens of ms), delay variance, i.e., delay jitter, and bandwidth demands (for video applications). It also has flexible loss sensitivity. In supporting synchronous or isochronous services, the network must be able to provide real or virtual connec- tions with guaranteed performance (quality) of service as negotiated at the call setup. The negotiated performance can be of various types, as for instance, an upper bound on call blocking probability, a lower bound on throughput, an upper bound on transit delay, or an upper bound on loss rate.

On the other hand, nonisochronous sources generate bursty information of random lengths at random times and, usually, with a low activity factor, i.e., the source is active only a fraction of the time. Most of the computer data applications

fall into this class. The variety of computer application models translate into a variety of performance needs expected from the underlying transport network. While the actual computing paradigms like timesharing, transaction and mainframe [23] are easily supported by a Connec- tion-Oriented (CO) or Connectionless (CL) transport network, mainly because of the reasonable services needed (low throughput and low latency to large throughput and moderate latency requirements), the picture is completely changed in the case of the emerging distributed, parallel processing environments. Thus, for the client- server computing paradigm, both the very low response time and large throughput are crucial. Here, remote software operations like Remote Procedure Call (RPC), Inter-Object Communica- tion (IOC) and Demand Paging (DP) invoke edge-to-edge network delays of tens, respectively hundreds of microseconds and data bursts with lengths in the range up to hundreds of kbits [23]. Very hard requirements! If, for instance, we consider a data burst of 100 kbits with a network transmission speed of 10 Gbit/s, that means a data transmission time of 10 /xs, i.e., 20 ~s for both sides of a link. Also, the propagation time on optical fiber (5 /~s/km) is comparable with these figures. The first conclusion is that such figures can be provided only by a connectionless network transport mechanism. The second one is that the underlying transport network must work in extreme conditions, such as the highest possible transmission speed, and must have the choice

Substations I

I

I

Media Media interface (station) 1 interface (station) n

"1 " r . . . . . . . . . . . . . .

~ t PHY I :°'r

Service bottlenec~

1 1 I I I I

i Multi-Gbit/s i t I

Optical fiber I

I

Electro-optic bottleneck

I

I

PHYI MAC IHigherl: ' .~ layer f , / ' l ' ° I ,

• ! Higher : | l aye r I~

: I / ~ o n ~ o ,

I . . . . . . . . . . . . . J

Electro-optic I bottleneck v Service bottleneck

Fig. 1. Multi-Gbit/s LAN configuration.

Substations I f

I

1093

of optimization for the physical/logical topology so as to reduce the propagation delay to the minimum possible [5]. The protocols (media access and transport) should also be simple and with adjustable parameters for minimum processing delay. The error recovery schemes should work on an edge-to-edge or network basis [4]. The flow control protocols should work on a preventive basis and be entirely dedicated to applications [25].

In order to cope with a diverse mix of broadband services up to hundreds of Mbit/s rates and with conflicting performance requirements, as mentioned above, simply making use of the existing architectures and media interfaces do not provide any benefit. In fact, the integration of broadband services on the existing architectures adds more complexity to the already complex media access controllers (MACs). The added functionality further exacerbates the electro-optic and service bottleneck problems, resulting in even more inefficient resource utilization and unsatis- factory performance. Thus, there is a definite need to closely examine these two bottleneck issues and to explore new ways of dealing with them. These include proper architecture, as well as proper design of switching and transmission schemes, protocol mechanisms and network management policies.

3. Electronic bottleneck

Although the optical fiber technology may support traffic with capacities up to Tbit/s, the electronic components at the network nodes, which typically operate at rates up to about 1 Gbit/s, drastically limit the total throughput. For instance, the new systems being designed and de- veloped to take advantage of the lightwave technology, such as the Fiber Distributed Data Inter- face (FDDI), the Distributed Queue Dual Bus (DQDB) and the Asynchronous Transfer Mode (ATM) in combination with the Synchronous Op- tical Network (SONET) can provide network throughputs only up to hundreds of Mbit/s. Such networks are inherently limited by the use of electronic components at stations and are of architectures that do not take advantage of the very high bandwidth of optical fiber.

There are many approaches to open up this

electro-optic interface bottleneck and, among them, two seem most promising. The first one, the so-called "multihop" architecture, makes use of a new network architecture to achieve high capacity with existing devices. This architecture, used by several recently proposed networks, such as Manhattan Street Network (MSN) [24], Shuf- fieNet [1], Store-and-Forward With Integrated Frequency-Time (SWIFT) [8], Lightnet [9] and Wavelength-Division Optical Network (WON) [5], has a distributed topology with a distributed temporal switching function, in which each station has access to a small number of fixed-wavelength transmitters and receivers. These wavelengths are assigned to stations in a manner that allows any pair of users to communicate with each other either directly (without hopping) or through one or more intermediate stations (with wavelength hopping). This approach presents, nevertheless, the inherent drawback of a so-called "deloading factor" in the case of a realistic (nonuniform) load pattern and fixed routing [11]. The load imbalance, due to either traffic intensity variability or traffic pattern variability, has an effect to reduce throughput per station by a factor of 0.3 to 0.5 relative to the balanced-load situation, and this means that the load imbalance might have a serious negative impact as the network size in- creases. Alternatively, adaptive routing algorithms (like datagram or virtual circuit) can im- prove the network performance [18,24]. However, the fact that the stations have a store-and-forward configuration means that such networks present rather high latencies and jitter, which should be avoided for isochronous traffic and applications of client-server type.

The second solution for this bottleneck problem is to use a pure WDM technique for transmission, combined with some form of WDM, TDM (time division multiplexing) or SDM (space division multiplexing) for switching. Examples of such alternative architectures are Lambdanet, the Hypass and Bhypass Switches [15], the Photonic Knockout Switch [12], the Star-Track Switch [22], and the High Capacity MAN [2]. Although such networks are capable of supporting multi-Gbit/s traffic rates, they present fundamental difficulties when using distributed control. Typically, these networks have a star topology with a central passive node to and from which all communication takes place, and each user transmits its infor-

1094

mation on a unique wavelength. When setting up a connection, the transmitter and the receiver must first be tuned to the same wavelength. But in the absence of a central active controller, how do they find each other? A pretransmission coor- dination (i.e., a call/session setup procedure) between two users wishing to communicate is therefore required. Possible solutions to this are [2,12,17,22]:

(i) scanning of the wavelength bands by users listening for callers;

(ii) using a WDM switching function in a central active node; and

(iii) connection setup through a separate, common channel.

