SlideShare uma empresa Scribd logo
1 de 20
JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 18, NO. 12, DECEMBER 2000                                                                            1657




       WRAP: A Medium Access Control Protocol for
        Wavelength-Routed Passive Optical Networks
                         Michael J. Spencer, Member, IEEE, and Mark A. Summerfield, Member, IEEE


    Abstract—We describe the WDM Request/Allocation Protocol                  of the medium access control (MAC) protocols employed [34].
(WRAP), a media-access control protocol for wavelength-routed                 In broadcast networks, simplifying the node hardware by using
passive optical networks (WR-PONs) in which each node has a                   fixed transmitters or receivers limits the number of nodes to the
single fixed optical receiver and a single tunable optical trans-
mitter. The protocol does not require a carrier sensing capability, a         number of available wavelengths [1]. Furthermore, existing and
separate control channel, or any centralized control or scheduling.           emerging tunable receiver technologies do not provide the wide
Access to transmission channels is regulated by allocations made              tuning range and rapid tunability required to implement flexible
at destination nodes in response to requests made by source                   and efficient packet networks. Attempts have been made to
nodes. Computer simulation is used to investigate three different             work around tuning speed limitations using MAC protocols
allocation algorithms, one of which—the Preferential/Random
algorithm—is shown to provide significantly better performance                which aggregate packets for transmission, and use scheduling
than the alternatives. Simulations are presented comparing the                and “pipelining” techniques to mask the effects of slow-tuning
performance of WRAP to two previously proposed applicable pro-                components [12]–[14]. Such protocols mitigate the effects
tocols—the Interleaved Time Division Multiple Access (I-TDMA)                 of tuning time under conditions of heavy network loading,
protocol, and the FatMAC protocol. WRAP is shown to provide                   however they cannot overcome the fundamental limitations of
fair and flexible access to the transmission capacity, enabling high
network utilization to be achieved under a wide range of traffic              slow-tuning devices, and are characterized by high packet delay
conditions, while providing a guaranteed minimum bandwidth                    even under lightly-loaded conditions, and high delay variance
between each source–destination pair. We conclude that of the                 under heavy loading. Finally, many of the proposed protocols
three protocols considered here, WRAP is the best-suited to                   for WDM LANs require the use of an additional control
general-purpose data communications applications such as local,               channel and/or the provision of a “carrier-sense” capability,
campus, and metropolitan area networks.
                                                                              resulting in further cost and complexity in the implementation
  Index Terms—Media-access control (MAC) protocols, passive                   of the network nodes [1], [5], [15], [31]–[33].
optical networks, wavelength division multiplexing, wavelength
routing.                                                                         Recent advances in optical component technologies, and
                                                                              in particular the development and commercialization of
                                                                              wavelength-selective components such as fiber-Bragg grat-
                          I. INTRODUCTION
                                                                              ings (FBGs) and arrayed waveguide gratings (AWGs), make

I   N RECENT years, there has been a great deal of interest in
    the application of wavelength division multiplexing (WDM)
to local area networks (LANs) [1]–[5]. The principal advantage
                                                                              wavelength-routed PONs (WR-PONs) a practical alternative
                                                                              to broadcast topologies. Wavelength-routed star networks can
                                                                              be constructed by using an AWG in place of a wavelength-in-
of WDM is that it makes available the enormous bandwidth of                   dependent coupler—a technique which has been proposed
optical fiber while allowing the individual wavelength channels               for low-cost optical access networks [16]. Wavelength-routed
to operate at bit-rates attainable using affordable electronic                ring and bus networks can be built using WDM add-drop
components [6]. Most proposed WDM LANs have been based                        multiplexers (W-ADMs) in place of wavelength-independent
upon broadcast passive optical network (PON) architectures                    splitters [5], [17]. In a wavelength-routed network, the trans-
employing either a star topology based on a central passive                   mitter wavelength acts as a form of addressing, determining
coupler [7]–[9], or a ring topology with a simple optical                     the destination of a transmission over a single optical hop.
splitter at each node [10]. These broadcast PON architectures                 Potential advantages of this approach include improved power
are conceptually simple, and can be constructed from optical                  budget and superior security (since signals are not needlessly
components which have been commercially available for over                    and indiscriminately broadcast to nodes other than the intended
a decade. However, the penalty for this simplicity in the optical             destination) and the ability to support more nodes than there
components is generally either: an increase in the complexity of              are wavelengths available through the use of multihop paths
the network nodes, requiring, for example, tunable transmitters               [18]. Furthermore, recent advances in multiwavelength source
and tunable receivers at every node [1], [11]; and/or relatively              technology indicate that rapidly-tunable sources with wide
low channel utilization and throughput due to the inefficiency                tuning ranges will become a commercial reality in the foresee-
                                                                              able future [19]–[22], [35], [36]. Grating assisted codirectional
  Manuscript received April 3, 2000; revised September 26, 2000.              Coupler with rear Sampled grating Reflector (GCSR) tunable
  The authors are with Australian Photonics Cooperative Research Centre,      lasers have achieved tuning times of less than 5 ns between 40
Photonics Research Laboratory, Department of Electrical and Electronic        ITU channels [35]. Sampled grating distributed Bragg reflector
Engineering, The University of Melbourne, Victoria 3010, Australia (e-mail:
m.spencer@ee.mu.oz.au; m.summerfield@ee.mu.oz.au).                            tunable lasers with a tuning time of approximately 5 ns between
  Publisher Item Identifier S 0733-8724(00)10886-2.                           50 channels [36] and multifrequency lasers with expected
                                                           0733–8724/00$10.00 © 2000 IEEE
1658                                                                              JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 18, NO. 12, DECEMBER 2000




Fig. 1. Generic wavelength-routed passive optical network (WR-PON) architecture. Each node has a single tunable transmitter (TT), and a single fixed receiver
(FR). All filtering and optical routing is performed by passive optical elements within the network.


tuning times of less than 10 ns between any of 40 channels [22]                 presented previously in [27]. The present paper extends this pre-
have also been reported.                                                        vious work, including a more comprehensive description of the
   Therefore, WDM LANs based on WR-PONs and requiring                           design and performance issues, more detailed results and com-
only tunable transmitters represent a very promising approach to                parisons with previously proposed applicable protocols.
meeting future high-capacity data network requirements. To re-                     The structure of this paper is as follows. In Section II, we
duce the complexity and cost of the node hardware further, MAC                  provide an overview of the operation of WRAP, and describe
protocols which do not require any additional control channel or                the WDM network architectures to which the protocol is appli-
carrier-sense capability are highly desirable.                                  cable. Simulations of the protocol demonstrating the impact of
   In this paper, we consider the design and performance of a                   different allocation algorithms are described in Section III. In
MAC protocol, called simply the WDM Request/Allocation                          Section IV we present simulation results demonstrating the per-
Protocol (WRAP), applicable to WR-PONs with fast-tuning                         formance of WRAP as a function of the network and protocol
transmitters and fixed receivers. Our protocol does not re-                     parameters, and compare it with alternative applicable proto-
quire an additional control channel, performing all necessary                   cols.
signaling in-band. Unlike the previously-proposed FatMAC
protocol [24], which also uses in-band signaling, WRAP                          II. OVERVIEW OF THE PROTOCOL AND NETWORK TOPOLOGIES
does not require an optical broadcast facility, and is therefore                   Fig. 1 illustrates schematically the “generic” wave-
applicable to systems using DBR or GCSR tunable semicon-                        length-routed PON considered in this paper. Each node has
ductor lasers which can emit only one wavelength at a time.                     a single tunable transmitter (TT) and a single fixed optical
Furthermore, WRAP delivers similar performance to the ideal                     receiver (FR). Note that no optical filtering is employed at
interleaved time-division multiplexed (I-TDMA) scheme [25]                      the receivers—all necessary “filtering” is inherently provided
under heavy, uniform loading, while offering superior perfor-                   by the wavelength-routing functionality of the PON. Light
mance and capacity sharing under light and/or uneven loads.                     entering the PON at an input port is routed by passive optical
WRAP is robust in the presence of significant propagation                       components within the network to one and only one output
delays in transmission, and in many situations shows improved                   port. The particular output port selected may depend only
performance in the presence of high delay-bandwidth product                     upon the input wavelength, or it may depend upon both the
as a result of inherent pipelining within the network.                          wavelength and the input port. In either case, a transmitting
   The initial motivation for the development of WRAP was to                    node has the ability to select a specific destination node by
provide a MAC protocol for use in the MAWSON WDM net-                           tuning its transmitter to the corresponding wavelength. The
work demonstrator currently being developed within the Aus-                     network has no “master” or “controller” nodes, and allocation
tralian Photonics Cooperative Research Centre at the University                 of network resources is distributed amongst the nodes. The
of Melbourne [26]. MAWSON is a ring-based network, however                      lack of centralized control reduces the system complexity
as we discuss in Section II, WRAP is also applicable to other                   and eliminates the possibility of there being a single critical
wavelength-routed topologies such as the AWG-based star. This                   congestion point, thus making the network simpler to install
is in contrast to most previously proposed protocols for WDM                    and maintain.
LANs which are applicable only to star topologies. The basic                       The MAC protocol described in this paper operates in a
design of WRAP, and initial performance simulations have been                   slotted mode to maximize efficiency. Each slot is divided into
SPENCER AND SUMMERFIELD: MAC PROTOCOL FOR WAVELENGTH-ROUTED PONS                                                                     1659




Fig. 2.   WR-PON queuing model.


a number of shorter transmission periods known as minislots,        followed immediately by the transmission of the corresponding
some of which are used for signaling, and others for data           data (e.g., [29]); and those in which the reservations must be
transmission, as described below. The protocol is applicable to     received by all nodes before implicit slot allocations are made
a range of wavelength-routed optical network topologies, the        based on a common algorithm (e.g., [11], [24]). WRAP is
only requirement being that a mechanism for synchronizing           based on a three-stage scheme in which a node wishing to send
transmissions at the slot boundaries must be provided. The          data first requests an allocation from the destination node. The
method of synchronization, and the effective point of synchro-      destination receives these requests from all source nodes, and
nization, depends upon the details of the physical topology. For    then sends back specific allocations to each source (hence the
some topologies this may have significant implementation and        name WDM Request/Allocation Protocol). Finally, the source
performance implications, as we show later when discussing          nodes send their data using the allocated slots.
the application of WRAP in a ring topology. In this section we         The WRAP scheme has a number of advantages over previ-
first present an overview of the operation of WRAP, followed        ously-proposed protocols:
by further details of its implementation in star and ring topolo-       • It does not require reservation requests from all source
gies. This paper is concerned only with the performance of the             nodes to be distributed globally within the network, nor
protocol in the simplest case of a single-hop topology in which            does it require a centralized scheduler to handle requests.
the number of nodes is equal to the number of wavelengths                  This means that it does not require a shared control
   , i.e.,         . Other, more complex, implementations are              channel, or a broadcast facility.
possible, including multihop topologies, and variations on the          • It is inherently fair, under any well-defined fairness crite-
protocol in which wavelengths may be shared (i.e.,             );          rion, since access to any specified destination can be reg-
however, these are beyond the scope of the present paper.                  ulated completely by the allocation algorithm used by the
                                                                           destination node.
A. Queuing Model and Protocol Description                               • With a suitable choice of allocation algorithm, perfor-
   In the absence of any carrier-sense capability there are two            mance can be made equivalent to that of I-TDMA under
basic classes of MAC protocol: those that allow collisions such            conditions of uniform heavy loading.
as ALOHA and slotted ALOHA [28]; and those that avoid all              The queuing model of the network is shown in Fig. 2. Each
collisions by the careful scheduling of transmissions. Protocols    node maintains a set of           queues, one for each of the dis-
which allow collisions are inherently inefficient and suffer        tinct destinations. The tunable transmitter at node can access
from poor throughput as a result of the high levels of data loss    the set of wavelengths           which, in the present work, is as-
which occur under moderate to heavy loading [28]. A simple          sumed to be sufficient to reach all other nodes in the network.
scheme based on the static preallocation of all data transmission   The transmitter can only tune to one wavelength in               at a
slots such as the I-TDMA protocol [25] provides a fair and          time. The receiver at node receives signals on the set of wave-
optimal use of network resources for uniform network loading        lengths           but makes no distinction between these wave-
and does not suffer any data collisions. Thus it is desirable       lengths. If two wavelengths in           are active simultaneously,
as the uniformly loaded limit for any collision-free protocol.      or the same wavelength is active via two sources simultaneously,
Dynamic sharing of transmission resources requires the use of       a collision occurs at node . The objective of the WRAP protocol
some kind of reservation scheme, usually based on a separate        is to avoid all such collisions.
control channel for exchanging the reservation information.            To implement WRAP, time is divided into a series of uni-
Most previously proposed reservation-based protocols fall into      form length slots, of duration        . The slot structure is shown
one of two basic categories: those in which the reservations are    in Fig. 3(a). The nodes are synchronized so that the start of a
1660                                                                                JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 18, NO. 12, DECEMBER 2000




