Daisy: a scalable all-optical packet network with ... - Semantic Scholar

Computer Networks and ISDN Systems 30 Ž1998. 1065–1082

Daisy: a scalable all-optical packet network with multifiber ring topology 1 M. Ajmone Marsan ) , A. Fumagalli, E. Leonardi, F. Neri, P. Poggiolini Dipartimento di Elettronica, Politecnico di Torino, 10129 Torino, Italy

Abstract This paper presents Daisy, a scalable all-optical packet network where each node is equipped with one wavelength-tunable transmitter and one fixed-wavelength receiver. Network scalability is obtained with a novel multifiber ring topology that allows spatial reuse of wavelengths and requires an optical transmitted power proportional typically to the square root of the number of nodes. The scalability of the proposed topology is analytically evaluated by taking into account the characteristics of state-of-the-art optical components. The topology of Daisy provides one logical channel per destination node. Each channel is shared in statistical time division by all nodes transmitting to a given destination. A channel inspection capability at each node allows the implementation of efficient slotted ordered access protocols. As an example, a simple and efficient protocol named SRR ŽSynchronous Round Robin. is described, and its performance is assessed by means of simulation. q 1998 Elsevier Science B.V. All rights reserved. Keywords: Photonic networks; All-optical networks; WDM rings; MAC protocols

1. Introduction According to the taxonomy introduced by Green in w1x, communication networks of the third generation use photonic technologies for the implementation of switching as well as transmission functions, and are thus termed all-optical. This alleviates the electronic bottleneck resulting from the mismatch between the limited speed of electronic switching and the much higher speed of optical transmission Žwhich is about four orders of magnitude faster. in

)

Corresponding author. http:rrwww1.tlc.polito.it. This work was supported in part by the Italian National Research Council through ‘‘Progetto Finalizzato Trasporti 2’’, and the Italian Ministry for University and Scientific Research. 1

the second generation networks of today. In all-optical networks, data generated in the electronic domain by the source user are converted into the optical domain at the source user-network interface ŽUNI., and transported to the destination with no back conversion into the electronic format before the destination UNI. A commonly used approach to exploit the enormous capacity of all-optical networks is to partition the optical bandwidth into a number of channels whose rate matches the speed of the electronic interface. This can be obtained with wavelength division multiplexing ŽWDM. w2,3x. Currently, the state-ofthe-art optical technology allows the separation of a number of WDM channels of the order of few dozens Žusual numbers range between 8 and 32; experiments were run with up to 100 wavelengths.,

0169-7552r98r$19.00 q 1998 Elsevier Science B.V. All rights reserved. PII: S 0 1 6 9 - 7 5 5 2 Ž 9 8 . 0 0 0 1 0 - 5

1066

M. Ajmone Marsan et al.r Computer Networks and ISDN Systems 30 (1998) 1065–1082

that by itself is not sufficient for the realization of large networks. Therefore, WDM must be combined with either space, or time diversity, or both, if a large user population must be served w3,4x. In the wide area network ŽWAN. scenario, WDM lightwave paths were proposed to provide point-topoint optical connections between randomly chosen node pairs in response to slow changes of traffic distribution w5–7x, and to implement logical topologies on top of arbitrary physical topologies w8x. These approaches provide for wavelength reuse in different portions of the network, thus allowing several thousands of users to be connected with a limited number of wavelengths Že.g., 8 in w7x.. In the local and metropolitan area network ŽLAN and MAN. scenarios, WDM-based networks were extensively investigated with the aim of providing very-high-speed all-optical packet communications among several users. Two survey papers describing the state of the art of WDM-based LANs were published by Mukherjee w9,10x. The first proposals of all-optical WDM networks mainly considered star topologies. In recent years, ring topologies have also become an attractive solution for all-optical WDM networks, thanks to the successful progress achieved in optical amplifiers, which can compensate for insertion losses at intermediate nodes. Rings offer a number of advantages over stars: they allow slot synchronization even at extremely high data rates, hence they offer an efficient and flexible use of the available optical bandwidth for packet communications. Moreover, the ring topology allows the implementation of collision-free access protocols with no need for distributed algorithms or fixed access schemes which are required in star topologies w9x. Finally, rings can allow larger throughputs to be achieved through spatial reuse of wavelengths. Experimental WDM all-optical ring networks are reported in the literature w12–18x. However, ring networks based on a single fiber are not easy to scale to large numbers of nodes, mainly due to two reasons: Ø The practical number of wavelength channels that can be multiplexed in a single fiber cannot be large with currently available devices, due to the limitations in optical amplifier bandwidth and the

need for safe interchannel spacing Žsee w4x.. Hence, if each node in the network is capable of exchanging an amount of traffic close to the capacity of one channel, then the number of nodes allowed in this network without significant throughput penalties cannot be much larger than the number of wavelengths. Ø Although optical amplifiers can compensate for the optical power loss incurred at each intermediate node traversed by the packet, there is a limitation to the maximum number of amplifiers that a packet can cross before the accumulated Amplified Spontaneous Emission ŽASE. noise impairs the correct reception of the packet content w19x. This second drawback severely limits the maximum number of nodes in all-optical ring topologies, regardless of the number of wavelength channels and the amount of traffic generated by each node. In this paper we present Daisy, a novel all-optical network with a multifiber ring topology that permits network scalability through space diversity in the following sense: Ø The multifiber approach of the Daisy topology can accommodate a number of nodes proportional to the number of wavelengths times the number of fibers. Ø In Daisy, only a fraction of the intermediate nodes encountered by a packet on its path to the destination introduces substantial insertion loss penalties, thus relaxing the constraint on the maximum number of nodes in the network imposed by the noise generated by optical amplifiers. Daisy jointly exploits wavelength, space, and time diversity. Indeed, only one ‘‘bi-channel’’ Ždefined by a pair of fibers and a wavelength – see the next section. is provided for, and dedicated to, transmissions to a given destination. Hence, all source nodes that need to transmit to a specific destination node must share the corresponding bi-channel through a slotted MAC protocol. The selection of the wavelength-fiber pair Žbichannel. leading to the desired destination is done at the source node using one fast wavelength-tunable transmitter and one fast space switch. Both these devices were already demonstrated in laboratory setups Žsee for example w20–23x.. The tuning delays


feasible today are compatible with the requirements of Daisy: for example, in w23x an optical transmitter capable of tuning to one of 16 channels, each operating at 1 Gbitrs, in less than 5 ns is described. To allow for multiple access to the same bi-channel in a collision-free manner, many slotted MAC protocols can be designed exploiting the channel inspection capability provided by the Multiple Subcarrier Signaling ŽMSS. technique w24–27x. Several different MAC protocols were designed for Daisy, and their efficiency was assessed with both analytical and simulation approaches. Results were reported elsewhere w28–31x. In this paper we describe a rather simple, but efficient MAC protocol, and we illustrate its performance with some simulation results. The paper is organized as follows. The next section contains a description of the Daisy topology. Section 3 illustrates the structure of the optical portion of each node of Daisy networks. Section 4 comments on the advantages of the chosen topology, on its scalability and on the complexity of the node implementation. A simplified version of the multiring topology is also presented in Section 4; this simpler topology may be easier to implement, but it is shown to be substantially less scalable than Daisy. Section 5 illustrates the SRR ŽSynchronous Round Robin. MAC protocol and presents some performance results obtained by simulation. Finally, Section 6 contains some concluding remarks. It is worth emphasizing that, unlike most papers on optical networks that just address one or at most very few aspects of an optical network design, this

1067

paper considers Daisy from a number of different viewpoints, referring to topology, optical components, scalability resulting from transmission impairments, and MAC protocol effectiveness.