Despite differences in operation and performance, all these procedures are time-consuming and cause high access delays [7,17]. Therefore, they are not consistent with bursty traffic needs in terms of low access delay requirements. More- over, wavelength switching is a time-consuming process as well. The multiwavelength switching can be performed in two ways, either by tunable receivers/optical filters (multiwavelength broad- cast and select) or by tunable lasers (active wavelength routing). Both of them require wavelength- agile components which are not yet available in a rapidly-tunable form, and nor are likely to be in the near future [6,19,21,35].

Based on the above considerations, we ad- vance a new concept for opening up this bottleneck. We propose a WDM based network architecture, in which the wavelengths are no longer dedicated to different users, but to different applications, such as isochronous traffic, nonisochronous traffic, dedicated control mechanisms for media access, error supervision/recovery, flow control, network clock distribution, etc. The users have permanent access to the wavelengths of interest. Thus, at least two wavelengths can be dedicated to each application or traffic c lass--a very high capacity channel to carry user information and a low capacity channel to carry control information for media access. But since the needed bandwidth for control purposes would generally be quite small, several control channels with close characteristics can also be combined into a TDM format on one or more (but less than the number of traffic classes or applications) channels on a fewer number of wavelengths. The same considerations are true

for the other control mechanisms (error and flow control, etc.) as well. These concepts lead to a network architecture which can provide very high throughputs with simple and low delay control mechanisms. In addition, there is no need for wavelength agility as in the previous cases. Also, the a priori partitioning of the network resources and the separation of the protocols for different traffic classes mean that the impact of one traffic class on the performance of the other classes is greatly diminished or even eliminated when compared to single-class based network architectures mentioned above. Furthermore, because of a large amount of bandwidth available on the network, there is no need to employ complex mechanisms for resource allocation to different traffic classes, as the percentage improvement in network utilization will be small at the expense of protocol complexity. In fact, each control channel can fol- low its own simple control mechanism, depending on the performance needs of the particular traffic class it supports. In contrast to this, the WDM networks reported in the literature convey all traffic on the same wavelength. Because of this, the problem of contention resolution for the common network resources between different traffic streams, with contradictory performance requirements, becomes complex. Moreover, the contention resolution mechanisms must operate on a time scale consistent with the very high speed of the WDM network.

4. Service bottleneck

As mentioned in Section 2, the service bottleneck problem arises because different service applications or traffic classes have different, and often conflicting, performance requirements. Three main points to be discussed below reflect some of the fundamental technological premises that form the bases for the new approach to very high speed integrated LAN design. On these bases, we examine three main components of a LAN, namely, transmission techniques, switching techniques and access mechanisms, and demon- strate how the SUPERLAN architecture provides a solution to the service bottleneck problem.

An integrated network, in which all traffic from diverse applications is carried in a single

network, offers several advantages, which include flexibility, bit-rate transparency, universal access with bandwidth on demand, and a single network solution. In spite of these conceptual advantages, such an approach poses tremendous technical difficulties in the dimensioning of network elements and allocation of resources to competing user applications. Further, a myriad of interacting preventive and reactive mechanisms are needed that concern with:

(i) guaranteeing contracted quality of service (QOS) for each traffic class while preventing one user from degrading the performance of other users;

(ii) efficient utilization of the available resources; and

(iii) providing fair access to the network resources for all users.

The resulting equipments (hardware and real- time software) will turn out to be very complex, if all these mechanisms were to be ever realized. Moreover, electronic processing being the performance bottleneck, the goal of the network design should be to conserve processing by reducing the functional complexity of network elements to the necessary minimum, and by trading off communication capacity and buffers. Thus, there is a need to reconsider the approach to very high speed LAN design, which integrates diverse service applications on to a single network.

Secondly, the use of optical fiber as a transmission medium has provided much lower loss and error rates. Now error rates of 10-12 and less are easily obtainable, so most of the losses are due to buffer overflows in congested processors. This effect, together with the significant increase in propagation delay and processing time with respect to the cel l /packet transmission time, makes it necessary to reexamine the actual existing error and flow control schemes. The high available bandwidth resource and the very low error rate mean that the transmitted data blocks (windows) can be increased so as to entirely match the application needs without packetizing/depacke- tizing [16]. This motivates error control schemes that operate on an edge-to-edge or network basis, so that the delay performance is met [4]. Also, the mechanisms for flow control to prevent buffer overflow may be dedicated to individual applications and they should work on a preventive basis. Therefore, a data block acknowledgment scheme

1095

operable on an edge-to-edge basis with delay as the main performance criterion is considered here.

Finally, the third and the last point that must be taken into consideration is concerned with the increased nodal processing power, storage capacity and control data transfer rate capability required for very high speed networks. Taken together, these enhanced requirements mean an increased need for network resources (bandwidth and buffer) that must be provided in each node for control purposes. To this end, it must be pointed out that the increased nodal resource needs for control traffic should not lead to performance bottlenecks for other user traffic classes. Therefore, minimization of the impact of the control mechanisms on the network performance should be an important consideration in network design.

There are two types of contentions in an integrated network: (i) contention among classes of traffic for com-

mon resources; and (ii) contention among users/hosts inside the

same class of traffic for common resources. In a LAN, the medium itself is a transport

resource as well as a switch. There are two ways of providing transmission resource to different classes of traffic: • all traffic classes compete for a common resource, in time or in frequency/wavelength domain, based on some priority policy. We ignore other alternatives, like common resources in code domain, because of their poorness in terms of spectral efficiency, which results in low throughput [33]. The main drawback of this contention policy stems from the need to employ complex and high speed mechanisms for resource allocation to different traffic streams, with different performance requirements. This heavily compli- cates the access protocol. The case of common resources in time domain, best illustrated by the evolving ATM networks, means that the stations have their transmissions scheduled to take place at different segments of time. The most serious drawback of such an approach is the high impact of one traffic class on the performances of the other traffic classes [36]. The case of common resources in wavelength domain, the so-called wavelength dedicated to user approach, is exempli- fied by networks like ShuffleNet [1], SWIFT [8],