             N                                                       N0                                             M
Fig. 3. WDM Request/Allocation Protocol (WRAP) slot structure and WDM transmission schedule. (a) Slot structure as viewed from a single transmitter or
                                                                           1 Request/Allocation (R/A) minislots, and data minislots. Transmitted minislots are
                         t                       t
receiver in an -node network consists of a synchronization minislot,
separated by a guard time , and the slot period is    . (b) Sample multichannel transmission schedule in a three-node network. Each row represents transmissions
to a single destination node; the numbers in the R/A minslots represent the source node; allocation of data minislots to source nodes is determined by the WRAP
protocol.


slot occurs at the same time on all the wavelengths in the net-                  resynchronization. They may also be the result of small vari-
work at the synchronization point. As described later, the syn-                  ations in propagation delays caused by the effects of dispersion
chronization points are at the network hub for a star topology                   on different wavelengths and by temperature changes and other
and at the WDM ADMs for a ring topology. As mentioned pre-                       environmental fluctuations in the fibers. The framing overhead
viously, WRAP is based on a three-stage bandwidth allocation                     enables nodes to perform clock and data recovery on each in-
scheme. In the first stage source nodes with data to transmit send               coming minislot. This is needed for correct reception of the data
requests to the destination nodes. In the second stage the desti-                contained within the minislot. The framing is also needed to
nation nodes independently make allocations of data minislots                    byte-align and delineate the data within each minislot.
and send them to the source nodes. In the final stage the source                    The synchronization minislot is followed by                re-
nodes transmit the data to the destinations nodes in the allocated               quest/allocation (R/A) minislots and        data minislots. The
data minislots. In the simplest and most efficient implementa-                   R/A minislots are preallocated on each wavelength in an
tion, these three procedures should occur in consecutive slots. It               interleaved TDMA manner in such a way that:
is thus apparent that if the slot period cannot be too short oth-
erwise, for example, requests will not arrive in time for alloca-                    • each source node is allocated one R/A minislot to each
tions to be sent out in the next slot. This implies that, regardless                   destination node in each slot; and
of the topology, there is a minimum slot length. As we shall see,                    • no source node is required to transmit R/A minislots to
the ring topology results in even more severe constraints on the                       two different destinations at the same time.
allowable slot length. The constraints on the slot length are dis-                  This is illustrated for the simple case of a three-node network
cussed in more detail in the following sections.                                 in Fig. 3(b). The length of the data minislots is a selectable pa-
   As shown in Fig. 3(a), each slot is further divided into a se-                rameter of the protocol, and is denoted by           (bits).
ries of minislots, whose lengths depend on their functions. The                     Once per slot, a node examines the state of its queues, and de-
synchronization minislot is used to maintain slot synchroniza-                   cides how many data minislots to request, up to the maximum
tion between nodes. The length of the synchronization minislot                   of per destination. Thus, a request requires at most
is       bits. A practical implementation also requires a guard                  bits. Each node also examines the requests received from all
time, , between minislots and an additional framing overhead                     other nodes during the previous slot, and makes specific, col-
of bits per minislot. The guard time accommodates the laser                      lision-free allocations of data minislots which are effective for
tuning time as well as any timing variations. These timing vari-                 the next slot. A number of different allocation algorithms are
ations may be caused by a drift in the node clocks between                       possible, as discussed in Section III. Allocations therefore re-
SPENCER AND SUMMERFIELD: MAC PROTOCOL FOR WAVELENGTH-ROUTED PONS                                                                                1661



quire    bits (one per data minislot). The minimum number of
bits     required in each R/A minislot (excluding overheads) is

                                                                 (1)

   Finally, each node examines the allocations it received during
the previous slot (based on requests it made in the slot prior to
that), and transmits data from each of its queues during the data
minislot periods allocated to it by the corresponding destination
nodes.
   Note that although there is a minimum three-slot delay be-
tween the arrival of data in a queue, and its delivery to the des-
tination, under heavier loading (when there is usually data in
the queues), the protocol is expected to remain efficient due
to the effects of pipelining, in which requests, allocations and
data transmissions take place continually over long sequences of
contiguous slots. Furthermore, as we will demonstrate, in prac-
tical implementations the slot length is often comparable to the
unavoidable propagation delays within the network, implying
that the built-in protocol delay is of the same order as the prop-
agation time.
   For an -node, -wavelength network, the total theoretical             Fig. 4. Four-node wavelength-routed ring network. WDM add–drop
data transmission capacity of a slot in bits is given by                multiplexers (ADMs) drop one wavelength out of the ring at each access point,
                                                                        while allowing any wavelength to be added.


                                                                        the length of a slot is constrained to be a submultiple of the
                                                                 (2)    ring round trip propagation delay,           , to ensure simple node
                                                                        synchronization, high utilization and collision free operation
The first expression gives an upper bound on the transmission           [27]. As described in the previous section, it is necessary for
capacity obtained by subtracting the protocol and network over-         the nodes to be able to send requests, return allocations and
heads for the total slot capacity, while the second expression is       transmit data in consecutive slots. This is possible so long as
the actual capacity achieved given a particular choice of data          1) the slot period is not less than the ring round trip propagation
minislot length        . Differences between these two quantities       delay, and 2) the largest distance from a node to an access point
may arise in topologies in which            is constrained, resulting   is less than half the distance around the ring. Thus the only
in a period of “dead time” at the end of each slot. Equation (2)        viable choice for a slot length is        , i.e.
describes the fixed relationships between slot length, number
                                                                                                                                                 (3)
of nodes, and number of data minislots. There are tradeoffs
amongst these parameters, which are investigated in more detail         Each node must transmit its minislots at such a time that they
through simulations in Section IV. Briefly, protocol overheads          arrive at the ADM and are inserted into the ring at exactly
are more significant for small slots and/or large and , while           the right time (to within the accuracy allowed by the guard
the choice of the number of data minislots is a tradeoff between        times) to avoid collisions with adjacent minislots on the same
the impact of protocol overheads for large            and the wasted    wavelength. This requires that each node be aware of all the
capacity of small      if the traffic patterns lead to many partially   propagation delays within the network, and maintain an internal
filled data minislots. Furthermore, as we discuss later, the imple-     timer to control transmissions. Global synchronization of the
mentation of an efficient allocation algorithm may place further        node timers can be maintained using the synchronization min-
constraints on .                                                        islot provided at the beginning of each slot. We have developed
                                                                        detailed cold- and warm-start procedures for establishing and
B. Ring Topology                                                        maintaining synchronization which have been verified by sim-
   The basic ring topology is illustrated for the case of a             ulation, and which will be implemented within the MAWSON
four-node network in Fig. 4. It consists of a fiber optic ring with     demonstrator. A full description of these procedures is beyond
a WDM add/drop multiplexer (ADM) located at each access                 the scope of the present paper.
point. In the simplest configuration each ADM drops a single,              The constraint [see (3)] on the slot length in the ring topology
fixed wavelength that is different from those dropped at other          implies that the maximum available slot capacity is dependent
access points. The ADM also allows any one wavelength, or               upon the ring length, which is not a dynamically-selectable pa-
any number of wavelengths simultaneously, to be added. This is          rameter. However, the approximate length can be selected when
the architecture of the MAWSON network demonstration being              the network is installed, and the ring length may be increased
developed by the Australian Photonics Cooperative Research              by the addition of spooled fiber if the geographical size is found
Centre at the University of Melbourne. In the ring topology,            to be too small. Equation (2) shows that the maximum capacity
1662                                                                               JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 18, NO. 12, DECEMBER 2000




Fig. 5. Four-node wavelength routed star network based on an arrayed waveguide grating (AWG) hub router. Signals are routed uniquely from source to destination
based on their wavelength, and the port at which they are input to the AWG.


per slot increases linearly with the slot length, and decreases                  within the accuracy allowed by the guard times) to avoid colli-
approximately linearly with the number of nodes and/or data                      sions with adjacent minislots on different wavelengths, destined
minislots. For practical values of the network and protocol over-                for the same output port. In this way, transmissions on all wave-
heads, a rule-of-thumb which can be derived from (2) is that at                  lengths are fully synchronized at the hub, as shown in Fig. 3(b).
1 Gb/s data rates, a maximum capacity of greater than 90% (i.e.,                 This requires that each node be aware of the propagation delay
900 Mb/s per wavelength) can be maintained if the length of the                  between itself and the hub within the network. As in the case
ring is greater than            m (i.e., 1 km per node). Note that               of the ring topology, global synchronization of the node timers
the maximum capacity tends toward 100% as the physical net-                      can be maintained using the synchronization minislot provided
work size or the bit-rate increases, the key parameter being the                 at the beginning of each slot. A technique for determining the
data-storage capacity          of the ring itself.                               node-hub propagation delay, and achieving global synchroniza-
                                                                                 tion, known as Distributed Slot Synchronization (DSS) has been
C. Star Topology                                                                 developed for use in the CORD network testbed at Stanford,
                                                                                 which is based on a broadcast star [4]. This technique may be
   The basic star topology is illustrated for a four-node network                readily adapted for use in the wavelength-routed star topology,
in Fig. 5. It consists of a passive optical wavelength-routing hub               although a full description of the scheme is beyond the scope of
(which may be implemented using a single AWG, or equiva-                         the present paper.
lent structure) to which the nodes are connected by fiber pairs.                    Once again it is necessary for the nodes to be able to send re-
Light entering the hub on any given wavelength and input fiber                   quests, return allocations and transmit data in consecutive slots.
is routed to a specific corresponding output port. The routing                   For the star topology this implies that the slot length must be
function is such that light on the same wavelength but arriving                  more than twice the longest propagation delay between the hub
from different input ports must always be routed to different                    and a node. Other than this requirement,           is a freely-se-
output ports. Under this condition, the routing function is unique               lectable parameter of the protocol, along with the number of data
and introduces no optical loss in principle, although any prac-                  minislots, . These parameters may be established by negoti-
tical implementation of the passive hub will exhibit some finite                 ation among the nodes in the network. Analysis and simulation
insertion loss. (For AWG-type devices, the insertion loss is typ-                of the network performance is therefore important to assist in
ically 3–6 dB, and is independent of the number of ports [30].)                  the development of formulae and/or heuristics that enable these
Fig. 5 illustrates the specific routing properties of a four-port                parameters to be optimized during network initialization.
AWG. Entering the hub from Node 1, light of wavelength
is routed back to Node 1; wavelength           is routed to Node 2;
                                                                                        III. INVESTIGATION OF ALLOCATION ALGORITHMS
and so on. However, wavelength           is routed to Node 2 when
entering the hub from Node 2; to Node 3 when entering from                          A node must consider the amount of transmission bandwidth
Node 3; and so on. The routing permutations follow the same                      requested by other nodes once per slot and make appropriate
pattern for the other wavelengths.                                               allocations to the requesting nodes. The goal of the algorithms
   With a star topology a central synchronization point at the                   considered here is to make this allocation fair, in the sense that,
hub of the star is used. Each node transmits its minislots at such               all else being equal, a receiving node will on average allocate
a time that they arrive at the hub at exactly the right time (to                 bandwidth to requesting nodes in proportion to their demands,
SPENCER AND SUMMERFIELD: MAC PROTOCOL FOR WAVELENGTH-ROUTED PONS                                                                         1663