2. The Daisy topology Daisy is an all-optical multifiber ring network with M s N = K nodes, 2 N being the number of unidirectional fibers that run parallel to one another Že.g., clockwise., and K being the number of available wavelength channels per fiber. A special case of Daisy is obtained when N s K, so that M s N 2 . The 2 N fibers in the ring are logically divided into two groups: the group of inner fibers A s a 0 , . . . ,a Ny14 , and the group of outer fibers B s b 0 , . . . ,bNy1 4 . Let the ith node be named n i , with 0 F i F M y 1. Each inner fiber a j g A is connected to all nodes n 0 ,n1 , . . . ,n My1. The outer fiber bj g B is connected to any node n i such that < i < N s j, where < P < N is the modulo N operator. In Fig. 1 two possible topology configurations for a 16-node Daisy are shown: on the left-hand side with N s 2 and K s 8 Ž4 fiber rings., and on the right-hand side with N s K s 4 Ž8 fiber rings.. It is interesting to note how nodes connected to fiber b < jq1 < N can be simply obtained with a clockwise rotation of one node position of the nodes connected to fiber bj . The Daisy topology shares some similarity with the meshed ring topologies Žalso called chordal rings. discussed in w32–38x, but in those cases only one

Fig. 1. A 16-node Daisy topology with N s 2 and K s 8 Žleft., and one with N s K s 4 Žright..

1068


Fig. 2. The paths of the optical signal from the source node n s to the destination node n d through the intermediate node n x in the two 16-node Daisy topologies of Fig. 1.

inner ring is present, and outer connections only provide increased reliability and smaller hop counts. The multiring topology was also considered in w39x, but for a multihop Žhence not all-optical in our sense. network where nodes are equipped with multiple transmitters and multiple receivers. Transmissions are slotted, and nodes transmit fixed length packets Žor cells. that fit exactly into one slot. All packets are transmitted using one fiber in group A Žsay a i ., and are received from one fiber in group B Žsay bj .. Both fibers are determined by the packet destination. A packet is switched from fiber a i , on which it was transmitted, to fiber bj , from which it will be received at the destination node, by a passive device Žcoupler. operating in the optical domain, i.e., packets need not be converted

into an electronic signal, and the switching operation requires no active device. We call channel the ordered pair Ž f, l., where f is the fiber, and l is the wavelength. In all, a total of 2 NK s 2 M channels exist in the network. Each node n d is equipped with a single receiver, which receives packets from channel Ž b < d < N , l? d r N @ .. We call bichannel the pair of channels wŽ a < d < N , l ? d r N @ ., Ž b < d < , l? d r N @ .x through which node n d can be N reached. In all, a total of NK s M bi-channels exist in the network. The mapping of bi-channels onto nodes is such that each bi-channel is in one-to-one correspondence with a receiver. When node n s wants to transmit to node n d , it has to transmit onto channel Ž a < d < N , l? d r N @ .. Thus, in order to reach any given node n d , the source node

Fig. 3. The trees showing the paths followed by all sources to reach node n 0 in the two 16-node Daisy topologies of Fig. 1.


needs to inject the optical signal onto the appropriate inner fiber a < d < N , and tune its transmitter to the appropriate frequency l? d r N @ . When the packet reaches an intermediate node n x Žthe first node with < x < N s < d < N encountered by the packet in its path., it is statically routed to channel Ž b < d < N , l? d r N @ ., which finally leads to destination. Note that in this routing operation no wavelength conversion is necessary. As an example, Fig. 2 shows the two fibers a < d < N and b < d < N in the 16-node Daisy topologies of Fig. 1. Denoting by n s the source node, n x the intermediate node, and n d the destination node, the packet propagation follows the path outlined in the figure Žthick curve. assuming that optical signals propagate clockwise in the multiring. Fig. 3 shows the tree formed by the paths followed by all sources to reach destination n 0 in the two 16-node Daisy topologies of Fig. 1. The maximum number of nodes crossed between source and destination is 7 in the topology on the left, and 5 in the topology on the right. In the general case of a network with N inner and outer fibers, and K wavelengths, the maximum number of nodes

1069

crossed between source and destination is Ž N y 1. q Ž K y 2.. All packets transmitted in each bi-channel are removed at the receiver, thereby avoiding optical signal recirculation. This means that the topology can be logically viewed as a set of staggered busses, each bus corresponding to one bi-channel. A generalization of the Daisy topology is possible, introducing a third group of fibers Žand then a fourth, and so on. that allows even longer jumps along the ring. Although this might lead to scalability advantages, it surely leads to more complex node architectures, that are probably beyond what may be practical with present optical technology. These generalizations are left for future research.

3. The node structure Fig. 4 shows the structure of node n 4 in the 16-node Daisy topology with N s 2 and K s 8 Žshown on the left-hand side of Fig. 1..

Fig. 4. The architecture of node n 4 in a 16-node Daisy topology with N s 2 and K s 8, and its connections to the multifiber ring.

1070


Consider first the upper right part of the figure. As a destination node Ž n d ., n 4 receives every packet arriving from channel Ž b 0 , l2 .. To implement this operation, a fixed optical filter which extracts wavelength l2 is placed on fiber b 0 . Note that this extraction avoids any problem relating to the recirculation of signals and noise, contrary to other optical ring topologies Žin Daisy, any packet transmitted on any wavelength cannot travel farther than one whole circulation around the ring, because any information travelling on a given wavelength on an outer fiber is extracted at the destination identified by the wavelength-fiber pair.. As a source node Ž n s ., n 4 may need to transmit either to fiber a 0 or to fiber a1 , depending upon the packet destination. As shown in Fig. 4, a 0 can be accessed directly, whereas fiber a1 is accessed through coupler C1. As an intermediate node Ž n x ., n 4 routes packets arriving from fiber a 0 to fiber b 0 using coupler C0 . Note that packets either routed to fiber b 0 from fiber a 0 or arriving from fiber b 0 are all directed to nodes n d , with < d < 2 s 0, which includes node n 4 itself. The specific destinations are selected through the K s 8 available wavelengths. Consider now the upper left part of Fig. 4. Four splitters Ž S a 0 ,S a1,Sb 0 ,Sb 1 . tap a small fraction of light from each fiber. The tapped light is received by two photodetectors, D 0 and D 1 , and the resulting electronic signals are further processed to detect packet arrivals from the two pairs of fibers. Since the access control logic is implemented in the electronic domain, delay lines are used to delay the remaining portion of the optical signals to provide enough time for the node electronics to decide whether or not to transmit a new packet in the slot passing by. Detection of packet arrivals is obtained through the Multiple Subcarrier Signaling ŽMSS. technique proposed in w24–27x. In MSS one of K available subcarriers Ž s0 , s1 , . . . , sKy1 . is used to encoderdecode the packet header information, while the packet payload is transmitted at baseband. In Daisy the only useful information of the packet header is the destination, which is thus simply encoded by a subcarrier tone, i.e., a packet directed to node n d is transmitted with subcarrier tone s ? d r N @ . The source node simultaneously transmits the packet payload and the subcarrier tone associated with the destination node using the