1096

WON [5], Lambdanet [15], etc. Here the optical fiber bandwidth is split up into multiple disjoint frequency bands, and the stations are allocated one or more such channels for transmission purposes. They transmit all their traffic in a TDM form on these wavelength channels, so we still have contention among various traffic classes for common resources in time domain, with the dis- advantage mentioned above. • separate resources for different traffic classes. Given the large amount of resource available in a multi-Gbit/s network, the advantages of the former policy, expressed in terms of service flexibility and network efficiency, do not motivate the increased complexity of the protocol processing, congestion control and packet switching. There is no need to make use of sophisticated mechanisms for media access, contention resolution, switching, etc., that work under very hard time constraints and are common for all traffic classes, since the available bandwidth resource is huge. It appears reasonable to segregate the bandwidth resource in a clever way and employ less complex protocols and switches which work under performance constraints decided entirely by the applications and not by the network. Also, better user performance are to be expected when compared with the former policy where different traffic classes with different priorities compete for a common resource, and therefore low priority traffic can be severely degraded. We consider, therefore, that an a priori (segregation) partitioning of the network resources for different traffic classes should be used for very high speed networks. There are two choices to share the media resources among various traffic classes, in time domain or in wavelength domain. The first alternative, used by most actual networks like FDDI, MAGNET [20], ATM [3], DTM (Dynamic Time division Multiplexing) [34], etc., presents fundamental limits when WDM is employed in con- junction with sharing of resources in the time domain to expand the network throughput over the speed of electronics, i.e., over 1 Gbit/s. Typi- cally, these networks make use of an in-band model for integrating different traffic in the same high-speed TDM frame. Here the control and the signaling data are time-multiplexed with the information data in a very high speed TDM frame. These two sets of data have different requirements. The former has a higher priority and

needs more reliable transmission and greater processing capabilities. For the latter it is just the opposite, i.e., it requires high speed and high throughput. However, this model has a limited bandwidth available for the control channel(s), and this insufficiency severely restricts the network throughput for information data. The maximum achievable throughput is limited in this case by the product of the packet/cell length of data and the bit rate of the control channel [7,17], and not by the technological limitations (like power- budget) on optical networks. This means that it is almost impossible to make use of a multiwavelength network to its fullest in the case of processing and switching of small cells/packets like, for instance, ATM cell. This model performs well only with extremely long packets (tens to hundreds of kbits), where the optical limitations are approached. Another drawback is the depen- dency of very high speed components. This network model needs very high speed components not only for general purposes, such as clock synchronization, but also for some other particular functions, such as frame/token synchronization, differentiation of control data and information data, etc. These problems can be obviated by using resource partitioning in the wavelength domain. This approach, the so-called wavelength dedicated to application, is proposed for SUPER- LAN (described in the next section). Here the wavelengths are dedicated to different traffic classes, including one or more wavelengths dedicated for control/access purposes. The time syn- chronism among channels belonging to the same application, but placed in different wavelengths, is provided. The stations have permanent access to all wavelengths of interest and the access for both user information data and control data does not occur at the very high speed electro-optical station interface with the optical fiber, but at the low speed electronic interface between the station/media interface and the users attached to it. This provides simplicity and low processing rates for the media access protocol, as well as a minimal number of high speed components. More- over, this approach has the advantage to provide bandwidth resource for control purposes that is limited only by the electronic speed, i.e., up to about 1 Gbit/s. Therefore, a multiwavelength network can be used to its fullest, that is given by the optical technological limits. This is shown in

the SUPERLAN case, where the network performance (throughput, number of stations, distance between stations, etc.) is restricted only by optics and not by electronics or protocol parameters (like packet/ceU size).

The next issue that must be reconsidered in the network design is how to resolve contention among the users, i.e., what switching technique should be employed to switch information among users inside the same class of traffic. This can be provided in the time domain (like in ATM, FDDI, DQDB, MAGNET, etc.) or in the wavelength domain (like in Lambdanet, the Photonie Knock- out Switch, etc.). In spite of the very high switching bandwidth (multi-Gbit/s) provided by the second approach, this has, nevertheless, the dis- advantage that it needs wavelength agile components and, as a result, is not suitable for bursty data. We chose, therefore, a time switching technique for SUPERLAN. Also, because of the fact that the future multi-Gbit/s LAN must provide switching support to a heterogeneous mixture of isochronous and nonisochronous sources, a proper switching approach must be chosen. This means that the switching architecture must provide transfer of data from one user to another user, so that the desired grade of service, such as throughput, delays or blocking performance, are met for all traffic classes. There are two main switching techniques: (i) circuit-switching, with poor bandwidth utiliza-

tion but with relative simplicity of the access mechanism; and

(ii) packet-switching, with high service flexibility and network efficiency but with sophisticated control algorithms.

Taking into consideration the fact that mixed networks (circuit-switching and packet-switching) are very likely to exist for a long time to come, as well as the fact that the internetworking between the existing networks and the future multi-Gbit/s network must be as simple as possible, with no economic or performance penalty, the use of both switching approaches are suggested for the very high speed networks. SUPERLAN has, therefore, a mixed switching architecture, with time domain circuit-switching services provided in one wavelength band for isochronous traffic, and with time domain packet-switching services provided in another wavelength band for nonisochronous traffic.

1097

Finally, the last element of the service bottleneck that must be reconsidered is the access control. At the present time there are a multitude of access control mechanisms in three different categories (fixed assignment, demand assignment or random access), which are provided in various domains (time, frequency/wavelength, etc.), and acting at different temporal scales (call/session, burst, cell/packet). A common feature of most existing networks is the fact that they make use of a common media access procedure for all types of traffic, and this procedure acts either at the cell/packet level or at the cell and call/session levels. This approach is disadvantageous because the control mechanism becomes quite complex. For instance, it is not necessary to have an access mechanism acting at the cell/packet level for isochronous traffic. Here an access mechanism acting at the call or burst level is sufficient. Also, for nonisochronous traffic, an access mechanism acting at the burst level will be simpler than if we chop (packetize) the data block. Because of this complexity, the access control mechanism becomes a real performance bottleneck for broadband services with hundreds of Mbit/s rates. We consider, therefore, that the future very high speed network must have highly simple access protocols. Based on these considerations, we propose a network model in which each traffic class/application has its own application-oriented access protocol, with no interference from other applications. The access protocols can be separated in time or in wavelength domain, and their parameters should be decided based entirely on the application needs. For instance, the temporal frame size for access control should be chosen with respect to the (largest) temporal pe- riodicity of the application in the case of isochronous traffic, but with respect to the lowest (admissible) delay in the case of nonisochronous traffic. Similar considerations are true for contention resolution mechanisms. Schemes based on service denial work well in the case of isochronous traffic type CBR, where the parameter of interest is mean blocking probability. On the other hand, schemes based on scheduling are recommended for nonisochronous traffic. By this concept we get simpler and low speed access protocols. For SUPERLAN we propose, therefore, dedicated medium access protocols for the isochronous traffic (acting at the call/burst level)

1098

and for the nonisochronous traffic (acting at the burst level), that are placed in different wavelengths. The penalty of this approach is hardware replication in each node.