while providing some guaranteed bandwidth to all nodes. The                                       TABLE I
node must also consider the independent allocations received                SYMMETRIC TRAFFIC LOAD DISTRIBUTION FOR SIMULATIONS   OF A
                                                                                             THREE-NODE NETWORK
from other nodes, as they may clash, and make a choice of which
one to use for each data minislot. Three allocation algorithms
have been developed and have been compared using simulations
of a 3-node            symmetrically loaded network.
   The simulations were performed using event driven simu-
lation code written in C. A slot period of 20 s, a data rate of
1 Gb/s, and four data minislots per slot
were assumed. Finite buffers of 40 data minislots per node
per destination were used. Realistic protocol overheads were
accounted for in the simulation as follows. A conservative
guard-time of              ns was assumed, which allows for the
relative drift of node clocks between synchronization events,
as well as differences in propagation delays on different wave-         access to the network capacity. Furthermore, under symmetric
lengths due to fiber dispersion and environmental fluctuations.         heavy loading, the cyclic allocation algorithm may result in re-
The guard time thus contributes 100 bits overhead per minislot.         ceiving nodes making allocations which are identical to the slot
The total of all framing and protocol overheads, each of which          allocations in the I-TDMA protocol, which is known to be op-
is rounded to the next 8-bit multiple (i.e., assuming a byte            timal under these loading conditions.
is the minimum transmission unit), is 996 bits per slot. This              Fig. 6(a)–(d) shows the distribution of delays experienced by
comprises 124 bits for the synchronization minislot (100 for the        data arriving at the source nodes obtained from a number of
guard time, 16 for the framing and 8 for the protocol as a slot         simulation runs at total offered loads of 50%, 65%, 80%, and
identifier), 156 bits for each R/A minislot (100 for the guard          125%, respectively. At offered loads less than the network ca-
time, 16 for framing and 40 for protocol, including 8 each for          pacity most data is queued only as a result of the protocol delay,
the requests and allocations) and 140 bits for each data minislot       thus resulting in short queue lengths, as shown in Fig. 6(a). The
(100 for the guard time, 16 for framing and 24 for protocol             variability of the data arrivals sometimes results in more instan-
overheads). Inclusion of the overheads results in an effective          taneous load than the network can handle, and this results in the
reduction of capacity, as described by (2). Data arrivals to the        excess being queued. As the offered load is increased the queue
network were modeled as packets of uniformly distributed                lengths increase, as shown in Fig. 6(b) but they are limited by
length of 0 to 16 000 bits arriving under a Poisson process.            the buffer capacity. When the buffers are full, further data ar-
The network loading was varied from 50 to 150% by changing              rivals result in loss. Thus, under overloaded conditions the queue
the mean inter-arrival time appropriately. Network overload             lengths approach the buffer capacity. The average rate at which
(greater than 100% loading) was considered to enable us to              the queues are served (i.e., the average number of data minis-
investigate the limiting behavior of the protocol, however we           lots transmitted per second) is proportional to the throughput.
recognize that real networks are unlikely to be operated in this        Data arriving at the queue must wait for all the data already in
regime due to the substantial data losses incurred. Simulation          the queue to be served before it will be sent. Thus the mean
runs of 100 000 slot periods (2 s) were repeated eight times at         queuing delay experienced by data is proportional to the mean
each of these values of loading.                                        queue length and inversely proportional to the mean throughput.
   The network was loaded symmetrically as shown in Table I,            Also, the delay is bounded, as the buffers are of finite length
with 50% of the traffic at each node sent to and received from the      and there is a nonzero worst case network throughput. Protocol
other two nodes. Unity (100%) loading of a source/destination           and transmission delays are incurred in addition to the queuing
route corresponds to a mean arrival rate equal to the raw capacity      delay, however pipelining results in the removal of most of the
of a wavelength neglecting overheads, i.e., 1 Gb/s.                     protocol delay for heavy loading conditions. Thus for the net-
                                                                        work conditions simulated the buffer size of 40 minislots and
                                                                        worst case throughput of one minislot per slot bound the delay
A. Cyclic Allocation                                                    to just above 40 slots. This bound on the delay is clearly visible
   In the cyclic allocation algorithm, each receiving node con-         in Fig. 6(b) and (c).
siders the requests from transmitting nodes in a round robin               The cyclic allocation algorithm is simple, and allows each
fashion [27]. Individual data minislots are allocated sequen-           node to make allocations independently of its location in the
tially, to nodes that have made requests until either there are         network. However, Fig. 6(c) and (d) demonstrate an undesir-
no more requests or no more unallocated minislots. This cyclic          able property of the algorithm which only became apparent
round-robin allocation continues in the next slot, recommencing         during these simulation runs. The graphs clearly show the
from the node following the last one considered in the previous         development of a bimodal delay distribution under conditions
slot. When a transmitting node receives more than one alloca-           of moderate-to-heavy loading. The observed results are due
tion for the same data minislot from different destination nodes,       to an undesirable synchronization and lock-up property of
the conflict is resolved in a similar way, i.e., by cyclic allocation   the cyclic allocation algorithm. Fig. 6(d) shows that under
of such minislots in a round-robin fashion. When executed iden-         very heavy load, there is a tendency for the protocol to
tically at every node, these algorithms result in fair and flexible     fall into one of two possible states, indicated by the gray
1664                                                                                JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 18, NO. 12, DECEMBER 2000




                                                 (a)                                                          (b)




                                                 (c)                                                          (d)
Fig. 6. Distribution of packet delays for the cyclic allocation algorithm in a three-node network under symmetric loading, for average network loads of (a) 50%,
(b) 65%, (c) 80%, and (d) 125%.


histogram (State 1) and the black histogram (State 2) respec-                    Fig. 6(c), the protocol falls in and out of the lock-up states in
tively. This behavior may be explained as follows. As the                        response to short-term peaks in demand, and the observed delay
network loading increases, the amount of data in the queues                      distribution is the time-averaged effect of the protocol spending
will tend to exceed the maximum capacity of a slot. This                         periods of time in both states.
will result in more outstanding requests every slot than there                      Fig. 7 shows the average delay and packet loss under the
are available data minislots, as a result of which the source                    cyclic allocation algorithm as a function of the mean network
nodes will consistently make identical allocations to the des-                   loading. The key feature of these results is the “knee” in the
tination nodes. The cyclic round robin allocation method is                      delay characteristic, which occurs at an average load of about
completely deterministic, and generates the same allocations                     60%. This is preceded by a dramatic rise in the packet loss. The
for as long as this overloaded state exists. State 1, which                      minimum delay under light loads is 110 s (5.5 slots). The delay
has lower mean delay, corresponds to the fortuitous circum-                      results under heavy loading conditions have been averaged over
stance in which the allocations received back at each source                     the length of the simulation runs, and they represent the average
node from the two destination nodes do not clash, i.e., both                     effect of both states. Delay in State 1 limits to 19.1 slots, while
destinations have allocated different data minislots, and the                    State 2 has an average delay of 30.4 slots (a result of the bimodal
transmitter is able to use all allocations. State 2, with higher                 distribution with peaks near 20 and 40 slots delay).
mean delay, corresponds to a situation in which many of the
allocations clash, and the source node is unable to use both                     B. Random Allocation
since it can only transmit to one destination at a time. In the                     One strategy for avoiding the undesirable synchronization of
3-node network, this state corresponds to a drop in available                    the allocations which occurs with the cyclic algorithm is to allo-
capacity to 2/3 of the maximum capacity (            in the case                 cate the minislots randomly instead of sequentially. Similarly, if
of an -node network). It therefore corresponds to a form of                      more than one allocation is received at the source node from two
congestion collapse for the cyclic allocation algorithm, and                     (or more) destination nodes, then a random selection is made.
must be avoided.                                                                 This algorithm is simple, it obviously avoids any deterministic
   Under overloaded conditions, the protocol tends to fall into                  lock-up states, and it retains the property that allocations can
one state or the other and stay there, and the two histograms in                 be made independently of a node’s location within the network.
Fig. 6(d) show the results of separate simulation runs in which                  However, it clearly does not limit to the ideal I-TDMA schedule
each state was observed. Under moderate loading, shown in                        under uniform heavy loading.
SPENCER AND SUMMERFIELD: MAC PROTOCOL FOR WAVELENGTH-ROUTED PONS                                                                                             1665




Fig. 7.   Average delay and packet loss as a function of offered load for the cyclic allocation algorithm in a three-node network under symmetric loading.




                                                  (a)                                                          (b)




                                                  (c)                                                          (d)
Fig. 8. Distribution of packet delays for the random allocation algorithm in a three-node network under symmetric loading, for average network loads of (a) 50%,
(b) 65%, (c) 80%, and (d) 125%.


   Fig. 8(a)–(d) shows the delay distributions for the random al-                  lots per slot) as a result of the random allocation clashes and thus
location algorithm obtained from a number of simulation runs                       the delay limits to an average value of 520 s (26 slots—slightly
at total offered loads of 50%, 65%, 80%, and 125%, respec-                         less than the theoretical 26-2/3 as the buffers never fill com-
tively. The queues evolve in a similar manner to the cyclic allo-                  pletely). The evolution to this condition is clearly visible in
cation algorithm, however in this case the bimodal distribution                    Fig. 8(c) and (d). Fig. 9 shows average delay and packet loss of
does not emerge. As a result, the upper tail of the distribution                   the random allocation algorithm as a function of mean network
at 65% loading shown in Fig. 8(b) is less heavy than the corre-                    loading. The knee of the delay curve occurs at a slightly higher
sponding distribution for the cyclic allocation algorithm. Under                   loading of about 65% compared with the cyclic allocation algo-
heavy loading the peak average capacity is 3/4 (i.e., 1-1/2 minis-                 rithm, and there is a corresponding improvement in packet loss.
1666                                                                                JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 18, NO. 12, DECEMBER 2000




Fig. 9.   Average delay and packet loss as a function of offered load for the random allocation algorithm in a three-node network under symmetric loading.