same Žlarge bandwidth. transmitter laser so that no additional transmitter is necessary for signalling. The lower left part of the figure shows the two photodiodes D 0 , and D 1 , receiving the optical signals tapped by the 4 splitters Ž S a 0 ,S a1,Sb 0 ,Sb 1 .. The light signals arriving from fibers a j and bj are merged, and detected by photodiode Dj . Only one photodiode output is selected at a time, depending upon the destination of the packet that node n 4 wishes to transmit next. The selected output is then filtered by a tunable passband filter to isolate the subcarrier tone that identifies the destination of the packet that node n 4 wishes to transmit. A threshold comparator is finally used to detect the effective presence of the subcarrier tone, which disactivates the enable signal. Note that both the selector and the filter are set by the controller on a packet-bypacket basis. The transmitter components of the node are depicted on the lower right part of Fig. 4. The node controller is informed of the destination Žnode n d . of the packet selected for transmission and stored in buffer data. The wavelength for the tunable laser, the subcarrier tone for the local oscillator, and the output for the switch Žspace multiplexer., are selected by the controller according to the packet awaiting transmission, and to the MSS technique described above. Packet transmission is finally triggered by the enable signal generated by the controller which also frees the transmitter buffer data. The node structure in Daisy is a function of the network size and of the values chosen for N and K. For instance, Fig. 5 depicts the structure of node n 4 in the 16-node Daisy topology with N s K s 4. Compared to the node architecture shown in Fig. 4, this second node structure on the one hand reveals an increased complexity due to the larger number of fibers in the network, but on the other hand requires a smaller number of wavelengths, hence a simpler tuning hardware. The optimal node configuration is thus practically determined by the actual technology, which makes one approach easier than the other. Note that the use of slotted MAC protocols requires the distribution of slot timing information Ži.e., the definition of slot boundaries.. The node architecture that was just described does not account for the additional hardware needed for the generation and the extraction of this timing information. Several


1071

Fig. 5. The architecture of node n 4 in a 16-node Daisy topology with N s K s 4, and its connections to the multifiber ring.

approaches can be used for this purpose, possibly again exploiting the subcarrier signaling technique. The extraction of a slot clock can be based on one or more of the signals tapped by splitters S a i and Sb i . The generation of a slot clock can either be implemented by a specific device inserted along the multiring, or by one or more nodes, in a distributed fashion.

4. Network scalability As already pointed out, the two main factors limiting the scalability of optical ring networks are the number of wavelengths per fiber and the attenuation of the optical signal.

Optical losses can be compensated for through amplification. Erbium-doped fiber amplifiers ŽEDFAs. are good candidates in WDM networks thanks to their large bandwidth, but there is an upper limit to the total amount of loss that can be recovered before noise builds up and submerges the signal. The most important noise source with respect to this phenomenon is the amplified spontaneous emission ŽASE. noise produced by optical amplifiers. A single Erbium-doped fiber amplifier ŽEDFA. features approximately a 35 nm bandwidth. This value substantially shrinks when several amplifiers are cascaded w40,41x. Moreover, current WDM technology is believed to allow only for a safe minimum channel spacing of 1 nm w4x. The combination of these factors leads to the widely accepted estimate of

1072


an upper limit of about 16 to 32 WDM channels per fiber, when several cascaded amplifiers are used w40x. Larger values are conditioned to advances in dense WDM technology andror flatter gain of cascaded fiber amplifiers w42x. Another potential source of limitations to the scalability of the considered networks is the onset of non-linear effects at high per-channel powers. In this work we neglect these effects since we assume to operate at per-channel power levels that cannot give rise to substantial non-linear degradation. Several experimental WDM ring architectures were recently reported Žsee for example w11–18x.. Some of these require as many wavelengths as the number of nodes M and this, in view of the small number of available WDM channels, severely curtails the scalability of those approaches. In Daisy, the number of needed wavelengths is equal to K s MrN and therefore, by adding fibers, the problem of wavelength availability can be solved by trading frequency diversity for space diversity.

Since a substantial contribution to optical loss is given by the traversing of a node, the fact that in Daisy the outer fibers only go through K s MrN nodes largely reduces the maximum amount of loss that must be compensated for through amplification Žfor a given total number of nodes M .. As a result, less noise is accumulated, and, from this viewpoint too, Daisy is more scalable than all those topologies where a single fiber Žor several fibers. goes through all nodes. In this section, we assess quantitatively the scalability potential of Daisy through an analysis that takes into account both the limited number of available WDM channels and the accumulation of ASE noise. We also present a comparison of Daisy with the multiring system whose node structure is depicted in Fig. 6. The hardware structure of the node in such a multiring is directly derived from Daisy by removing the outer fibers and by coupling the transmitter light directly into the N inner fibers. The reader can easily verify that this network can be run

Fig. 6. The node architecture in the simplified multiring topology, with N s 4, and its connections to the multifiber ring.


using the same protocols that have been designed for Daisy, so that any comparison between this reduced multiring structure and Daisy needs to be made considering only scalability or hardware complexity issues. This simpler system, which is essentially a subset of the Daisy topology, trades off some of the scalability of Daisy for a lower overall complexity. Like Daisy, it only requires K s MrN wavelengths, but its scalability is not as good because of the higher total loss that light signals incur. To perform the scalability analysis we specialize to both network structures a mathematical method that was developed in w27x to analyze generic WDM packet networks employing subcarrier signaling and cascaded optical amplification. In Appendix A we report all the data that would enable the reader to perform the same specialization and carry out the calculations himself. Several assumptions had to be made on component performance and device structure: they are in line with current state-of-the-art proÕen technologies. We also imposed very conservative constraints on required system margins. One important assumption is that amplification takes place at each node only on those fibers whose light goes through the node receiver, as shown by the dashed boxes labeled EDFA in Figs. 4–6. This means that the number of amplifiers on each fiber is equal to K, and that the total number of amplifiers in Daisy is equal to 2 M, while it is equal to M in the simplified multiring of Fig. 6.