5. SUPERLAN: a model for very high speed LANs

SUPERLAN [27] is an integrated very high speed local area network and is based on a WDMA architecture. The total user traffic on the network is separated into two classes, isochronous and nonisochronous, each of which is allocated two or more wavelengths for the purposes of control and data transfer. Time-synchronous channels are provided in different wavelengths for the transport of isochronous data in w_iso_d and the associated control information (multiaccess, signaling and frame synchronization) in w iso_c, and for the transport of nonisochronous data in w_niso_d and the associated control information (multiaccess and frame synchronization) in w_niso_c. Besides these, there are four more wavelengths. They are dedicated to:

(i) network clock distribution in w_cl; (ii) network management and control in w_mng,

with provisions for decision, activation and deactivation of parameters for PHY, MAC, error control and flow control, under the control of applications;

(iii) error recovery/supervision in w_err; and (iv) flow control in w_tic (for nonisochronous

traffic). The attached stations always have access to all

wavelengths of interest in the network and they provide separation of the local traffic into traffic classes (isochronous and nonisochronous) and different control data before data are transmitted on the network.

SUPERLAN is capable of supporting isochronous and nonisochronous traffic with arbi- trary bit rates, both narrowband and broadband. It provides, in the first alternative, circuit-switching services for isochronous traffic, with perfect recovery at destination, i.e., the contention mechanism is based only on blocking. Moreover, in this model each channel is able to adapt its speed and capacity to the specific bandwidth need of the particular application it supports. This is particu- larly useful for many types of isochronous traffic where the channel speed can be linked to the

coding rate of the signal, reducing or eliminating the need for rate adaptation and multiplexing/ demultiplexing functions inside the network. It also provides packet/burst-switching services for nonisochronous traffic, that are based on delay- throughput trade-off. The network can support switched services at rates up to 10 Gbit/s for each traffic class, for a total network throughput of over 20 Gbit/s. It is also able to provide communicative (for individual communication) and distributive (for mass communication) services that may include audio, video and data signals, and that need point-to-point and/or mul- tipoint communications among a preliminary maximum number of 64 stations with up to 16 substations connected to each station. Each of these stations provide a flexible interconnection to different networks and devices, such as classi- cal data networks that support OSI type protocols, ISDN and B-ISDN/ATM networks, other special purpose systems (e.g., high speed printers, highly parallel computing engines) etc., with throughputs independent of the data rates of SUPERLAN.

The network has, in its first phase, a physical ring configuration with a master station '(MS) and a number of up to 63 ordinary stations (OS). It also has a logical hybrid topology, i.e., logical ring topologies for the w iso d and w niso_d data transport subnetworks, and for the w niso_c and w t i c subnetworks, and logical bus topologies for the other subnetworks. The master station has the responsibility for network supervision (such as frame size determination, clock and frame generation, total loop-length adjustments, etc.) and for network operation control (such as resource allocation and management for isochronous traffic). It also provides communication channels for the local substations.

SUPERLAN is, therefore, comprised of eight logically separate subnetworks, but provides users with the functionality of a single, integrated multi-Gbit/s network. It makes use of eight parallel, wavelength-separated channels with time synchronization provided among subnetworks belonging to the same user traffic class, as described below [28-31]: • the w_cl subnetwork, that provides the distribution of the basic network clock. It has a logical bus topology where the master station generates the basic clock signal of 100 MHz, and the other

stations extract this clock signal and use it for network operation. • the w _ i s o d isochronous data subnetwork, with a maximum capacity of 10 Gbi t /s , that provides the transfer of isochronous user traffic. It has a logical ring topology with an insertion access method based on serial-parallel and parallel-serial registers. In this subnetwork all the stations have the same configuration and functions. The transmission form is by constant-length (125 /xs) TDM frames, but with variable-capacity, ftred-time slots. The slot length is of 10 ns duration, and each slot provides a variable rate ranging from 0.5 Gb i t / s up to 10 Gbi t /s . Thus, as shown in Fig. 2, the duration of one bit and the total number of bits in each time slot can vary. The minimum bit duration is restricted entirely by the modulation bandwidth of the semiconductor lasers used in the system. Accordingly, the maximum number of bits in one time slot is 96 in the case of 10 Gb i t / s rate, or 48 at 5 Gbi t / s , or 24 at 2.5 Gbi t / s , etc. However, the actual number of bits in each time slot is decided only in terms of the bandwidth need of the particular call allocated to that slot. There is no explicit reading of addresses in the block of user information data; they are implicitly contained in the slot

t.. 10 ns b=

l I

I I

~synch Bt B2 Bn =synch

I I I I I I I I I T I i = = = i = i i = t

I . I I I " I I I I t - I

--x GuardV, - - ~Gua rd ~ ' band t = band I

where Bn = 8, with a capacity of 64 kbit/s or 16, with a capacity of 128 kbit/s

or 24, with a capacity of 192 kbit/s



T = 50 ps, for 10 Gbit/s

or 100 ps, for 5 Gbit/s

or 200 ps, for 2.5 Gbit/s, etc.

Fig. 2. Data slot configuration.