C. Preferential/Random Allocation                                                 minslot will be preferentially allocated. For example, Node
                                                                                  0 will preferentially allocate the first data minislot to Node
   The cyclic allocation algorithm was designed to converge to                    3, the second to Node 2, the third to Node 1, and so on.
the I-TDMA schedule under uniform heavy loading, but was                          The set of preferential allocations actually made is shown in
found to suffer from undesirable lock-up states. The random al-                   Fig. 10(c)—these allocations are deterministic, and are derived
location algorithm avoids these lock-up states, resulting in im-                  directly from Fig. 10(a) and (b). Each node then makes random
proved average performance, but still underutilizes the available                 allocations of any unallocated data minislots to nodes with
bandwidth under heavy loading. The ideal allocation algorithm                     additional requests. One such random allocation is shown in
would avoid the lock-up states and converge to the I-TDMA                         Fig. 10(d). At Node 0, Data Minislots 0 and 3 are not needed
schedule under uniform heavy loading. We have developed such                      for preferential allocations, and in this example, Data Minislot
an algorithm, which we call the preferential/random (P/R) algo-                   0 has been randomly allocated to Node 2. Likewise, at Node
rithm. Under this scheme the fixed I-TDMA schedule is consid-                     1, Node 0 has been randomly allocated two minislots chosen
ered to be the preferred allocation at the receiving nodes. When                  from Data Minislots 2, 4, and 5. Fig. 10(e) summarizes the
requests arrive from the source nodes, allocations are made ini-                  allocations received by each node, with a cross representing
tially from the I-TDMA schedule. Once this has been done, if                      an allocation. An example of the allocations which may
there are further requests from some source nodes and there are                   subsequently be used, after resolving any clashes, is given in
remaining free data minislots, then these are allocated by con-                   Fig. 10(f). At Node 0, Data Minislot 1, the random allocation
sidering the requests from each source node in a round-robin                      from Node 3 clashes with the preferential allocation from Node
manner (to ensure that the allocation of the remaining capacity                   2, thus the preferential allocation is used. The clash which
is fair on average), and assigning to each source one of the re-                  occurs at Node 0, Data Minislot 4, is between two random
maining data minislots chosen at random (to avoid any deter-                      allocations (from Nodes 1 and 3). One of these is chosen at
ministic lock-up effects). Similarly, if a source node receives al-               random—in this example it is the allocation from Node 1.
locations for the same data minislot from two (or more) different                 Similar choices apply to all the other allocation clashes present.
destinations, if one of these is the preferred I-TDMA allocation,                    The main disadvantage of the P/R algorithm is that it requires
that one is used. If all are the result of random allocations, one                each node to “know” its place in the I-TDMA schedule, and thus
is chosen at random.                                                              it is not independent of the node’s location in the network. In
   An example of the P/R allocation scheme for a 4-node                           practice, the labeling scheme used for nodes and I-TDMA slots
           network with six data minislots              is given in               is arbitrary (i.e., the physical ordering of nodes in the network is
Fig. 10. Fig. 10(a) shows a table of all the requests made for the                not significant), and can easily be assigned as part of the cold-
slot under consideration. The numbers in the table indicate the                   and warm-start procedures which are required in any case. For
number of data minislots requested for transmission by each                       the preferred I-TDMA schedule to map fairly into each slot, the
source node to each destination node. For example, Node 0 is                      additional constraint is imposed that                     , where
requesting four minislots to Node 1, one minislot to Node 2,                      is any positive integer.
and five minislots to Node 3. With             , if Node 0 receives                  The effect of the P/R algorithm is that each node is guaran-
allocations for all 10 requests, it will obviously be able to                     teed at least                of its transmission capacity to every
use six at most, however in the absence of any information                        other node. A node may be able to obtain up to the entire avail-
regarding the requests made by other nodes, the R/A protocol                      able capacity, depending on the requests made by other nodes.
considers all requests independently. The preferential source                     The access to this capacity is shared in a statistically fair way
nodes for the data minislots at each of the destination nodes                     among those requesting it. Under heavy uniform loading the net-
are shown in Fig. 10(b), in which the number in each box                          work limits to the optimal fixed TDMA schedule and does not
represents the node number to which the corresponding data                        suffer the instabilities of the cyclic algorithm. It can be shown,
SPENCER AND SUMMERFIELD: MAC PROTOCOL FOR WAVELENGTH-ROUTED PONS                                                                                           1667




                                                             N                                         M
Fig. 10. Example of the P/R allocation scheme for a 4-node ( = 4) network with six data minislots ( = 6). (a) Requests made to the nodes. (b) Preferential
sources for the data minislots at each node. (c) Preferential allocations made. (d) Possible random allocations made. (e) Summary of the allocations received by
each node. (f) Actual allocations used.


although space does not permit its inclusion in this paper, that                    Simulation of the P/R allocation algorithm resulted in the
all three allocation algorithms produce the same number of allo-                 delay distributions shown in Fig. 11(a)–(d). The guaranteed ca-
cations, the difference between them being in the data minislots                 pacity of two minislots per slot and the buffer size of 40 min-
chosen. Furthermore, allocation clash resolution by the P/R al-                  islots per node per destination ensure that the worst case delay
gorithm can be shown to result in a mean performance that is                     is limited to just above 20 slots. The protocol delay is always
always better than the random algorithm’s, and that under cer-                   present, but is masked by pipelining at high loading.
tain loads (including uniform) the P/R algorithm achieves the                       Fig. 12 shows the average delay and packet loss under the
maximum possible minislot usage.                                                 P/R allocation algorithm as a function of the mean network
1668                                                                                 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 18, NO. 12, DECEMBER 2000




                                                  (a)                                                         (b)




                                                  (c)                                                         (d)
Fig. 11. Distribution of packet delays for the P/R allocation algorithm in a three-node network under symmetric loading, for average network loads of (a) 50%,
(b) 65%, (c) 80%, and (d) 125%.




Fig. 12.   Average delay and packet loss as a function of offered load for the P/R allocation algorithm in a three-node network under symmetric loading.


loading. The maximum average delay under overload conditions                      considered. Therefore, in the remainder of the paper only the
is 360 s (18 slots), comparable to the desirable behavior of the                  P/R algorithm is considered.
cyclic scheme. The knee of the delay characteristic occurs at a
mean network loading level of 80%. At high loading there is full                                    IV. PERFORMANCE SIMULATIONS
utilization of the available network capacity under the I-TDMA
schedule.                                                                         A. Scaling Performance of the Protocol
   The results in this section illustrate the clear superiority of                   By choosing a value for the slot period      in conjunction
the P/R allocation algorithm. It is stable under heavy loading                    with the number of nodes      a tradeoff can be made between
and overload conditions, and exhibits the lowest delay, delay                     the minimum delay at low loads, and the cost of the per-slot
variance and highest average throughput of the three algorithms                   overheads. As discussed previously, in practice,      is a free
SPENCER AND SUMMERFIELD: MAC PROTOCOL FOR WAVELENGTH-ROUTED PONS                                                                                           1669




Fig. 13. Average packet delay for the WRAP protocol under symmetric loading, as a function of total offered network load, for a channel capacity of B      =  1
Gb/s per wavelength and a fixed slot duration of 15 s. Results are shown for three-node (squares), five-node (diamonds), seven-node (triangles), and nine-node
(circles) networks.




Fig. 14. Total throughput for the WRAP protocol under symmetric loading, as a function of total offered network load, for a channel capacity of B = 1 Gb/s per
wavelength and a fixed slot duration of 15 s. Results are shown for three-node (squares), five-node (diamonds), seven-node (triangles), and nine-node (circles)
networks.


parameter in star networks but may be constrained by the                         the overall network capacity grows as           but the number of
fiber length in ring networks. Simulations were carried out                      source/destination pair routes grows as              . These obser-
for          , 5, 7, and 9 and              , 25, 35, and 45 s                   vations are confirmed in the throughput results in Fig. 14, which
with symmetric loading to provide insight into this tradeoff.                    show clearly that the fractional utilization of the raw bandwidth
Symmetric loading was used as it fully utilizes the network                      is reduced for higher numbers of nodes. In this sense, the uni-
capacity under high loads, and is therefore a case of particular                 form symmetric loading example represents a worst-case for the
interest in which the results are relatively easy to interpret.                  protocol.
The number of minislots was chosen to be                                            The relative effect of the increasing overheads with increasing
and the total buffer capacity was fixed at 1 Mbyte per node,                     number of nodes can be mitigated by making corresponding
divided equally amongst all possible destinations. All network                   changes to the slot period         , at the expense of some addi-
and protocol overheads were again included, using the same                       tional delay under light loading. The effect of changing both the
parameters discussed previously.                                                 slot length and the number of nodes, while keeping them propor-
   In the first set of simulations the number of nodes was in-                   tional was simulated. The delay as a function of the total offered
creased while keeping        constant at 15 s. The average delay                 load to the network is shown in Fig. 15, and the corresponding
as a function of the total offered load to the network is shown                  throughput is shown in Fig. 16. The overheads remain at about
in Fig. 13, and the corresponding total throughput is shown in                   the same fraction of the network capacity and as a result the
Fig. 14. The worsening delay performance, with the knee of the                   knees of the delay curves occur at roughly the same fractional
curve occurring at a relatively small fraction of the total net-                 loading in all cases. Similarly, the fractional utilization under
work capacity as the number of nodes increases, is due mainly                    overload conditions is approximately the same in all cases. The
to the effect of increasing overheads. The increase in the max-                  slight increase in delay under light loading is difficult to see in
imum delay under overload is due to the reduced capacity be-                     Fig. 15, as it is small compared to the queuing delays which
tween each source/destination pair, resulting from the fact that                 occur under moderate to heavy loads.
1670                                                                               JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 18, NO. 12, DECEMBER 2000




Fig. 15. Average packet delay for the WRAP protocol under symmetric loading, as a function of total offered network load, for a channel capacity of B    =
                                                                                                                                                         1
Gb/s per wavelength and a slot duration which is increased in proportion to the number of nodes. Results are shown for three-nodes and t = 15 s (squares),
five-nodes and t    = 25 s (diamonds), seven-nodes and t         = 35 s (triangles), and nine-nodes and t     = 45 s (circles).




Fig. 16. Total throughput for the WRAP protocol under symmetric loading, as a function of total offered network load, for a channel capacity of B = 1 Gb/s per
wavelength and a slot duration which is increased in proportion to the number of nodes. Results are shown for three-nodes and t  = 15 s (squares), five-nodes
and t    = 25 s (diamonds), seven-nodes and t          = 35 s (triangles), and nine-nodes and t      = 45 s (circles).