1073

Table 1 shows the maximum number M of nodes in Daisy for different values of the individual node bit rate and for different numbers of fibers N, under the constraint that at most 16 wavelengths can be simultaneously present in each fiber Ži.e., K F 16.. For a given value of N, M s NK is obtained by deriving the largest allowable value of K as follows. The maximum permitted value of K Ži.e., K s 16. is first introduced in the model outlined in Appendix A, and the bit error probability is estimated Žtaking into account both optical noise and electrical noise within receivers.. If the resulting value is less than 10y9 , the value K s 16 is accepted. Otherwise, the value of K is progressively decremented until the bit error probability becomes smaller than 10y9 . For a given node bit rate, the maximum value in the corresponding row of Table 1 is emphasized: this is the maximum number of nodes that Daisy can support at that bit rate. It corresponds to the couple Ž N, K . which can be easily identified Ž N is the column index, while K s MrN is given within parentheses.. Similar tables can be computed for different constraints on the maximum number of wavelengths in each fiber, both for Daisy and for the simplified multiring of Fig. 6. Table 2 shows the largest permitted values of M, together with the corresponding values of N and K, at different node bit rates, for Daisy and the simplified multiring, under the two constraints K F 8 and

Table 1 Maximum number of nodes M, and the corresponding value of K, in Daisy, for different values of the individual node bit rate and for different numbers of fibers, under the constraint that at most 16 wavelengths can be simultaneously present in each fiber Gbitrs

1 2 2.5 3 4 5 6 7 8 9 10

M ŽK. Ns5

Ns6

Ns7

Ns8

Ns9

N s 10

N s 11

N s 12

80 Ž16. 80 Ž16. 80 Ž16. 80 Ž16. 80 Ž16. 80 Ž16. 75 Ž15. 70 Ž14. 65 Ž13. 60 Ž12. 60 Ž12.

96 Ž16. 96 Ž16. 96 Ž16. 96 Ž16. 96 Ž16. 90 Ž15. 78 Ž13. 72 Ž12. 72 (12) 66 Ž11. 60 Ž10.

112 Ž16. 112 Ž16. 112 Ž16. 112 Ž16. 105 (15) 91 (13) 84 (12) 77 Ž11. 70 Ž10. 70 (10) 63 Ž9.

128 Ž16. 128 Ž16. 128 (16) 120 (15) 104 Ž13. 88 Ž11. 80 Ž10. 80 (10) 72 Ž9. 64 Ž8. 64 (8)

144 Ž16. 144 (16) 126 Ž14. 117 Ž13. 99 Ž11. 90 Ž10. 81 Ž9. 72 Ž8. 72 Ž8. 63 Ž7. 63 Ž7.

160 Ž16. 130 Ž13. 120 Ž12. 110 Ž11. 90 Ž9. 80 Ž8. 80 Ž8. 70 Ž7. 60 Ž6. 60 Ž6. 60 Ž6.

176 (16) 121 Ž11. 110 Ž10. 99 Ž9. 88 Ž8. 77 Ž7. 66 Ž6. 66 Ž6. 55 Ž5. 55 Ž5. 55 Ž5.

156 Ž13. 108 Ž9. 96 Ž8. 96 Ž8. 72 Ž6. 72 Ž6. 60 Ž5. 60 Ž5. 48 Ž4. 48 Ž4. 48 Ž4.


1074

Table 2 Maximum number of nodes M, and the corresponding values of N and K, in Daisy and in the simplified multiring for different values of the individual node bit rate, under the constraint that at most 8 or 16 wavelengths can be simultaneously present in each fiber Gbitrs

M Ž N, K . K F8

1 2 2.5 3 4 5 6 7 8 9 10

K F16

Daisy

Multiring

Daisy

Multiring

112 Ž14,8. 104 Ž13,8. 96 Ž12,8. 96 Ž12,8. 88 Ž11,8. 80 Ž10,8. 80 Ž10,8. 72 Ž9,8. 72 Ž9,8. 64 Ž8,8. 64 Ž8,8.

72 Ž9,8. 64 Ž8,8. 64 Ž8,8. 64 Ž8,8. 56 Ž7,8. 56 Ž7,8. 56 Ž7,8. 49 Ž7,7. 49 Ž7,7. 49 Ž7,7. 48 Ž6,8.

176 Ž11,16. 144 Ž9,16. 128 Ž8,16. 120 Ž8,15. 105 Ž7,15. 91 Ž7,13. 84 Ž7,12. 80 Ž8,10. 72 Ž6,12. 70 Ž7,10. 64 Ž8,8.

112 Ž7,16. 98 Ž7,14. 96 Ž6,16. 96 Ž6,16. 84 Ž6,14. 80 Ž5,16. 75 Ž5,15. 70 Ž5,14. 65 Ž5,13. 65 Ž5,13. 60 Ž5,12.

K F 16. Fig. 7 plots the same values of M. As discussed above, the chosen constraints on the numbers of wavelengths represent reasonable values given the current status of WDM technology and the bandwidth of EDFA’s w4,3,40x. Choosing N lower or greater than the optimal value would bring about a lower supportable number of nodes, because of either wavelength unavailability or detection impairments due to excessive ASE or receiver noise. Note that K is also the number of amplifiers in each fiber: small values of K lead to more nodes,

hence larger insertion losses, between adjacent amplifiers, thus exacerbating the noise problem Žpower losses are recovered less often.. The fact that the maximum number of nodes with 16 wavelengths is not twice as much as with 8 wavelengths stems from the fact that the total power of the optical signals in the amplifiers must be limited. This means that with 16 wavelengths each channel can use less power, and less signal to noise margin is available. For instance, with 16 wavelengths Ždashed curve with black markers., Daisy can host 176 nodes Žusing N s 11 and K s 16. at 1 Gbitrs, or 91 nodes at 5 Gbitrs Žusing N s 7 and K s 13.. With K F 8, less nodes can be accommodated in the network, and the difference with K F 16 becomes small for higher bit rates: while at 1 Gbitrs Daisy can accommodate 176 nodes with Ž N, K . s Ž11,16., or 112 nodes with Ž N, K . s Ž14,8., at 10 Gbitrs a maximum of 64 nodes is possible with K s 8 and N s 8. When the Ždashed. curves for Daisy are compared with the Žsolid. curves for the simplified multiring of Fig. 6, it can be noted that Daisy with no more than 8 wavelengths can support about the same number of users as the multiring with no more than 16 wavelengths Žthe two curves practically overlap.. The fact that in the multiring network of Fig. 6 the light needs to go through M transmission couplers that are not present on the outer fibers of Daisy gives a considerable scalability edge, particularly below 5 Gbitrs per channel. As an example, at 1 Gbitrs, the scalability of Daisy is over 50% greater than that of the multiring. Clearly, acomplexity argument can be raised here, and if maximum scalability is not an absolute premium, then the simpler multiring structure could be preferred, also in view of the fact that it only requires N fibers and M amplifiers as opposed to the 2 N fibers and 2 M amplifiers necessary in Daisy. The fact that the two networks are completely identical from a logical and protocol viewpoint, allows this choice to be made only based on the users’ constraints in terms of either scalability or complexity. 5. The SRR multiple access protocol

Fig. 7. Maximum number of nodes in Daisy Ždashed lines. and in the multiring Žsolid lines. vs. the individual node rate, for K F8 Žwhite markers. and K F16 Žblack markers..