1099

............ ! I- ,0os , (,

~ - - 64 kbit/s slot

~ - - 768 kbit/s slot

W - - 384 kbit /s slot

- ~ - - 128 kbit/s s lot

Fig. 3. Example of a data frame structure for isochronous traffic.

positions in each frame, which are decided through a call setup procedure. Figure 3 shows an example of a w_iso_d TDM frame. There are 12500 slots in a 125 /zs frame and every station reads a n d / o r overwrites information data in pre- allocated time slots each time the w_iso_d temporal frame passes through the station. • the w_niso_d nonisochronous subnetwork, with a fixed rate of 10, or 5, or 2.5 Gbi t / s , etc., that provides the transfer of nonisochronous user information. It is a subnetwork similar to the w iso_d subnetwork, in the sense that it has the same logical topology and a similar TDM frame model, but with fixed-capacity, fltxed-time slots. Once again, the slot length is of 10 ns duration, but now each slot carries a fixed number of bits corresponding to the subnetwork bit rate. More- over, the temporal frame size in w _ n i s o d (and, accordingly, the number of time slots in a frame) is adjustable. This is determined by a bidding (claim) procedure on the w_mng subnetwork according to delay-throughput trade-off needs of different nonisochronous traffic carried on w_niso _d. • the w_iso_c isochronous control subnetwork, with a capacity of 100 Mbit /s , that supports resource management information flows as well as connection management (signaling) and control information flows for isochronous traffic. It is a subnetwork running in a time-synchronous fashion with the w _ i s o d subnetwork, and it has a logical bus topology, with an insertion access method based on serial registers. The MS node acts as a transmitter and receiver, and the OS nodes act as repeaters and drop/ inser t intermediate stations. A synchronous TDM transport mode with constant-length frames of 125 /.ts is

1100

used for the transmission of control and signaling data.

A train of 125/zs TDM frames with 100 Mbi t / s rate and (n + 2) time slots in a frame flows con- tinuously along the bus, originating and ending at the MS node. The first slot serves as a frame synchronization header (SH) that contains a fixed-bit pattern used for establishing frame synchronization in each node. The next n time slots contain network control information for actual communication among the network users. They provide the transport of special dedicated control units, called cells, and the number n is chosen accordingly to the media access mechanism. The last time slot (T) is used for auxiliary functions (bits for subnetwork frame synchronization, reservation bits, etc.). Every station has free access to a certain number of time slots in a frame, i.e., they can read a n d / o r overwrite control data in different time slots every time the w _ i s o c temporal frame passes through the node. • the w_niso_c nonisochronous control subnetwork, with a capacity of 100 Mbit /s , that supports resource management information flows as well as connection management and control information flows for nonisochronous traffic. It is a subnetwork running in a time-synchronous fashion with the w_niso_d and w_tic subnetworks. It has a logical ring topology, with an insertion access method based on serial registers. A synchronous TDM transport mode with constant- length frames is used. The size of the w_niso_c TDM frame is exactly the same as of the w_niso_d frame, and is decided on the basis of the bidding procedure. The form of the w_niso_c TDM frame is similar to that of the w iso_c frame, i.e., it contains a first synchronization time slot (for frame synchronization), m time slots (control information) and a last trailer slot (auxiliary functions). • the w_mng, w_err and w_tic subnetworks, that make use of synchronous TDM transport mode similar to the w_iso_c subnetwork. They are still under study and will be reported later.

As mentioned above, the purpose of media admission control is to provide fair access to the network resources for all users, as well as to ensure that the contracted QOS requirements are met for each admitted connection. Based on that, a class of new media access mechanisms, devel- oped for different applications, are reported in

[29,31] for SUPERLAN, and is briefly described below.

A Connection-Oriented (CO) procedure with a centralized media access protocol, that is based on a global scheduling multiple access scheme, and with a contention resolution of type service denial (blocking), is provided for the isochronous traffic. There are three phases in a CO communication procedure: connection establishment; data transfer; and disengagement/termination. These phases have different requirements and they are supported by different subnetworks in SUPER- LAN. The connection and the termination phases are supported by the w_iso_c control subnetwork, and the data transfer by the w_iso_d isochronous data subnetwork. In the first stage, unknown statistics are considered for isochronous traffic. Therefore, an admission control acting at the call level and based on the use of peak rate for different isochronous traffic classes, both CBR and VBR, is used. The isochronous bandwidth resource available in SUPERLAN (up to 10 Gbit /s) is partitioned into separate bandwidth pools, dedicated to different isochronous traffic classes, so as to provide equalization (fairness) of the blocking probabilities among different traffic with different offered loads and bandwidth requirements. Time slots in w_iso_d are assigned to calls based on the peak transfer rate of the call.

There are two choices in allocation of w_iso_d slots to isochronous traffic: (i) an integer number of time slots is allocated to

one call; and (ii) an integer number of slots a n d / o r fractions

of slot are allocated to one call. The first alternative provides a better resource

utilization of the w_iso_c subnetwork at the expense of lower utilization of the w_iso_d subnetwork resources. For the latter, however, it is just the opposite, i.e., it provides a better resource utilization in w_iso_d subnetwork and a worse one in w_iso_c subnetwork. In the future, separation of CBR from VBR traffic (in time or in wavelength) will be considered, and admission control policies based on statistical multiplexing acting at burst level will also be taken into ac- count.

A Connectionless (CL) procedure, with a distributed media access protocol working at a burst level, and with a congestion control mechanism

based on dynamic resource adjustment (increase/decrease), is provided for the nonisochronous traffic. Here, the incoming bursty traffic is ordered into priority classes based on their delay-throughput needs. The most stringent needs are met in the case of distributed computa- tion models, where the transmission of short messages is more crucial to the performance than the transmission of long messages [37]. Fairness is enforced, in this case, by ordering the nonisochronous applications according to their admissible latencies. A dynamic resource-sharing mechanism that is based on a preemptive priority policy is used for allocation of w_niso_d resources to different classes of nonisochronous traffic. Inside the same traffic class, fairness is provided on a first-come-first-served (FCFS) basis. Also, the most stringent latency requirements are used in the determination of media access protocol parameters, such as the temporal frame size in w_niso_c and w_niso_d. This parameter, which represents the maximum access delay in w_niso_c subnetwork, has typical values in the range of 10 tzs up to hundreds of milliseconds. In order to provide a minimum access delay, a reservation scheme [32] is used for the access protocol. The individual stations send reservation messages across the w_niso_c subnetwork to inform the network about their needs for resources in one, or more, of the next frame(s) in the w_niso_d subnetwork. The removal of these reservation messages is done at the source stations. A contention resolution mechanism based on an increase in bandwidth for lower priority traffic is used when different bursts of different priorities are present and the bandwidth resource available in one frame is less than the total requested bandwidth. In this case, the resource is increased by means of buffers inserted in the temporal ring. These buffers are used only for the lower priority traffic in case of congestion. They provide alternative routes for traffic of lower priority, and let the traffic of higher priority pass through the station with no extra delay penalty (temporal multihop procedure). Each station is, therefore, provided with a set of reserve buffers dedicated to traffic of different priorities, except for the traffic of highest priority, which never use them. That means that the latency is traded-off for throughput improvements in case of congestion, i.e., data bursts of different priorities experience