   We conclude that the WRAP protocol scales reasonably well                    topology. In addition, FatMAC has been specifically optimized
with increasing numbers of nodes, so long as the slot period                    for multiprocessor interconnection, and as such tends to trade
is similarly scaled. This scaling property is the result of the                 off bandwidth efficiency in order to obtain lower minimum
fact that each node brings with it additional capacity, in the                  delay. The FatMAC protocol is a time-slotted protocol with
form of an added wavelength. The additional delay caused by                     two phases. In the first, control slots containing the destination
the longer slot period under lightly loaded conditions must be                  of the packet at the head of the queue are broadcast by each
considered. If the geographical size of the network increases as                node. In the second phase the packet from the head of each
nodes are added then the minimum delay remains proportional                     node’s queue is sent, with the order of transmissions to a
to the inherent propagation delays. Since propagation delays are                destination determined by the number of control slots with that
unavoidable, the penalty due to the protocol delays in such a                   destination’s address in the previous phase. This two phase
case is minimal. This additional delay for light loads is small                 procedure is then repeated. A simulation comparison between
in comparison with queuing delays and it is deterministic, and                  WRAP, I-TDMA, and FatMAC has been carried out in order to
thus unlikely to adversely affect multimedia traffic. The effect                assess the relative performance of our protocol.
of this delay on applications using the network is an important                    In the first set of simulations, all protocols were simulated
topic, but is beyond the scope of the present paper.                            for the symmetrically-loaded 3-node case, and for loads ranging
                                                                                from 10 to 150%. In order to make the comparisons as fair as
B. Comparison with Other Applicable Protocols                                   possible, the I-TDMA and FatMAC slot lengths were set equal
                                                                                to the WRAP minislot length. To accommodate FatMAC, the
  Two previously-proposed MAC protocols which could be                          network topology is assumed to be a star, the optical transmit-
applied to WR-PONs of the type considered here are I-TDMA                       ters are assumed to have a WDM broadcast facility, and propa-
[25] and FatMAC [24]. The former protocol is applicable to                      gation delays in the network are assumed to be negligible, i.e.,
both stars and rings, while the latter is only practical in a star              less than the transmission time of the smallest data or protocol
SPENCER AND SUMMERFIELD: MAC PROTOCOL FOR WAVELENGTH-ROUTED PONS                                                                                  1671




                                                                           (a)




                                                                           (b)
Fig. 17. (a) Comparison of the average packet delay for the WRAP (squares), I-TDMA (diamonds), and FatMAC (triangles) protocols for three-node networks
under symmetric loading, as a function of the offered load. (b) Corresponding comparison of packet loss.


unit. While both WRAP and I-TDMA can accommodate prop-                                                   TABLE II
                                                                                 ASYMMETRIC TRAFFIC LOAD DISTRIBUTIONS FOR SIMULATIONS OF AN
agation delays within their slot structures, FatMAC cannot tol-                                     EIGHT-NODE NETWORK
erate propagation delays which exceed its small control slot du-
ration. The delay and loss of these protocols as a function of the
fractional offered load to the network are shown in Fig. 17(a)
and (b) respectively. As expected, FatMAC gives low delays
under light load, but has poor performance as the loading is in-
creased. I-TDMA has similar but better performance to WRAP
for the symmetric loading, as it does not incur as much protocol
delay and has lower protocol overheads. While I-TDMA is op-
timal for symmetric loading with small traffic arrival variations,
it is not flexible, in that the maximum available capacity for any
source/destination route is limited to              of that obtain-
able under WRAP.
    To illustrate this, an eight-node network with traffic loads
as shown in Table II was simulated. Equivalent slot/minislot
lengths were used for the different protocols, and buffer
size per source/destination route chosen to be 40 times the
capacity of a slot/minislot. Loading from 10 to 150% of
the tabulated load was applied, this factor being called the                 this throughput for an example node, Node 4, into the traffic
“Traffic Matrix Scaling Factor” in the figures. The nodes                    sent to the four destinations making up the total (i.e., Nodes
have been numbered 0 through 7. For each protocol we have                    0, 3, 5, and 6); and 3) the corresponding average delays
plotted: 1) the throughput per node, measured as the total                   experienced by traffic sent from Node 4 to each of the four
output to all other nodes in the network; 2) a breakdown of                  destinations. Fig. 18 shows the throughput per node for a)
1672                                                                                JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 18, NO. 12, DECEMBER 2000




                                                                              (a)




                                                                              (b)




                                                                              (c)
Fig. 18. Total throughput achieved by each node in an eight-node network under the asymmetric traffic pattern shown in Table II as a function of the factor by
which the traffic matrix is scaled. Simulation results are shown for (a) I-TDMA, (b) FatMAC, and (c) WRAP.




I-TDMA, b) FatMAC, and c) WRAP; Fig. 19(a)–(c) shows                            tination capacity is sought. Only Node 0 is able to obtain all
the breakdown for traffic sent from node 4; and Fig. 20(a)–(c)                  of the desired throughput as the loading is increased, as indi-
shows the corresponding delays.                                                 cated by the linear growth of throughput as a function of loading
   It can be seen from Fig. 18(a), Fig. 19(a), and Fig. 20(a) that              in Fig. 18(a). This is because the total output from Node 0 is
the fixed bandwidth allocations of I-TDMA lead to low limiting                  uniformly distributed to all seven remote nodes, a circumstance
throughput and high delays if more than the available per-des-                  which is perfectly accommodated by the I-TDMA schedule. The
SPENCER AND SUMMERFIELD: MAC PROTOCOL FOR WAVELENGTH-ROUTED PONS                                                                                        1673




                                                                              (a)




                                                                              (b)




                                                                              (c)
Fig. 19. Breakdown of the throughput achieved by Node 4 to each of the four destination nodes, Node 0, Node 3, Node 5, and Node 6, corresponding to the total
throughput shown in Fig. 18. Simulation results are shown for (a) I-TDMA, (b) FatMAC, and (c) WRAP.


throughput of all other nodes is severely limited due to exhaus-                exhausts the capacity of the slots corresponding to those desti-
tion of the capacity of the per-destination slots in the transmis-              nations in the schedule. The onset of this capacity exhaustion is
sion schedule. This is demonstrated by the flattening out of the                accompanied by the large increases in queuing delays observed
throughput curves in Fig. 18(a). For the particular case of Node                in Fig. 20(a).
4, shown in Fig. 19(a) and 20(a), the small proportion of traffic                  Fig. 18(b) shows that FatMAC obtains better overall
destined to Nodes 0 and 3 is able to be transmitted. However the                throughput than I-TDMA. Fig. 20 also shows that it has the
majority of traffic, which is destined for Nodes 5 and 6, rapidly               lowest delays of all three protocols under light loading and
1674                                                                                 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 18, NO. 12, DECEMBER 2000




                                                                         (a)




                                                                         (b)




                                                                               (c)
Fig. 20. Breakdown of the average delay experienced by packets sent from Node 4 to each of the four destination nodes, Node 0, Node 3, Node 5, and Node 6,
corresponding to the throughput per destination shown in Fig. 19. Simulation results are shown for (a) I-TDMA, (b) FatMAC, and (c) WRAP.


overload conditions. However, Fig. 19(b) demonstrates that                       to Nodes 0, 3, and 6 could be improved, at the expense of a
although Node 4 obtains higher overall throughput under                          reduction in throughput to Node 5. An additional problem
FatMAC than I-TDMA, this is the result of higher throughput                      with FatMAC is that it suffers from head-of-line blocking as it
to Node 5, with traffic to Nodes 0, 3, and 6 fairing somewhat                    only uses one queue per node. The loading levels at which the
worse. This is the result of a fairness problem in FatMAC                        throughput to each destination begins to limit is more clearly
which we have not addressed, although a possible solution                        observed from the locations of the knees in the delay curves of
is discussed in [24]. If this problem were solved, throughput                    Fig. 20(b).
SPENCER AND SUMMERFIELD: MAC PROTOCOL FOR WAVELENGTH-ROUTED PONS                                                                                1675



   Fig. 18(c) shows that the WRAP protocol achieves the highest       the lowest delay and delay variance, and the highest average
maximum throughput for all nodes (note that the vertical scale is     throughput of the three algorithms considered.
twice that of the other two graphs in the Figure). Furthermore, it       The scaling behavior of the protocol for increasing number
shares the available capacity most efficiently amongst the nodes,     of nodes was investigated by simulation, and found to be ac-
as evidenced by the fact that the throughput increases linearly       ceptable provided that the slot length is increased in proportion
with offered load for all nodes to a scaling factor of over 50%.      the number of nodes. This maintains the same relative protocol
When the throughput does level out, it is due to the fact that        overheads but does so at the expense of an increase in the in-
the total offered traffic from all sources to a particular desti-     trinsic protocol delay.
nation exceeds the total capacity available to that destination.         Simulations have shown that WRAP provides both fair
For example, the total traffic offered to the network destined for    and flexible access to the network. It allows bandwidth to be
Node 6 is the highest, so nodes transmitting to that node (i.e.,      allocated according to demand, while providing a guaranteed
Nodes 4 and 7) are the first to be affected. Although Node 0 is       minimum bandwidth allocation between each source-destina-
also transmitting to Node 6, it is not affected in this example, as   tion pair. Comparisons were performed with two alternative
its demand never exceeds the guaranteed minimum provided by           protocols, I-TDMA and FatMAC. For symmetric network
the preferential I-TDMA allocations. This is further illustrated      loading I-TDMA performs optimally, and has no additional
in Fig. 19(c) which shows that the throughput to Node 6 is the        protocol overheads. Under these conditions, WRAP limits
first to level out, at a scaling factor of about 60%. As the total    to the I-TDMA behavior under heavy loads, but has higher
network load is increased, throughput to Node 6 from Node 4           overheads, and larger intrinsic protocol delays. FatMAC has
actually declines, due to increased traffic demands from other        the lowest intrinsic delay, however this is obtained at the
source nodes, however it can never fall below the guaranteed          expense of network utilization. Under asymmetric traffic
minimum. Once again, the knees in the delay curves shown in           conditions WRAP provides the best throughput performance
Fig. 20(c) are the clearest indicator of the loading level at which   in comparison to FatMAC and I-TDMA, offering the I-TDMA
the throughput reaches a maximum.                                     capacity as a guaranteed minimum, and maintaining low delays
   Based on the simulation results in this section, we conclude       for comparatively higher offered loads. We thus conclude that
that I-TDMA is only appropriate in an environment where the           of the three protocols considered here, WRAP is the best-suited
total traffic load on the network is distributed evenly amongst       to general-purpose data communications applications such as
all source-destination pairs. This is unsurprising, as it utilizes    local, campus and metropolitan area networks.
a fixed TDMA schedule. FatMAC exhibits the lowest delays at
low network loading, and it is therefore most appropriate for ap-
                                                                                                ACKNOWLEDGMENT
plications such as the multiprocessor interconnect for which it
was designed, where low latency is more important than high              The authors gratefully acknowledge M. Roughan from the
utilization. The WRAP protocol represents a compromise in             Department of Electrical and Electronic Engineering at the Uni-
which latency at low network loads is traded off for higher band-     versity of Melbourne, for helpful discussions and comments on
width utilization and greater flexibility, while retaining the max-   this work.
imum throughput of I-TDMA as a minimum guaranteed perfor-
mance.
                                                                                                    REFERENCES
                                                                        [1] V. W. S. Chan, K. L. Hall, E. Modiano, and K. A. Rauschenbach, “Ar-
                                                                            chitectures and technologies for high-speed optical data networks,” J.
                        V. CONCLUSION                                       Lightwave Technol., vol. 16, pp. 2146–2168, Dec. 1998.
                                                                        [2] J. M. Senior, M. R. Handley, and M. S. Leeson, “Developments in wave-
   This paper has presented WRAP—a collision-free media ac-                 length division multiple access networking,” IEEE Commun. Mag., pp.
                                                                            28–36, Dec. 1998.
cess protocol for wavelength routed passive optical networks.           [3] L. G. Kazovsky, T. Fong, and T. Hofmeister, “Optical local area network
We have discussed the application of this protocol to ring and              technologies,” IEEE Commun. Mag., pp. 50–54, Dec. 1994.
star topologies. Each node requires a single fixed receiver and         [4] R. T. Hofmeister, L. G. Kazovsky, C. L. Lu, P. Poggliolini, and F. Yang,
                                                                            “CORD: Optical packet-switched network testbed,” Fiber Integr. Opt.,
a single wavelength tunable transmitter. WRAP is a slotted pro-             vol. 16, pp. 199–219, 1997.
tocol that utilizes in-band signaling to send requests to other         [5] M. A. Marsan, A. Bianco, E. Leonardi, M. Meo, and F. Neri, “MAC pro-
nodes and to return the allocations made independently by them.             tocols and fairness control in WDM multirings with tunable transmitters
                                                                            and fixed receivers,” J. Lightwave Technol., vol. 14, pp. 1230–1243, June
As a consequence the protocol does not require any additional               1996.
control wavelength or network, nor does it require a centralized        [6] “Special issue on multiwavelength technology and networks,” J. Light-
controller. Several possible allocation algorithms have been de-            wave Technol., vol. 14, pp. PAGE NOS?–, June 1996.
                                                                        [7] M. S. Goodman, H. Kobrinski, M. P. Vecchi, R. M. Bulley, and J. L.
scribed and their performance compared using computer simu-                 Gimlett, “The LAMBDANET multiwavelength network: Architecture,
lation. It was found that a cyclic round-robin algorithm exhibits           applications and demonstrations,” IEEE J. Sel. Areas Commun., vol. 8,
lock-up states under heavy loads, leading to instability and a              pp. 995–1003, Aug. 1990.
                                                                        [8] T. K. Chiang, S. K. Agrawal, D. T. Mayweather, D. Sadot, C. F. Barry, M.
form of congestion collapse. The lock-up states can be avoided              Hickey, and L. G. Kazovsky, “Implementation of STARNET: A WDM
by using a random allocation algorithm, however this was not                computer communications network,” IEEE J. Sel. Areas Commun., vol.
found to result in a significant improvements in utilization. A             14, pp. 824–839, June 1996.
                                                                        [9] N. R. Dono, P. E. Green Jr., K. Liu, R. Ramaswami, and F. F.-K. Tong,
hybrid scheme, which we have called the preferential/random                 “A wavelength division multiple access network for computer commu-
allocation algorithm, was developed. This algorithm exhibits                nication,” IEEE J. Sel. Areas Commun., vol. 8, pp. 983–993, Aug. 1990.
1676                                                                                 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 18, NO. 12, DECEMBER 2000