In Daisy, since node n d receives data only from fiber b < d < N at wavelength l? d r N @ , no receiver con-


tention occurs at the destination node. Instead, since several source nodes may wish to transmit to one destination in an arbitrary slot, collisions between packets directed to the same destination node must be avoided both at the source node, and at the intermediate node where the optical signal is switched from fiber a < d < N to fiber b < d < N . A collision at the source node n s occurs if node n s transmits packet p 2 directed to node n d in the same slot in which another packet Ž p 1 ., also directed to node n d , is passing at node n s on fiber a < d < N . A collision at the intermediate node n x occurs if packet p 2 , switched at node n x from fiber a < d < N to fiber b < d < N , interferes with packet p 1 arriving at node n x on fiber b < d < N in the same slot and at the same wavelength, because both packets are directed to the same destination node. In either collision case, packet p 1 entered the network before packet p 2 , i.e., packet p 1 was sent by a source node upstream with respect to the source of packet p 2 . It is interesting to note that, when packet p 2 is ready for transmission, packet p 1 is passing at the source node of packet p 2 on either fiber a < d < N , or fiber b < d < N . Therefore, in order to avoid packet collision, the source node, prior to transmitting a packet whose destination is node n d , senses the activity on bi-channel wŽ a < d < N , l? d r N @ ., Ž b < d < N , l? d r N @ .x as shown in Figs. 4–6. If both channels are free, then the new packet is transmitted, and it will reach its destination without collision. Otherwise, the new packet transmission is deferred, and the in-transit packet is given precedence. It can be observed that, due to ring symmetries, each node has a better-than-average access to the channels leading to some destinations, and a worsethan-average access to other channels, leading to other destinations. If we assume node numbers to be increasing in the transmission direction, the channel on which node n i , i s 0,1, . . . , M y 1, has the best access chance is the one leading to node n < iy1 < M , since to access this channel, node n i needs not defer to transmissions of any other node; the channel on which node n i has the worst access chance is the one leading to node n < iq1 < M , since, to access this channel, node n i must defer to the possible transmissions of M y 2 upstream nodes. We assume nodes not to transmit to themselves. We say that the traffic directed to node n j from

1075

node n i has lower access priority than traffic directed to n j from nodes n < jq1 < M ,n < jq2 < M , . . . ,n < iy2 < M , n < iy1 < M . In particular, we say that node n i has access priority < i y j < M when transmitting to node n j , 1 being the highest access priority, and M y 1 the lowest. When a packet is ready for transmission from node n i to node n j , the access must be delayed if the channel leading to n j is already occupied with a transmission by one of the nodes with higher access priority with respect to destination n j . Many different collision-free MAC protocols can be designed, exploiting the channel inspection capability of Daisy. With the node architectures shown in Figs. 4–6, only the information resulting from the inspection of the bi-channel on which the packet transmission must be attempted is brought into the electronic portion of the node. This is because such node architectures were designed for protocols, which we call a priori protocols, where the packet is selected independently of the information obtained by the channel inspection. With more complex node architectures, which allow the information resulting from the inspection of all bi-channels to be brought to the node electronics, it is possible to implement more complex protocols where the packet to be transmitted is selected according to such information Ž a posteriori protocols.. Note however that the increased complexity of a posteriori algorithms, and the fact that the packet selection can only result from the algorithm itself, introduce much longer delays than a priori protocols; thus, the convenience of a posteriori protocols becomes questionable. Several alternative a priori protocols with increasing complexity and effectiveness were devised and evaluated with both analytical and simulative approaches w28–31x. Instead, the design of a posteriori protocols started just recently. Observe finally that the choice of the MAC protocol does not influence the attenuation along the ring, and is thus independent of the scalability results that were presented in the previous section. Here we just illustrate a rather simple and efficient a priori protocol named SRR Žsynchronous round robin., and assess its performance by means of simulation results. The goal of the SRR MAC Protocol is to force the network behavior under heavy load to approach TDMA Žtime division multiple access. on each bi-

1076


channel, while allowing a random access to slots under low load. In order to achieve this goal, nodes must keep one separate packet queue for every possible destination. Thus, M y 1 FIFO queues exist at each node. Queues are cyclically scanned by the node, looking for a packet to transmit. In an arbitrary time slot s, node n i deterministically schedules for transmission a packet from a specific queue. If we introduce the integer variable k s < s < My 1 , that cycles over the range 0,1,2, . . . , M y 2, the queue with packets directed to destination n < iqkq1 < M is selected for transmission in time slot s. If such a queue is empty, the transmission of the first packet from the longest queue is attempted at time slot s. In any case, if transmission in time slot s is not possible because it would generate a collision, in the following time slot Žtime slot s q 1. a new packet is selected for the transmission attempt according to the SRR algorithm. By so doing, the SRR MAC protocol maintains the schedule among transmissions, and its behavior under heavy load conditions is very close to TDMA. The selection of the next packet to be transmitted in SRR requires the information about the lengths of all queues, as well as a common knowledge by all nodes about the current value of s, i.e., a global Žnetwork-wide. synchronization on the slot sequence. Moreover, the ring latency must be equalized in order to be a multiple of the SRR frame duration Ži.e., a multiple of M y 1 slots.. Note also that a packet which was selected but not successfully transmitted in the current slot is sent back to its queue. 6. Performance results This section presents numerical results for a 16node Daisy topology, assuming that the buffer space within each node is unlimited. Since the network performance depends on the number of nodes, but is completely independent of the topology, having fixed M s 16, results apply to both topologies shown in Fig. 1 and, more generally, to any 16-node system devoting one logical channel to every destination Že.g., see the proposal in w26x.. Numerical results for throughput and packet delays are obtained with a custom simulation program written in Simula.