1101

different delays in their propagation through the network. However, the network is dimensioned such as the admissible latencies for all nonisochronous traffic classes are met. In this way, delay-throughput performance requirements are provided for all traffic classes. In a later phase, segregation of nonisochronous traffic classes in different wavelengths, and with distinct multiaccess mechanisms, will be considered.

6. Synchronization issues in SUPERLAN

The feasibility of the multi-wavelength network architecture proposed in Section 5 crucially depends on synchronized transmission on the control and the associated information data subnetworks. Here, we briefly discuss the issues in- volved in synchronizing subnetworks across different wavelengths of the same optical fiber and provide the solution chosen for SUPERLAN. References [28,30] deal with these issues in com- plete detail.

Network synchronization is concerned with problems of distributing time and frequency among remote nodes such as to provide a correct operation of the network. Since the transmission of information by light pulses undergoes distortions of different sorts in the propagation through optical fiber, some form of ranging system is required in order to compensate for these distortions. For instance, the fiber can attenuate, delay or spread the transmitted pulses. Further uncer- tainties in the electrical signal are introduced due to the statistical nature of the opto-electric detection process (based upon photon counting), which can, in addition, be corrupted by additive noise of different nature after photodetection [26]. For SUPERLAN this picture is further complicated by the presence of different subnetworks, placed in different wavelengths, that must work in a time-synchronous fashion. Also, there is a supple- mentary requirement for ring topologies (like w_iso d and w_niso_d) that an integral number of fixed-length data frames/slots be in transit around the circumference of the ring. SUPER- LAN must, therefore, provide detection and synchronization capabilities for error-free transmission in each subnetwork, as well as synchronization facilities at the network level, i.e., among subnetworks belonging to the same application.

1102

Moreover, these facilities should be provided at bit, slot and frame levels.

The very high bit rate (up to 10 Gbit/s) carried on w_iso_d and w_niso_d subnetworks puts the most severe requirement on the choice of detection and synchronization solutions. The very high bit rate limits the intersymbol interference (ISI) distortion and the timing jitter in all stations to less than 50 ps. Given a basic application (that includes data and control channels placed in different wavelengths), the synchronization solution must provide: • a very low bit error rate (BER), of less than 10-12, in each channel, and • compensation of the difference in propagation delays between the data and control subnetworks. This compensation must react on time scales consistent with the processing times in the data and control subnetworks, i.e., slot and frame time scales.

The error performance of a point-to-point optical fiber communication system employing self- time repeaters depends on two components--the signal-to-noise ratio (SNR) in the detector branch, denoted by a (power-budget limit), and the SNR, denoted by /3, in the phase-locked-loop (PLL) branch (spectral limit). Based on the application (network configuration), one of these parameters can dominate and, therefore, specific solutions are needed to provide error-free transmission. For instance, if a is small compared to /3, the cumulative effect--due to pulse interference intersymbol (ISI) in optical fiber, shot noise in avalanche photodetector and thermal noise in receiver/amplifier--is large, and the error performance depends mainly on the choice of the detector solution. We ignore the crosstalk noise across different wavelengths as in the SUPER- LAN case all channels have the same direction of transmission and, therefore, the power differences between them are minor. On the other hand, a low/3 value means that the timing jitter (random and systematic)--due to the combined action of data stream randomness (self-noise) [14], linear distortions of the signal (interference intersymbol), imperfections of the retiming circuit (PLL) and noise in the channel (thermal and shot) [10]--has values approaching the pulse in- terval. In this case the error performance depends mainly on the synchronization solution. Appropriate solutions for detection and time syn-

chronization are, therefore, provided in SUPER- LAN, so as to provide a and /3 values corresponding to the error performance of less than 10-12 at rates up to 10 Gbit/s.

An incoherent transmission technique based on on-off modulation with bit duration at the transmitter, combined with direct detection at the receiver, is used for transmission in SUPERLAN. Multimode Fabry-Perot lasers and avalanche photodiodes (APD) are used for the data channels. For the other channels (clock distribution, media control, flow control, etc.), with speeds up to 100 Mbit/s, p.i.n, diodes and monomode lasers type distributed feedback (DFB) or distributed Bragg reflector (DBR), can be used as well. A minimum of 8 wavelength-multiplexed channels are provided on the fiber. The repeater sections are limited (by ISI distortion) to 2.5 km in the case of 10 Gbit/s rate, or 5 km in the case of 5 Gbit/s rate, etc. A large a margin of 14 dB is left for wavelength mux/demux and other unac- counted (coupling, etc.) losses in the case of data transmission in the 1300 nm window on a maximum repeater section of 5 km. Also, /3 values larger than a are obtained in SUPERLAN with a specific solution that eliminates jitter accumula- tion around the ring and, therefore, the number of stations is not limited by jitter, i.e., it can be unlimited. A jitter-free timing information is provided (for transmission of data and control information onto optical fiber) in all stations, by means of a dedicated channel placed in wavelength w_cl. Master Station generates, with a very precise and stable generator, an analogue sinus signal of 100 MHz, cl_tr, that is transmitted around the w_cl subnetwork. A firm frequency component of 100 MHz (synchronization pulses) is provided in transmitted data (Fig. 2), with guard bands, between the synchronization pulse and the adjacent data pulses, that are sufficiently large to completely eliminate the self-noise and ISI distortions on synchronization pulses. Because of the presence of synchronization pulse together with its guard bands, the bandwidth resource available for data information in w_iso_d and w_niso_d is diminished by 4%, i.e., from 10 Gbit/s down to 9.6 Gbit/s, or from 5 Gbit/s to 4.8 Gbit/s, etc. Low jitter peaks of 20 ps or less are easily obtainable in the receive clock signal cl_rc, for recep- tion of data and control information from optical fiber.