 [10] I. Chlamtac, A. Fumagalli, L. G. Kazovsky, and P. T. Poggiolini, “A           [25] K. Bogineni, K. M. Sivalingam, and P. Dowd, “Low-complexity mul-
      contention/collision free WDM ring network for multi gigabit packet                tiple access protocols for wavelength-division multiplexed photonic net-
      switched communication,” J. High Speed Networks, vol. 4, pp. 201–219,              works,” IEEE J. Select. Areas Commun., vol. 11, pp. 590–604, May
      1995.                                                                              1990.
 [11] D. A. Levine and F. Akyildiz, “PROTON: A media access control pro-            [26] M. A Summerfield, “Wavelength division multiplexing for local area
      tocol for optical networks with star topology,” IEEE/ACM Trans. Net-               networks,” in Proc. CLEO/Pacific Rim ’99, Seoul, Korea, August
      working, vol. 3, no. 2, pp. 158–167, April 1995.                                   30–September 3, 1999, Paper FN4, pp. 1215–1216.
 [12] M. S. Borella and B. Mukherjee, “Efficient scheduling of nonuniform           [27] J. Fransson, M. Johansson, M. Roughan, L. Andrew, and M. A.
      packet traffic in a WDM/TDM local lightwave network with arbitrary                 Summerfield, “Design of a medium access control protocol for
      transceiver tuning latencies,” IEEE J. Select. Areas Commun., vol. 14,             a WDMA/TDMA photonic ring network,” in Proc. IEEE Global
      pp. 923–934, June 1996.                                                            Telecommunications Conf. (GLOBECOM’98), Sydney, Australia,
 [13] H. Choi, H. A. Choi, and M. Azizoglu, “Efficient scheduling of trans-              November 8–12, 1998, Paper S89.6.
      mission in optical broadcast networks,” IEEE/ACM Trans. Networking,           [28] L. Kleinrock, Queuing Systems, Volume I: Theory. New York: Wiley,
      vol. 4, pp. 913–920, Dec. 1996.                                                    1976.
 [14] S. Trindandapani, J. S. Meditch, and A. K. Somani, “The MaTPi pro-            [29] M.-S. Chen, N. R. Dono, and R. Ramaswami, “A media-access protocol
      tocol: Masking tuning times through pipelining in WDM optical net-                 for packet-switched wavelength division multiaccess metropolitan area
      works,” in Proc. INFOCOM ’94, June 1994, pp. 1528–1535.                            networks,” IEEE J. Select. Areas Commun., vol. 8, pp. 1048–1057, Aug.
 [15] N. F. Maxemchuk, “Twelve random access strategies for fiber optic net-             1990.
      works,” IEEE Trans. Commun., vol. 36, pp. 942–950, Aug. 1988.                 [30] O. Katsunari, “Planar waveguide devices for WDM communications,”
 [16] S. L. Woodward, P. P. Iannone, K. C. Reichmann, and N. J. Frigo, “A                in Proc. CLEO/Pacific Rim ’99, Seoul, Korea, August 30–September 3,
      spectrally sliced PON employing Fabry–Perot lasers,” IEEE Photon.                  1999, Tutorial TuTA, pp. 112–113.
      Technol. Lett., vol. 10, pp. 1337–1339, Sept. 1998.                           [31] A. C. Kam, K.-Y. Siu, R. A. Barry, and E. A. Swanson, “A cell switching
 [17] R. D. T. Lauder, J. M. Badcock, W. T. Holloway, and D. D. Sampson,                 WDM broadcast LAN with bandwidth guarantee and fair access,” J.
      “WDM ring network employing a shared multiwavelength incoherent                    Lightwave Technol., vol. 16, pp. 2265–2280, Dec. 1998.
      source,” IEEE Photon. Technol. Lett., vol. 10, pp. 294–296, Feb. 1998.        [32] S. F. Su and R. Olshansky, “Performance of multiple access WDM
 [18] D. Guo and A. S. Acampora, “Scalable multihop WDM passive ring                     networks with subcarrier multiplexed control channels,” J. Lightwave
      with optimal wavelength assignment and adaptive wavelength routing,”               Technol., vol. 11, pp. 1028–1033, May/June 1993.
      J. Lightwave Technol., vol. 14, pp. 1264–1277, June 1996.                     [33] K. Kitayama, “Carrier sense using subcarrier-multiplexed signaling for
 [19] B. Glance, U. Koren, C. A. Burrus, and J. D. Evankow, “Discretely-tuned            optical frequency division multiple access,” J. Lightwave Technol., vol.
       N -frequency laser for packet switching applications based on WDM,”               15, pp. 758–765, May 1997.
      Electron. Lett., vol. 27, no. 15, pp. 1381–1383, July 18, 1991.               [34] J. P. Jue, M. S. Borella, and B. Mukherjee, “Performance analysis of
 [20] T.-P. Lee, C. E. Zah, R. Bhat, W. C. Young, B. Pathak, F. Favire, P. S.            the rainbow WDM optical network prototype,” IEEE J. Select. Areas
      D. Lin, N. C. Andreadakis, C. Caneau, A. W. Rahjel, M. Koza, J. K.                 Commun., vol. 14, pp. 945–951, June 1996.
      Gamelin, L. Curtis, D. D. Mahoney, and A. Lepore, “Multiwavelength            [35] O. A. Lavrova, G. Rossi, and D. J. Blumenthal, “Rapid tunable trans-
      DFB laser array transmitters for ONTC reconfigurable optical network               mitter with large number of ITU channels accessible in less than 5 ns,”
      testbed,” J. Lightwave Technol., vol. 14, pp. 967–976, June 1996.                  in Proc. ECOC 2000, vol. 2, Sept. 2000, pp. 169–170.
 [21] A. A. Saavedra, P.-J. Rigole, E. Goobar, R. Schatz, and S. Nilsson,           [36] B. Mason, G. A. Fish, S. P. DenBaars, and L. A. Coldren, “Widely tun-
      “Amplitude and frequency modulation characteristics of widely tunable              able sampled grating DBR laser with integrated electroabsorption mod-
      GCSR lasers,” IEEE Photon. Technol. Lett., vol. 10, pp. 1383–1385,                 ulator,” IEEE Photon. Technol. Lett., vol. 11, pp. 638–640, June 1999.
      Oct. 1998.
 [22] C. R. Doerr, C. H. Joyner, and L. W. Stulz, “40-wavelength rapidly digi-
      tally tunable laser,” IEEE Photon. Technol. Lett., vol. 11, pp. 1348–1350,
      Nov. 1999.
                                                                                   Michael J. Spencer (M’95) photograph and biography not available for the time
 [23] D. J. Blumenthal, A. Carena, L. Rau, V. Curri, and S. Humphries, “All-
                                                                                   of publication.
      optical label swapping with wavelength conversion for WDM-IP net-
      works with subcarrier multiplexed addressing,” IEEE Photon. Technol.
      Lett., vol. 11, pp. 1497–1502, Nov. 1999.
 [24] K. Sivalingam and P. W. Dowd, “A multilevel WDM access protocol
      for an optically interconnected multiprocessor system,” J. Lightwave         Mark A. Summerfield (M’96) photograph and biography not available for the
      Technol., vol. 13, pp. 2152–2167, Nov. 1995.                                 time of publication.

Mais conteúdo relacionado

Mais procurados

Relay Enhanced LTE-Advanced Networks – Resource Allocation and QoS provisioni...
Relay Enhanced LTE-Advanced Networks – Resource Allocation and QoS provisioni...Relay Enhanced LTE-Advanced Networks – Resource Allocation and QoS provisioni...
Relay Enhanced LTE-Advanced Networks – Resource Allocation and QoS provisioni...Eiko Seidel
 
Quality of service in wireless communication
Quality of service in wireless communicationQuality of service in wireless communication
Quality of service in wireless communicationPRADEEP Cheekatla
 
Comparative performance analysis of different modulation techniques for papr ...
Comparative performance analysis of different modulation techniques for papr ...Comparative performance analysis of different modulation techniques for papr ...
Comparative performance analysis of different modulation techniques for papr ...IJCNCJournal
 
Hybrid PAPR Reduction Scheme for Universal Filter Multi-Carrier Modulation in...
Hybrid PAPR Reduction Scheme for Universal Filter Multi-Carrier Modulation in...Hybrid PAPR Reduction Scheme for Universal Filter Multi-Carrier Modulation in...
Hybrid PAPR Reduction Scheme for Universal Filter Multi-Carrier Modulation in...CrimsonPublishersRDMS
 
Duplexing mode, ARB and modulation approaches parameters affection on LTE upl...
Duplexing mode, ARB and modulation approaches parameters affection on LTE upl...Duplexing mode, ARB and modulation approaches parameters affection on LTE upl...
Duplexing mode, ARB and modulation approaches parameters affection on LTE upl...IJECEIAES
 
SYNOPSIS ON IMPLEMENTING SC-FDMA AND OFDMA IN MATLAB
SYNOPSIS ON IMPLEMENTING SC-FDMA AND  OFDMA IN MATLAB  SYNOPSIS ON IMPLEMENTING SC-FDMA AND  OFDMA IN MATLAB
SYNOPSIS ON IMPLEMENTING SC-FDMA AND OFDMA IN MATLAB Murtadha ali shukur
 
Wavelet Packet based Multicarrier Modulation for Cognitive UWB Systems
Wavelet Packet based Multicarrier Modulation for Cognitive UWB SystemsWavelet Packet based Multicarrier Modulation for Cognitive UWB Systems
Wavelet Packet based Multicarrier Modulation for Cognitive UWB SystemsCSCJournals
 
WC & LTE 4G Broadband Module 2- 2019 by Prof.Suresha V
WC & LTE 4G Broadband  Module 2- 2019 by Prof.Suresha VWC & LTE 4G Broadband  Module 2- 2019 by Prof.Suresha V
WC & LTE 4G Broadband Module 2- 2019 by Prof.Suresha VSURESHA V
 
Channel Estimation Techniques in MIMO-OFDM LTE SystemsCauses and Effects of C...
Channel Estimation Techniques in MIMO-OFDM LTE SystemsCauses and Effects of C...Channel Estimation Techniques in MIMO-OFDM LTE SystemsCauses and Effects of C...
Channel Estimation Techniques in MIMO-OFDM LTE SystemsCauses and Effects of C...IJERA Editor
 