As in previous sections, packets Žor cells. are assumed to always exactly fit into one slot, and to be generated according to a Poisson process with rate l i at node n i . The total traffic in the network is L s Ýis0My 1 l i . The probability that a packet originated at node n i is directed to node n j is pi j . For obvious reasons pi i s 0. The network and traffic models considered in this section are quite simple, and their objective is to provide an indication of the performance and fairness achievable by Daisy with the SRR MAC protocol; more comprehensive results for SRR, and for other MAC protocols, considering more sophisticated traffic patterns, are presented in w28–31x. Two traffic scenarios are considered: Ø Uniform traffic: all source nodes generate the same amount of traffic Ž l i s LrM, ; i ., and destinations are equally likely Ž pi j s 1rŽ M y 1., ; i, j:i / j .; Ø Unbalanced traffic: one ‘‘server’’ node Žsay node n 0 . is present in the network, the remaining M y 1 ‘‘client’’ nodes direct half of their traffic to the server node, while the other half of the traffic is uniformly distributed among client nodes. The server generates an amount of traffic equal to half the total traffic generated by clients, and evenly distributes it to client nodes. In this case we use the arrival rates l0 s Lr3 and l i s Ž2r3. LrŽ M y 1. ; i / 0, while p 0 j s 1rŽ M y 1. ; j / 0, pi j s 1rw2Ž M y 2.x ; i, j / 0, i / j, and pi0 s 1r2 ; i / 0. Fig. 8 shows the average packet delay Ew W x ŽW is defined as the time from the packet generation until the beginning of its transmission. versus the overall network throughput in 16-node Daisy topologies with the SRR protocol under uniform traffic. Two curves are shown, one for source-destination pairs at the lowest priority Ž Ew Wi j x, with j s < i q 1 < M ., another for highest priority pairs Ž Ew Wi j x, with j s < i y 1 < M .; these curves average results for all source-destination pairs at the given access priority. Note however that, due to rotational symmetries, the average delay of all packets generated by each node is independent of the node index. These results clearly show that the protocol cannot completely recover the intrinsic unfairness of the topology, in spite of its TDMA behavior at high loads.


1077

Fig. 8. Average packet delays versus throughput for the SRR protocol under uniform traffic.

Fig. 9. Average packet delay for all source-destination pairs, with the SRR protocol, in the server scenario with infinite buffers, for a total traffic L s 2.7.

1078


It must be observed that, despite the delay unfairness shown by the above results, the throughput in uniform traffic is almost optimal Ži.e. very close to 1 with our normalizations. w28x. Fig. 9 shows results for the average packet delays for all source-destination pairs, in a 16-node Daisy topology under unbalanced traffic with the SRR protocol, with node 0 acting as a server and nodes 1 to 15 acting as clients, with a total traffic in the network equal to L s 2.7, which means that the bi-channel leading to the server is highly loaded Ž0.9., while the bi-channels leading to clients have a light load equal to 0.12. The unfairness of the SRR protocol is even more evident in this case. Increasing delays for decreasing access priorities are observed on the channel leading to the server. This is again due to the intrinsic characteristics of the considered topology, and little can be done against this behavior, unless a global fairness control algorithm is introduced Žsee w29,30x for a proposal.. Delays are instead small Žof the order of a few slot times. on the other channels: this is the main effect of maintaining different packet queues at each node, thereby avoiding head-of-theline blocking. In the case of just one packet queue per node, Žsee w31x., losses would be observed on the channel leading to the server. In order to adequately judge the values of the delays obtained with the SRR protocol, it must be considered that with the given load, the adoption of a fixed TDMA protocol in the bi-channel leading to the server would result in delays of the order of 70 slots, hence much larger than the largest delay obtained with the proposed MAC protocol.

7. Conclusions With the advent of fast tunable optical sources and the rapid success of optical amplifiers, WDM ring architectures are entering the arena of all-optical network architectures capable of providing information transfer support to future bandwidth-greedy applications. As a direct consequence, the well-known and well-established properties of classic ring networks can be exported to benefit third generation photonic networks.

This paper presented Daisy, a WDM all-optical collision-free packet network with multiring topology and attractive network scalability properties. Scalability was evaluated using a detailed analytical model that takes into account ASE noise and receiver noise, showing that, for example, 128 nodes can be all-optically interconnected in Daisy using only 16 wavelength channels, each at transmission speeds of 2.5 Gbitrs. Several different access protocols can be designed for Daisy; in this paper we presented the SRR protocol, and showed some simulation results about its performance.

Acknowledgements The authors should like to thank the anonymous referees for helpful comments that led to a significant improvement of this paper.

Appendix A. The scalability model of Daisy For the assessment of the scalability of Daisy we used the modelling approach originally described in w27x, which accounts for both optical noise and electrical noise within the receiver. In the calculations we imposed a conservative system margin of 6 dB, i.e., the required signal-to-noise ratios for both header detection Žto implement the MSS technique. and payload detection Žto receive data. are 6 dB higher than the ones that would ensure 10y9 error probability. This margin accounts for the unavoidable nonideality of the various network components. In both the Daisy topology and the simplified multiring we assume that amplification at each node takes place only on those fibers whose light goes through the node channel-drop filter. This means that, in Daisy at node n i , two optical amplifiers Ždashed boxes. are used on fibers a < i < and b < i < , as N N shown in Figs. 4 and 5, while one amplifier per node is used in the multiring ŽFig. 6.. On fibers bj g B, amplification is assumed to nominally compensate for the exact amount of optical loss between amplifiers. On fibers a j g A, the amplifier is meant to bring up the signal level as much as possible, to facilitate the coupling of light


Fig. 10. Computational model of a coupler. The light power is multiplied by the factors appearing in the boxes. Generally twoby-two devices are used to obtain this functionality, of which only three ports are used.

from the inner fiber into the outer fiber. In both cases, the per-WDM-channel power is limited by the optical amplifier saturation power. We assumed a conservative value of 10 mW for all amplifiers. Considering the worst-case of the simultaneous presence of packets on all the K wavelengths that each fiber carries, the available per-channel power after amplification is 10rK mW. All amplifiers were given a conservative noise figure of 4.5 dB. The header and payload receivers were assumed to use 50 V electronic amplifiers with an electrical noise figure of 4. The photodetectors were assumed to have a unit responsivity and an internal 50 V matching load. The node architecture in Daisy makes use of several different couplers. For all the couplers we assumed the model shown in Fig. 10. In all cases we chose a realistic loss parameter a of y0.25 dB. The value of the transmission parameter r is extremely critical because it greatly influences network scalability. Through iterative numerical analysis we found the following optimal values: r s y0.5 dB 2 , y2.7 dB, and y0.8 dB, for the header detector coupler Ž S a and Sb in Figs. 4 and 5., the inner-outer fiber i i coupler Ž C0 ., and the transmitter coupler Ž Ci , with i ) 0., respectively. The transmitter coupler value was constrained, among other factors, by the maxi-

2

To better clarify, a sy0.25 dB and r sy0.5 dB for the header detector coupler corresponds to a 1 dB attenuation for the through light, whereas about 10% of the input light is coupled into the header detector photodiode.

1079

mum available transmitter laser power, which was assumed to be 6 mW into its pigtail. Today, such an output power can be found in commercially available lasers. An even better scalability performance could have been achieved by making the r values dependent on the coupler position along the inner or outer fibers, but this option was not considered since it appeared unrealistic from the viewpoint of a practical implementation. The payload receiver filter was assumed to be implemented as shown in Fig. 11. It consists of a fiber grating acting as a selective reflector that can bounce back up to 99% of the incident light at a specific wavelength, within a bandwidth of a few tenths of a nanometer. The reflected light is gathered through a coupler ŽFig. 11., thus realizing a singlewavelength-extracting filter. This wavelength-extracting filter implementation has the advantage of being polarization insensitive and features a low attenuation on the wavelengths that are let through. This attenuation is due mostly to the light going through the coupler before reaching the grating. The best value for the r parameter of this coupler was found to be y1.3 dB. The back-reflected light that is not coupled out is blocked by the optical amplifier isolator. It should be mentioned here that some increase in scalability could be obtained by using more amplifiers Žeach having lower individual gain. spread throughout the network. However, we chose to constrain the design to only two amplifiers per node due to obvious cost and practicality issues. The placement of these two amplifiers in the node is not arbitrary; the proposed placement is the best among

Fig. 11. A possible structure for the receiver wavelength-extracting filter. A reflective fiber grating bounces back the desired wavelength. The reflected signal is then extracted with a coupler and sent to the payload data receiver.