1103

In order to provide time synchronization among channels placed in different wavelengths, compensation of the difference in optical transmission time (group delay), for wavelengths belonging to the same application, must be provided in stations as well. This compensation must be provided in each 10 ns time slot, so that the time slots in the data and the control channels are transmitted simultaneously onto the optical fiber (using cl_tr clock signal), irrespective of the group delays on the precedent regeneration section. This is done in SUPERLAN at the expense of extra cross-node delay on w_iso_d and w_niso_d channels. An insertion access method based on serial-parallel (S/P) and parallel-serial (P/S) registers is used on data channels for the transfer through stations. That means a cross- node delay of 25 ns, which is a sum total of the S /P conversion delay (10 ns), the cross-station delay (5 ns) and the P /S conversion delay (10 ns). This is 2.5 times larger than in the usual case of insertion access method based on serial register (used on the control channels). The cross-node delay in the control channel is given by the slot temporal length in w_iso_c or w_niso_c, and it is equal to 10 ns * n, where n is the number of bits in a control cell (which depends on application). Therefore, the fiber-to-station transfer is done in SUPERLAN by means of S /P register in data channel and serial register in control channel. The cl_rc clock signal, with phase dictated by the incoming data stream on the data channel, is used for this transfer in both channels. On the other hand, the station-to-fiber transfer is done by means of P /S register in the data channel and the same serial register in the control channel with the cl _tr clock signal used in both channels. For providing an error-free transfer through stations, a phase relationship between cl_rc and cl_tr is guaranteed (in stations). The best performance is obtained when cl_tr samples data exactly in the middle of their eye diagram, i.e., cl_ tr is delayed, by means of delay lines, to 5 ns with reference to cl_rc. However, this requirement limits the maximum allowable difference between group delays, during the propagation on the fiber (between two stations), on w_cl and the data or control channels (group delay limits). Similar to the power-budget limit or the spectral limit, mentioned before, group delay requirements must be provided in SUPERLAN for an

error-free slot synchronization among channels belonging to the same application but placed in different wavelengths. An error-free slot synchronization is completely guaranteed in the case of single-mode optical fibers whose dispersion characteristics are such that the group delay varia- tions are flattened to a limit of maximum 0.8 ns/km over a broad wavelength range, covering both 1300 nm and 1550 nm windows.

The last synchronization facility that must be provided in SUPERLAN is between the data and the control subnetworks (belonging to the same application) at a frame level, e.g., at 125 /zs temporal length for isochronous traffic. Because of different cross-node delays on the data and the control channels, higher transport speeds are expected for data channels in propagation around the ring. Therefore, different frame delimitation moments are obtained for the data and the control channels in different stations according to their position in ring against the Master Station (where these moments coincide). Moreover, based on the total network delays on the data and the control channel, more temporal loops may be available in the data channel during one temporal loop in the control channel. Because of that, each station, after receiving the synchronization header (frame delimitation moment) in the control subnetwork, must know exactly when it has to expect the next frame delimitation moment in the data subnetwork. It must also know the number of temporal loops available for the data channel during one frame. SUPERLAN is a network model where the frame synchronization is provided irrespective of the ratio of the frame size to total network delay in the data or control channel. This is done in the Master Station by continuous measuring (using the synchronization header in the control frame) of the total network delay in the control subnetwork and, based on this, by computing the number of loops available for data channel in a frame. Master Station also provides, by means of the trailer slot (T) in control frame, the Ordinary Stations with the needed information for frame synchronization. Based on this information, and on the frame delimitation moments (learned from the synchronization header in control channel), each station is able to provide with four dedicated counters right time slot(s) for different calls/sessions in progress on data channel.

1104

7. Conclusions

It is widely recognized that the two main bottlenecks that limit the performance of multi- Gbit/s LANs are the electro-optic and service bottlenecks. We have motivated and presented a novel architectural solution which opens up these bottlenecks and provides performance limited only by the optical fiber technology. The distin- guishing features of this architecture include: • utilization of a coarse WDM to open up the electro-optic bottleneck in the network nodes; • separation of the low rate, highly reliable control data from the very high speed information data, and of the isochronous traffic from the nonisochronous traffic; and • separate and simplified protocols which can support packet- and circuit-switching as dictated by the supported applications.

The novelty of this model is given mainly by the idea of allocating different wavelengths to different applications, thereby making use of the abundant bandwidth available on the fiber. This design concept shows that the logic complexity of a Gbit/s LAN can be greatly minimized and reduced to that of the low speed electronic elements. Equally important, this model is very suitable for applications that integrate narrowband and broadband services and that may include video (both CBR and VBR), image, data, voice, etc., each of them with its own traffic characteristics and performance requirements. Such a flexible arrangement permits growth in applications on the network without interfering with other existing applications, and avoids most of the performance pitfalls of other integrated networks mentioned earlier in the paper. Other advantages are due to the following: • it is possible to develop application-oriented protocols, with no interference from other applications; • there are no wavelength-agile components in the network; and • it is possible to develop networks with performance limited by technological limitations on optical fiber only, and not by electronics or control mechanisms.

A factor which works against such an architecture is hardware replication in each node for each wavelength. But perhaps the SUPERLAN architecture may not represent a cost penalty

over the single integrated LAN architectures when all the ancillary functions are considered.

The emerging multichannel systems based on dense WDM will favor this architecture as the number of applications on the network grows. Our future efforts will be directed towards a multi-Gbit/s network model with more subnetworks, each of them completely dedicated to a specific type of service (voice, video, graphics, data, facsimile, client-server, etc.), with optimized physical/virtual topology and simple, low-delay, application-oriented media access and transport protocols.

References

[1] A.S. Acampora, A multichannel multihop local lightwave network, Proc. IEEE GLOBECOM'87, Tokyo, Japan, November 1987, pp. 37.5.1-37.5.9.

[2] A.S. Acampora, A high capacity metropolitan area network using lightwave transmission and time-multiplexed switching, IEEE Trans. Comm. 38(10) (1990) 1761-1770.

[3] J.J. Bae and T. Suda, Survey of traffic control schemes and protocols in ATM networks, Proc. IEEE 79(2) (1991) 170-189.