Cognitive Radio 802.22 in NetSim
Cognitive Radio 802.22 in NetSimCognitive Radio 802.22 in NetSim
Cognitive Radio 802.22 in NetSimSANJAY ANAND
 
IRJET- Survey Paper on Performance Evaluation of 5G WiMAX (IEEE 802.16) Syste...
IRJET- Survey Paper on Performance Evaluation of 5G WiMAX (IEEE 802.16) Syste...IRJET- Survey Paper on Performance Evaluation of 5G WiMAX (IEEE 802.16) Syste...
IRJET- Survey Paper on Performance Evaluation of 5G WiMAX (IEEE 802.16) Syste...IRJET Journal
 
Performance Investigation of OFDM-FSO System under Diverse Weather Conditions...
Performance Investigation of OFDM-FSO System under Diverse Weather Conditions...Performance Investigation of OFDM-FSO System under Diverse Weather Conditions...
Performance Investigation of OFDM-FSO System under Diverse Weather Conditions...IJECEIAES
 
Implementation and Study of Universal Filtered Multi Carrier under Carrier Fr...
Implementation and Study of Universal Filtered Multi Carrier under Carrier Fr...Implementation and Study of Universal Filtered Multi Carrier under Carrier Fr...
Implementation and Study of Universal Filtered Multi Carrier under Carrier Fr...Editor IJAIEM
 
DISCRETE COSINETRANSFORM-II FOR REDUCTION IN PEAK TO AVERAGE POWER RATIO OF O...
DISCRETE COSINETRANSFORM-II FOR REDUCTION IN PEAK TO AVERAGE POWER RATIO OF O...DISCRETE COSINETRANSFORM-II FOR REDUCTION IN PEAK TO AVERAGE POWER RATIO OF O...
DISCRETE COSINETRANSFORM-II FOR REDUCTION IN PEAK TO AVERAGE POWER RATIO OF O...ijwmn
 

Mais procurados (20)

Relay Enhanced LTE-Advanced Networks – Resource Allocation and QoS provisioni...
Relay Enhanced LTE-Advanced Networks – Resource Allocation and QoS provisioni...Relay Enhanced LTE-Advanced Networks – Resource Allocation and QoS provisioni...
Relay Enhanced LTE-Advanced Networks – Resource Allocation and QoS provisioni...
 
5G
5G5G
5G
 
Quality of service in wireless communication
Quality of service in wireless communicationQuality of service in wireless communication
Quality of service in wireless communication
 
Comparative performance analysis of different modulation techniques for papr ...
Comparative performance analysis of different modulation techniques for papr ...Comparative performance analysis of different modulation techniques for papr ...
Comparative performance analysis of different modulation techniques for papr ...
 
Hybrid PAPR Reduction Scheme for Universal Filter Multi-Carrier Modulation in...
Hybrid PAPR Reduction Scheme for Universal Filter Multi-Carrier Modulation in...Hybrid PAPR Reduction Scheme for Universal Filter Multi-Carrier Modulation in...
Hybrid PAPR Reduction Scheme for Universal Filter Multi-Carrier Modulation in...
 
Duplexing mode, ARB and modulation approaches parameters affection on LTE upl...
Duplexing mode, ARB and modulation approaches parameters affection on LTE upl...Duplexing mode, ARB and modulation approaches parameters affection on LTE upl...
Duplexing mode, ARB and modulation approaches parameters affection on LTE upl...
 
SYNOPSIS ON IMPLEMENTING SC-FDMA AND OFDMA IN MATLAB
SYNOPSIS ON IMPLEMENTING SC-FDMA AND  OFDMA IN MATLAB  SYNOPSIS ON IMPLEMENTING SC-FDMA AND  OFDMA IN MATLAB
SYNOPSIS ON IMPLEMENTING SC-FDMA AND OFDMA IN MATLAB
 
Manmohan 15% PLAGIARISM REPORT M.TECH ( M S WORD FILE )
Manmohan 15% PLAGIARISM REPORT M.TECH ( M S WORD  FILE )Manmohan 15% PLAGIARISM REPORT M.TECH ( M S WORD  FILE )
Manmohan 15% PLAGIARISM REPORT M.TECH ( M S WORD FILE )
 
6 chapter 1,2,3,4,5 (15%) M.TECH ( M S WORD FILE )
6 chapter 1,2,3,4,5 (15%) M.TECH ( M S WORD FILE )6 chapter 1,2,3,4,5 (15%) M.TECH ( M S WORD FILE )
6 chapter 1,2,3,4,5 (15%) M.TECH ( M S WORD FILE )
 
6 chapter 1,2,3,4,5 (15%) M.TECH ( PDF FILE )
6 chapter 1,2,3,4,5 (15%) M.TECH ( PDF FILE )6 chapter 1,2,3,4,5 (15%) M.TECH ( PDF FILE )
6 chapter 1,2,3,4,5 (15%) M.TECH ( PDF FILE )
 
Wavelet Packet based Multicarrier Modulation for Cognitive UWB Systems
Wavelet Packet based Multicarrier Modulation for Cognitive UWB SystemsWavelet Packet based Multicarrier Modulation for Cognitive UWB Systems
Wavelet Packet based Multicarrier Modulation for Cognitive UWB Systems
 
WC & LTE 4G Broadband Module 2- 2019 by Prof.Suresha V
WC & LTE 4G Broadband  Module 2- 2019 by Prof.Suresha VWC & LTE 4G Broadband  Module 2- 2019 by Prof.Suresha V
WC & LTE 4G Broadband Module 2- 2019 by Prof.Suresha V
 
Channel Estimation Techniques in MIMO-OFDM LTE SystemsCauses and Effects of C...
Channel Estimation Techniques in MIMO-OFDM LTE SystemsCauses and Effects of C...Channel Estimation Techniques in MIMO-OFDM LTE SystemsCauses and Effects of C...
Channel Estimation Techniques in MIMO-OFDM LTE SystemsCauses and Effects of C...
 
Cognitive Radio 802.22 in NetSim
Cognitive Radio 802.22 in NetSimCognitive Radio 802.22 in NetSim
Cognitive Radio 802.22 in NetSim
 
IRJET- Survey Paper on Performance Evaluation of 5G WiMAX (IEEE 802.16) Syste...
IRJET- Survey Paper on Performance Evaluation of 5G WiMAX (IEEE 802.16) Syste...IRJET- Survey Paper on Performance Evaluation of 5G WiMAX (IEEE 802.16) Syste...
IRJET- Survey Paper on Performance Evaluation of 5G WiMAX (IEEE 802.16) Syste...
 
Performance Investigation of OFDM-FSO System under Diverse Weather Conditions...
Performance Investigation of OFDM-FSO System under Diverse Weather Conditions...Performance Investigation of OFDM-FSO System under Diverse Weather Conditions...
Performance Investigation of OFDM-FSO System under Diverse Weather Conditions...
 
Implementation and Study of Universal Filtered Multi Carrier under Carrier Fr...
Implementation and Study of Universal Filtered Multi Carrier under Carrier Fr...Implementation and Study of Universal Filtered Multi Carrier under Carrier Fr...
Implementation and Study of Universal Filtered Multi Carrier under Carrier Fr...
 
DISCRETE COSINETRANSFORM-II FOR REDUCTION IN PEAK TO AVERAGE POWER RATIO OF O...
DISCRETE COSINETRANSFORM-II FOR REDUCTION IN PEAK TO AVERAGE POWER RATIO OF O...DISCRETE COSINETRANSFORM-II FOR REDUCTION IN PEAK TO AVERAGE POWER RATIO OF O...
DISCRETE COSINETRANSFORM-II FOR REDUCTION IN PEAK TO AVERAGE POWER RATIO OF O...
 
OFDMA principles
OFDMA principlesOFDMA principles
OFDMA principles
 
Lte optimization
Lte optimizationLte optimization
Lte optimization
 

Destaque

Destaque (6)

Gcsr bank part 2
Gcsr bank part 2 Gcsr bank part 2
Gcsr bank part 2
 
Armenia Country (GCSR)
Armenia Country (GCSR)Armenia Country (GCSR)
Armenia Country (GCSR)
 
oman country GCSR by vishalkumar
oman country GCSR by vishalkumaroman country GCSR by vishalkumar
oman country GCSR by vishalkumar
 
GCSR
GCSRGCSR
GCSR
 
Gcsr part 1
Gcsr part 1Gcsr part 1
Gcsr part 1
 
Project front page, index, certificate, and acknowledgement
Project front page, index, certificate, and acknowledgementProject front page, index, certificate, and acknowledgement
Project front page, index, certificate, and acknowledgement
 

Semelhante a IEEE paper

Design of time division multiplexing/wavelength division multiplexing passiv...
Design of time division multiplexing/wavelength division  multiplexing passiv...Design of time division multiplexing/wavelength division  multiplexing passiv...
Design of time division multiplexing/wavelength division multiplexing passiv...IJECEIAES
 
10-Gb/S Transmission of Wdm Pon for Man with 50km Reach Based On Ftth
10-Gb/S Transmission of Wdm Pon for Man with 50km Reach Based On Ftth10-Gb/S Transmission of Wdm Pon for Man with 50km Reach Based On Ftth
10-Gb/S Transmission of Wdm Pon for Man with 50km Reach Based On FtthIJERA Editor
 
RECONFIGURABLE OPTICAL ADD AND DROP MULTIPLEXERS A REVIEW
RECONFIGURABLE OPTICAL ADD AND DROP MULTIPLEXERS A REVIEWRECONFIGURABLE OPTICAL ADD AND DROP MULTIPLEXERS A REVIEW
RECONFIGURABLE OPTICAL ADD AND DROP MULTIPLEXERS A REVIEWecijjournal
 
RECONFIGURABLE OPTICAL ADD AND DROP MULTIPLEXERS A REVIEW
RECONFIGURABLE OPTICAL ADD AND DROP MULTIPLEXERS A REVIEW RECONFIGURABLE OPTICAL ADD AND DROP MULTIPLEXERS A REVIEW
RECONFIGURABLE OPTICAL ADD AND DROP MULTIPLEXERS A REVIEW ecij
 
Performance evaluation of VLC system using new modulation approach
Performance evaluation of VLC system using new modulation approachPerformance evaluation of VLC system using new modulation approach
Performance evaluation of VLC system using new modulation approachjournalBEEI
 
Performance Analysis of Ultra Wideband Communication System
Performance Analysis of Ultra Wideband Communication SystemPerformance Analysis of Ultra Wideband Communication System
Performance Analysis of Ultra Wideband Communication SystemEditor IJMTER
 
optical space division multiplexing
optical space division multiplexingoptical space division multiplexing
optical space division multiplexingmohammedalimahdi
 
A simulation study of wi max based communication system using deliberately cl...
A simulation study of wi max based communication system using deliberately cl...A simulation study of wi max based communication system using deliberately cl...
A simulation study of wi max based communication system using deliberately cl...eSAT Publishing House
 
A simulation study of wi max based communication system using deliberately cl...
A simulation study of wi max based communication system using deliberately cl...A simulation study of wi max based communication system using deliberately cl...
A simulation study of wi max based communication system using deliberately cl...eSAT Journals
 
An adaptive channel estimation scheme based on redundancy minimization for fi...
An adaptive channel estimation scheme based on redundancy minimization for fi...An adaptive channel estimation scheme based on redundancy minimization for fi...
An adaptive channel estimation scheme based on redundancy minimization for fi...TELKOMNIKA JOURNAL
 
Performance analysis of multilayer multicast MANET CRN based on steiner minim...
Performance analysis of multilayer multicast MANET CRN based on steiner minim...Performance analysis of multilayer multicast MANET CRN based on steiner minim...
Performance analysis of multilayer multicast MANET CRN based on steiner minim...TELKOMNIKA JOURNAL
 