1080


all other possible configurations. This result was verified through the analysis. Identical assumptions on basic device performance and component structure were made when analyzing the simplified multiring structure. The same iterative device placement and coupler rparameter optimizations, that were numerically performed for Daisy, were also carried out for the simplified multiring topology. References w1x P.E. Green, The future of fiber-optic computer networks, IEEE Comput. 24 Ž9. ŽSeptember 1991. 78–87. w2x C.A. Brackett, Dense wavelength division multiplexing networks: principles and applications, IEEE J. Select. Areas Comm. 8 Ž6. ŽAugust 1990. 948–964. w3x P.E. Green, Optical networkig update, IEEE J. Select. Areas Comm. 14 Ž5. ŽJune 1996. 764–779. w4x A.S. Acampora, The scalable lightwave network, IEEE Comm. Mag. 32 Ž12. ŽDecember 1994. 36–42. w5x I. Chlamtac, A. Ganz, G. Karmi, Lightnet: lightpath based solutions for wide bandwidth WANs, Infocom’90, San Francisco, CA, June 1990; also in IEEErOSA J. Lightwave Technol. 11 Ž5,6. ŽMayrJune 1993. 951–961. w6x S.B. Alexander et al., A precompetitive consortium on wideband all-optical networks, IEEErOSA J. Lightwave Technol. 11 Ž5,6. ŽMayrJune 1993. 714–735. w7x C.A. Brackett, A.S. Acampora, J. Sweitzer, G. Tangonan, M.T. Smith, W. Lennon, K.-C. Wang, R.H. Hobbs, A scalable multiwavelength multihop optical network: a proposal for research on all-optical networks, IEEErOSA J. Lightwave Technol. 11 Ž5,6. ŽMayrJune 1993. 736–753. w8x R. Ramaswami, K.N. Sivarajan, Design of logical topologies for wavelength routed optical networks, IEEE J. Select. Areas Comm. 14 Ž5. ŽJune 1996. 840–851. w9x B. Mukherjee, WDM-based local lightwave networks – part I: single-hop systems, IEEE Netw. 6 Ž3. ŽMay 1992. 12–27. w10x B. Mukherjee, WDM-based local lightwave networks – part II: multihop systems, IEEE Netw. 6 Ž5. ŽJuly 1992. 20–32. w11x M.I. Irshid, M. Kavehrad, A fully transparent fiber-optic ring architecture for WDM networks, IEEErOSA J. Lightwave Technol. 10 Ž1. ŽJanuary 1992. 101–108. w12x W.I. Way et al., Self-routing WDM high-capacity SONET ring network, OFC’92, San Jose, ´ CA, February 1992, paper TuO2. w13x A. Hamel, D. Laville et al., Multilayer add-drop multiplexers in a self-healing WDM ring network, OFC’95, San Diego, CA, February 1995, paper TuQ2. w14x M.J. Chawki, V. Tholey et al., Demonstration of a WDM survivable open ring network using reconfigurable dropping receivers, ECOC’94, Firenze, Italy, September 1994. w15x Ying Cai et al., Demonstration of photonic packet-switched ring networks with optically transparent nodes, IEEE Photon. Technol. Lett. 6 Ž9. ŽMay 1994. 1139–1141.

w16x M.J. Chawki et al., Wavelength reuse scheme in a WDM unidirectional ring network using a proper fiber grating addrdrop multiplexer, OFC’95, San Diego, CA, February 1995, paper ThI3. w17x H. Toba, K. Oda, K. Inoue, K. Nosu, T. Kitoh, An optical FDM-based self-healing ring network employing arrayed waveguide grating filters and EDFA’s with level equalizers, IEEE J. Select. Areas Comm. 14 Ž5. ŽJune 1996. 800–813. w18x I. Chlamtac et al., CORD: contention resolution by delay lines, IEEE J. Select. Areas Comm. 14 Ž5. ŽJune 1996. 1014–1029. w19x K. Bala, C. Brackett, Cycles in wavelength routed optical networks, IEEE Summer Topical Meeting in Optical Networks and Their Enabling Technologies, Lake Tahoe, CA, July 1994. w20x N.K. Shankaranarayanan, U. Koren, B. Glance, G. Wright, Two-section DBR laser transmitters with accurate channel spacing and fast arbitrary-sequence tuning for optical FDMA networks, OFC’94, San Jose, ´ CA, February 1994, paper TuI2, pp. 36–37. w21x W.H. Nelson et al., Large-angle 1.3 mm InPrInGaAsP digital optical switches with extinction ratios exceeding 20 dB, OFC’94, San Jose, ´ CA, February 1994, paper TuM2, pp. 53–54. w22x S.-W. Seo, K. Bergman, P.R. Prucnal, Transparent optical networks with time division multiplexing, IEEE J. Select. Areas Comm. 14 Ž5. ŽJune 1996. 1039–1051. w23x C.-K. Chan, L.-K. Chen, K.-W. Cheung, A fast channel-tunable optical transmitter for ultrahigh-speed all-optical time-division multiaccess networks, IEEE J. Select. Areas Comm. 14 Ž5. ŽJune 1996. 1052–1056. w24x A. Budman et al., Multigigabit optical packet switch for self-routing networks with subcarrier addressing, OFC’92, San Jose, ´ CA, February 1992, paper TuO4. w25x S.-F. Su, R. Olshansky, Performance of multiple access WDM networks with subcarrier control channels, Summer Topical Meetings, Santa Barbara, CA, August 1992. w26x I. Chlamtac, A. Fumagalli, L.G. Kazovsky, P. Poggiolini, A multi-Gbitrs WDM optical packet network with physical ring topology and multi-subcarrier header encoding, ECOC’93, Montreux, Switzerland, September 1993. w27x P. Poggiolini, S. Benedetto, Theory of subcarrier encoding of packet headers in quasi-all-optical broadband WDM networks, IEEErOSA J. Lightwave Technol. 12 Ž10. ŽOctober 1994. 1869–1881. w28x M. Ajmone Marsan, A. Bianco, E. Leonardi, M. Meo, F. Neri, On the capacity of MAC protocols for all-optical WDM multi-rings, Infocom’96, San Francisco, CA, March 1996. w29x M. Ajmone Marsan, A. Bianco, E. Leonardi, M. Meo, F. Neri, MAC protocols and fairness control in WDM multirings with tunable transmitters and fixed receivers, IEEErOSA J. Lightwave Technol. 14 Ž6. ŽJune 1996. 1230– 1244. w30x M. Ajmone Marsan, A. Bianco, E. Leonardi, A. Morabito, F. Neri, SR3: a bandwidth-reservation MAC protocol for multimedia applications over all-optical WDM multi-rings, Infocom’97, Kobe, Japan, April 1997. w31x M. Ajmone Marsan, A. Fumagalli, E. Leonardi, F. Neri,