[4] J.J. Bae, T. Suda and N. Watanabae, Evaluation of the effects of protocol processing overhead in error recovery schemes for a high-speed packet switched network: link- by-link versus edge-to-edge schemes, IEEE J. Selected Areas Comm. 9(9) (1991) 1496-1509.

[5] J.A. Bannister, The wavelength-division optical network: architectures, topologies, and protocols, Ph.D. Disserta- tion, University of California, Los Angeles, 1990.

[6] C.A. Brackett, Dense wavelength division multiplexing networks: principles and applications, IEEE J. Selected Areas Comm. 8(6) (1990) 948-964.

[7] C.A, Brackett, On the capacity of multiwavelength optical-star packet switches, IEEE J. Lightwave Telecomm. Systems 2(2) (1991) 33-37.

[8] I. Chlamtac and A. Ganz, Toward alternative high speed networks: the SWIFT architecture, Proc. IEEE INFO- COM'87, San Francisco, CA, March 1987, pp. 1102-1108.

[9] I. Chlamtac, A. Ganz and G. Karmi, Circuit switching in multi-hop lightwave networks, Proc. ACM SIGCOMM'88, Stanford, CA, August 1988, pp. 188-199.

[10] D.L. Duttweiler, The jitter performance of phase-locked loops extracting timing from baseband data waveforms, Bell System Tech. Z 55(1) (1976) 37-58.

[11] M. Eisenberg and N. Mehravari, Performance of the multichannel multihop lightwave network under nonuniform traffic, IEEE Z Selected Areas Comm. 6(7) (1988) 1063-1078.

[12] K.Y. Eng, A photonic knockout switch for high-speed packet networks, [EEE J. Selected Areas Comm. 6(7) (1988) 1107-1116.

[13] G. Gallassi, G. Rigolio and L. Verri, Resource manage-

ment and dimensioning in ATM networks, IEEE Net- work 4(3) (1990) 8-17.

[14] F.M. Gardner, Self-noise in synchronizers, IEEE Trans. Comm. 28(8)(1980) 1159-1163.

[15] M.S. Goodman, Multiwavelength networks and new approaches to packet switching, IEEE Comm. Mag. 27(10) (1989) 27-35.

[16] P.E. Green, The future of fiber-optic computer networks, Computer 24(9) (1991) 78-87.

[17] P.S. Henry, High-capacity lightwave local area networks, IEEE Comm. Mag. 27(10) (1989) 20-26.

[18] M.J. Karol and S. Shaikh, A simple adaptive routing scheme for shuffieNet multihop lightwave networks, Proc. IEEE GLOBECOM'88, November 1988, pp. 1640-1647.

[19] H. Kobrinski and K. Cheung, Wavelength-tunable optical filters: applications and technologies, IEEE Comm. Mag. 27(10) (1989) 53-63.

[20] A.A. Lazar, A.T. Temple and R. Gidron, MAGNET II: a metropolitan area network based on asynchronous time sharing, IEEE J. Selected Areas Comm. 8(8) (1990) 1582- 1594.

[21] T.P. Lee and C. Zah, Wavelength-tunable and single- frequency semiconductor lasers for photonic communications networks, IEEE Comm. Mag. 27(10) (1989) 42-52.

[22] T.T. Lee, M.S. Goodman and E. Arthurs, A broadband optical multicast switch, Proc. ISS'90, Stockholm, Swe- den, May 1990, III (1990) 7-13.

[23] W.P. Lidinsky, Data communications needs, 1EEE Net- work 4(2) (1990) 28-33.

[24] N.F. Maxemchuk, Regular mesh topologies in local and metropolitan area networks, AT&T Tech. J. 64(7) (1985) 1659-1685.

[25] N.F. Maxemchuk and M. EIZarki, Routing and flow control in high-speed wide-area networks, Proc. IEEE 78(1) (1990) 205-221.

[26] S.D. Personick, Optical Fiber Transmission Systems (Plenum Press, New York, 1981).

[27] A. Popescu and R.P. Singh, SUPERLAN: a model for very high speed local area networks, in: M.J. Johnson, ed., Protocols for High-Speed Networks, If, Proc. IFIP

1105

WG6.1/WG6.4 2nd International Workshop, Palo Alto, CA, 27-29 November 1990 (North-Holland, Amsterdam, 1991).

[28] A. Popescu and R.P. Singh, Synchronization issues in Superlan, Technical Report TRITA-TCS-TS-9206-TR, Royal Institute of Technology, Stockholm, Sweden, 1992.

[29] A. Popescu and R.P. Singh, Media access mechanisms in Superlan, Technical Report TRITA-TCS-TS-9207-TR, Royal Institute of Technology, Stockholm, Sweden, 1992.

[30] A. Popescu and R.P. Singh, Synchronization issues in WDMA networks, submitted to the 4th IFIP WG 6.4 Conference on High Performance Networking, Liege, Belgium, December 1992.

[31] A. Popescu, R.P. Singh, H. Rajaei and Y. Ismailov, Media access mechanisms in multiwavelength networks based on the wavelength-dedicated-to-application concept, submitted to the 4th IFIP WG 6.4 Conference on High Performance Networking, Liege, Belgium, Decem- ber 1992.

[32] I. Rubin and J.E. Baker, Media access control for high- speed local area and metropolitan area communications networks, Proc. IEEE 78(1) (1990) 168-203.

[33] J.A. Salehi, Emerging optical code-division multiple access communications systems, IEEE Network 3(2) (1989) 21-30.

[34] D.R. Spears, Broadband ISDN switching capabilities from a service perspective, IEEE J. Selected Areas Comm. 5(8) (1987) 1222-1230.

[35] H. Toba, K. lnoue and K. Nosu, A conceptional design on optical frequency-division-multiplexing distribution systems with optical tunable filters, IEEE J. Selected Areas Comm. 4(9) (1986) 1458-1467.

[36] F. Vakil and H. Saito, On congestion control in ATM netwo(ks, IEEE Z Lightwace Telecomm. Systems 2(3) (1991) 55-65.

[37] H. Wu, L.-E. Thorelli and A.-H. Smai, Flow control support for efficient realization of the EDA model, J. Microprocessmg and Microprogramming 34 (1992) 175- 178.

an alternative solution to the electro-optic and service bottleneck problems in integrated...

Documents