w32x

w33x w34x

w35x w36x

w37x

w38x

w39x

w40x

w41x

w42x

Modeling slotted multi-channel ring all-optical networks, Int. Symp. on Modeling, Analysis and Simulation of Computer and Telecommunication Systems ŽMASCOTS’97., Haifa, Israel, January 1997. E. Hafner, Z. Nenadal, Enhancing the availability of a loop system by meshing, 1976 Int. Zurich Seminar on Digital Communications, Zurich, Switzerland, March 1976. A. Grnanov, L. Kleinrock, M. Gerla, A highly reliable loop network architecture, FTCS-10, Kyoto, Japan, October 1980. C.S. Raghavendra, M. Gerla, A. Avizienis, Reliable loop topologies for large local computer networks, IEEE Trans. Comput. C-34 Ž1. ŽJanuary 1985. 46–55. B.W. Arden, H. Lee, Analysis of chordal ring network, IEEE Trans. Comput. C-30 Ž4. ŽApril 1981. 291–295. K.W. Doty, New designs for dense processor interconnection networks, IEEE Trans. Comput. C-33 Ž5. ŽMay 1984. 447– 455. Y. Pan, On efficient distributed elections in clustered chordal rings, Proc. 1992 Int. Conf. on Parallel Processing, II, Software ŽAugust 1992. II-141–II-144. C.S. Raghavendra, J.A. Silvester, A survey of multi-connected loop topologies for local computer networks, Comput. Netw. ISDN Syst. 11 Ž1986. 29–42. D. Guo, W. Guo, A.S. Acampora, Scalable multihop WDM passive ring with optimal wavelength assignment and adaptive wavelength routing, Globecom’95, Singapore, November 1995, pp. 2195–2204. M.A. Ali et al., Performance of erbium-doped fiber amplifier cascades in WDM multiple access lightwave networks, IEEE Photon. Technol. Lett. 6 Ž9. ŽSeptember 1994. 1142–1145. E.L. Goldstein, A.F. Elrefaie, N. Jackman, S. Zaidi, Multiwavelength fiber-amplifier cascades in unidirectional interoffice ring networks, OFC’93, San Jose, ´ CA, February 1993, paper TuJ3. D. Bayart et al., Experimental investigation of the gain flatness characteristics for 1.55 mm erbium-doped fluoride fiber amplifiers, IEEE Photon. Technol. Lett. 6 Ž5. ŽMay 1994. 613–615.

1081

Marco Ajmone Marsan is a Full Professor at the Electronics Department of Politecnico di Torino, in Italy. He was born in Torino, Italy, in 1951. He holds a Dr. Ing. degree in Electronic Engineering from Politecnico di Torino, and a Master of Science from the University of California, Los Angeles. Since november 1975 to october 1987 he was at the Electronics Department of Politecnico di Torino, first as a Researcher, then as an Associate Professor. Since November 1987 to October 1990 he was a Full Professor at the Computer Science Department of the University of Milan, in Italy. During the summers of 1980 and 1981 he was with the Research in Distributed Processing Group, Computer Science Department, UCLA. He has coauthored about 200 journal and conference papers in the areas of Communications and Computer Science, as well as the two books Performance Models of Multiprocessor Systems published by the MIT Press, and Modelling with Generalized Stochastic Petri Nets published by John Wiley. He received the best paper award at the Third International Conference on Distributed Computing Systems in Miami, Fla., in 1982. His current interests are in the fields of performance evaluation of communication networks and their protocols. M. Ajmone Marsan is a Senior Member of IEEE. Andrea Fumagalli received the Dr. Ing. degree in electronic engineering and the Ph.D. degree in electronic engineering from the Politecnico di Torino, Italy, respectively in 1987 and 1992. He did postdoctoral research in the field of all-optical networks at the University of Massachusetts at Amherst. Since 1992 he has been with the Department of Electronics at the Politecnico di Torino as a Researcher. Currently he is a Visiting Associate Professor at the Erik Jonsson School of Engineering and Computer Science at the University of Texas at Dallas. Dr. Fumagalli’s research interests include aspects of mobile and optical networks, related protocol design and performance evaluation.

1082


Emilio Leonardi received a Dr. Ing. degree in Electrical Engineering in 1991, and a Ph.D. in Telecommunications in 1996, both from Politecnico di Torino, Turin, Italy. He is currently with the Networking Research Group of Politecnico di Torino, where he holds a PostDoctoral fellowship. From September 1994 to July 1995 he was at the Computer Science Department of University of California, Los Angeles, ŽUCLA., working on Deadlock-Free Routing and Flow Control Techniques for Asynchronous High-speed Wormhole Networks. His current research interests are in the fields of performance evaluation of high-speed networks, all-optical networks, wireless access protocols, queueing theory. Fabio Neri is an Associate Professor at the Electronics Department of Politecnico di Torino, Turin, Italy. He was born in Novara, Italy, in 1958. He received his Dr. Ing., Ph.D. degree in Electrical Engineering from Politecnico di Torino in 1981 and 1987, respectively. From 1991 to 1992 he was with the Information Engineering Department at University of Parma, Parma, Italy, as an Associate Professor. From 1982 to 1983 he was a visiting scholar at George Washington University in Washington, DC. In the summer of 1995 he was visiting researcher at the Computer Science Department of the University of California in Los Angeles ŽUCLA.. His research interests are in the fields of performance evaluation of communication networks, high-speed and all-optical networks, discrete event simulation, and queuing theory. He has co-authored over 50 papers published in international journals and presented in leading international conferences.

Pierluigi Poggiolini was born in 1963 in Torino, Italy. He received the MS degree Žsumma cum laude. in 1988 and the Ph.D. degree in 1993, both from Politecnico di Torino. From 1988 to 1989 he was with the Italian State Telephone Company research center CSELT, working on performance analysis and computer simulation of lightwave transmission systems and optoelectronic devices. He also worked in the field of polarization scrambling-spreading techniques, where he holds an international patent. From 1990 to 1992 he was with the Optical Communications Research Laboratory at Stanford University where he worked on polarization modulation and was involved in the STARNET broadband optical network project. From 1993 to 1995 he was a postdoctoral fellow at Stanford University, where he worked on the ARPA-sponsored optical broadband packet network project CORD. He is currently a research assistant at Politecnico di Torino. His research interests include the theoretical analysis and experimental implementation of polarization modulated optical systems, the study and simulation of long-haul optical transmission systems and the design and performance evaluation of broadband packet-switched optical networks.