Dynamic Spectrum Allocation in Wireless sensor Networks
Dynamic Spectrum Allocation in Wireless sensor NetworksDynamic Spectrum Allocation in Wireless sensor Networks
Dynamic Spectrum Allocation in Wireless sensor NetworksIJMER
 
Radio network dimensioning and planning for wimax networks
Radio network dimensioning and planning for wimax networksRadio network dimensioning and planning for wimax networks
Radio network dimensioning and planning for wimax networksPfedya
 
Wireless Communication fundamentals and 4G Technology
Wireless Communication fundamentals and 4G TechnologyWireless Communication fundamentals and 4G Technology
Wireless Communication fundamentals and 4G TechnologyKiranShanbhag9
 
Analysis of System Capacity and Spectral Efficiency of Fixed-Grid Network
Analysis of System Capacity and Spectral Efficiency of Fixed-Grid NetworkAnalysis of System Capacity and Spectral Efficiency of Fixed-Grid Network
Analysis of System Capacity and Spectral Efficiency of Fixed-Grid NetworkIJCNCJournal
 
Analysis of System Capacity and Spectral Efficiency of Fixed-Grid Network
Analysis of System Capacity and Spectral Efficiency of Fixed-Grid NetworkAnalysis of System Capacity and Spectral Efficiency of Fixed-Grid Network
Analysis of System Capacity and Spectral Efficiency of Fixed-Grid NetworkIJCNCJournal
 

Semelhante a IEEE paper (20)

Design of time division multiplexing/wavelength division multiplexing passiv...
Design of time division multiplexing/wavelength division  multiplexing passiv...Design of time division multiplexing/wavelength division  multiplexing passiv...
Design of time division multiplexing/wavelength division multiplexing passiv...
 
10-Gb/S Transmission of Wdm Pon for Man with 50km Reach Based On Ftth
10-Gb/S Transmission of Wdm Pon for Man with 50km Reach Based On Ftth10-Gb/S Transmission of Wdm Pon for Man with 50km Reach Based On Ftth
10-Gb/S Transmission of Wdm Pon for Man with 50km Reach Based On Ftth
 
RECONFIGURABLE OPTICAL ADD AND DROP MULTIPLEXERS A REVIEW
RECONFIGURABLE OPTICAL ADD AND DROP MULTIPLEXERS A REVIEWRECONFIGURABLE OPTICAL ADD AND DROP MULTIPLEXERS A REVIEW
RECONFIGURABLE OPTICAL ADD AND DROP MULTIPLEXERS A REVIEW
 
RECONFIGURABLE OPTICAL ADD AND DROP MULTIPLEXERS A REVIEW
RECONFIGURABLE OPTICAL ADD AND DROP MULTIPLEXERS A REVIEW RECONFIGURABLE OPTICAL ADD AND DROP MULTIPLEXERS A REVIEW
RECONFIGURABLE OPTICAL ADD AND DROP MULTIPLEXERS A REVIEW
 
Performance evaluation of VLC system using new modulation approach
Performance evaluation of VLC system using new modulation approachPerformance evaluation of VLC system using new modulation approach
Performance evaluation of VLC system using new modulation approach
 
Performance Analysis of Ultra Wideband Communication System
Performance Analysis of Ultra Wideband Communication SystemPerformance Analysis of Ultra Wideband Communication System
Performance Analysis of Ultra Wideband Communication System
 
R04504114117
R04504114117R04504114117
R04504114117
 
optical space division multiplexing
optical space division multiplexingoptical space division multiplexing
optical space division multiplexing
 
A simulation study of wi max based communication system using deliberately cl...
A simulation study of wi max based communication system using deliberately cl...A simulation study of wi max based communication system using deliberately cl...
A simulation study of wi max based communication system using deliberately cl...
 
A simulation study of wi max based communication system using deliberately cl...
A simulation study of wi max based communication system using deliberately cl...A simulation study of wi max based communication system using deliberately cl...
A simulation study of wi max based communication system using deliberately cl...
 
An adaptive channel estimation scheme based on redundancy minimization for fi...
An adaptive channel estimation scheme based on redundancy minimization for fi...An adaptive channel estimation scheme based on redundancy minimization for fi...
An adaptive channel estimation scheme based on redundancy minimization for fi...
 
Performance analysis of multilayer multicast MANET CRN based on steiner minim...
Performance analysis of multilayer multicast MANET CRN based on steiner minim...Performance analysis of multilayer multicast MANET CRN based on steiner minim...
Performance analysis of multilayer multicast MANET CRN based on steiner minim...
 
Dynamic Spectrum Allocation in Wireless sensor Networks
Dynamic Spectrum Allocation in Wireless sensor NetworksDynamic Spectrum Allocation in Wireless sensor Networks
Dynamic Spectrum Allocation in Wireless sensor Networks
 
Radio network dimensioning and planning for wimax networks
Radio network dimensioning and planning for wimax networksRadio network dimensioning and planning for wimax networks
Radio network dimensioning and planning for wimax networks
 
media_radio_mobile_wi_max_
 media_radio_mobile_wi_max_ media_radio_mobile_wi_max_
media_radio_mobile_wi_max_
 
07114069
0711406907114069
07114069
 
Wireless Communication fundamentals and 4G Technology
Wireless Communication fundamentals and 4G TechnologyWireless Communication fundamentals and 4G Technology
Wireless Communication fundamentals and 4G Technology
 
Ew31992997
Ew31992997Ew31992997
Ew31992997
 
Analysis of System Capacity and Spectral Efficiency of Fixed-Grid Network
Analysis of System Capacity and Spectral Efficiency of Fixed-Grid NetworkAnalysis of System Capacity and Spectral Efficiency of Fixed-Grid Network
Analysis of System Capacity and Spectral Efficiency of Fixed-Grid Network
 
Analysis of System Capacity and Spectral Efficiency of Fixed-Grid Network
Analysis of System Capacity and Spectral Efficiency of Fixed-Grid NetworkAnalysis of System Capacity and Spectral Efficiency of Fixed-Grid Network
Analysis of System Capacity and Spectral Efficiency of Fixed-Grid Network
 

Último

Apidays New York 2024 - Accelerating FinTech Innovation by Vasa Krishnan, Fin...
Apidays New York 2024 - Accelerating FinTech Innovation by Vasa Krishnan, Fin...Apidays New York 2024 - Accelerating FinTech Innovation by Vasa Krishnan, Fin...
Apidays New York 2024 - Accelerating FinTech Innovation by Vasa Krishnan, Fin...apidays
 
Rising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdf
Rising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdfRising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdf
Rising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdfOrbitshub
 
Apidays New York 2024 - APIs in 2030: The Risk of Technological Sleepwalk by ...
Apidays New York 2024 - APIs in 2030: The Risk of Technological Sleepwalk by ...Apidays New York 2024 - APIs in 2030: The Risk of Technological Sleepwalk by ...
Apidays New York 2024 - APIs in 2030: The Risk of Technological Sleepwalk by ...apidays
 
Connector Corner: Accelerate revenue generation using UiPath API-centric busi...
Connector Corner: Accelerate revenue generation using UiPath API-centric busi...Connector Corner: Accelerate revenue generation using UiPath API-centric busi...
Connector Corner: Accelerate revenue generation using UiPath API-centric busi...DianaGray10
 
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...Orbitshub
 
TrustArc Webinar - Unlock the Power of AI-Driven Data Discovery
TrustArc Webinar - Unlock the Power of AI-Driven Data DiscoveryTrustArc Webinar - Unlock the Power of AI-Driven Data Discovery
TrustArc Webinar - Unlock the Power of AI-Driven Data DiscoveryTrustArc
 
Six Myths about Ontologies: The Basics of Formal Ontology
Six Myths about Ontologies: The Basics of Formal OntologySix Myths about Ontologies: The Basics of Formal Ontology
Six Myths about Ontologies: The Basics of Formal Ontologyjohnbeverley2021
 
Finding Java's Hidden Performance Traps @ DevoxxUK 2024
Finding Java's Hidden Performance Traps @ DevoxxUK 2024Finding Java's Hidden Performance Traps @ DevoxxUK 2024
Finding Java's Hidden Performance Traps @ DevoxxUK 2024Victor Rentea
 
Exploring Multimodal Embeddings with Milvus
Exploring Multimodal Embeddings with MilvusExploring Multimodal Embeddings with Milvus
Exploring Multimodal Embeddings with MilvusZilliz
 
ICT role in 21st century education and its challenges
ICT role in 21st century education and its challengesICT role in 21st century education and its challenges
ICT role in 21st century education and its challengesrafiqahmad00786416
 
Boost Fertility New Invention Ups Success Rates.pdf
Boost Fertility New Invention Ups Success Rates.pdfBoost Fertility New Invention Ups Success Rates.pdf
Boost Fertility New Invention Ups Success Rates.pdfsudhanshuwaghmare1
 
AWS Community Day CPH - Three problems of Terraform
AWS Community Day CPH - Three problems of TerraformAWS Community Day CPH - Three problems of Terraform
AWS Community Day CPH - Three problems of TerraformAndrey Devyatkin
 
MINDCTI Revenue Release Quarter One 2024
MINDCTI Revenue Release Quarter One 2024MINDCTI Revenue Release Quarter One 2024
MINDCTI Revenue Release Quarter One 2024MIND CTI
 
Web Form Automation for Bonterra Impact Management (fka Social Solutions Apri...
Web Form Automation for Bonterra Impact Management (fka Social Solutions Apri...Web Form Automation for Bonterra Impact Management (fka Social Solutions Apri...
Web Form Automation for Bonterra Impact Management (fka Social Solutions Apri...Jeffrey Haguewood
 
Apidays New York 2024 - Scaling API-first by Ian Reasor and Radu Cotescu, Adobe
Apidays New York 2024 - Scaling API-first by Ian Reasor and Radu Cotescu, AdobeApidays New York 2024 - Scaling API-first by Ian Reasor and Radu Cotescu, Adobe
Apidays New York 2024 - Scaling API-first by Ian Reasor and Radu Cotescu, Adobeapidays
 
DBX First Quarter 2024 Investor Presentation
DBX First Quarter 2024 Investor PresentationDBX First Quarter 2024 Investor Presentation
DBX First Quarter 2024 Investor PresentationDropbox
 
CNIC Information System with Pakdata Cf In Pakistan
CNIC Information System with Pakdata Cf In PakistanCNIC Information System with Pakdata Cf In Pakistan
CNIC Information System with Pakdata Cf In Pakistandanishmna97
 
Apidays New York 2024 - Passkeys: Developing APIs to enable passwordless auth...
Apidays New York 2024 - Passkeys: Developing APIs to enable passwordless auth...Apidays New York 2024 - Passkeys: Developing APIs to enable passwordless auth...
Apidays New York 2024 - Passkeys: Developing APIs to enable passwordless auth...apidays
 

Último (20)

Apidays New York 2024 - Accelerating FinTech Innovation by Vasa Krishnan, Fin...
Apidays New York 2024 - Accelerating FinTech Innovation by Vasa Krishnan, Fin...Apidays New York 2024 - Accelerating FinTech Innovation by Vasa Krishnan, Fin...
Apidays New York 2024 - Accelerating FinTech Innovation by Vasa Krishnan, Fin...
 
Rising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdf
Rising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdfRising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdf
Rising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdf
 
Apidays New York 2024 - APIs in 2030: The Risk of Technological Sleepwalk by ...
Apidays New York 2024 - APIs in 2030: The Risk of Technological Sleepwalk by ...Apidays New York 2024 - APIs in 2030: The Risk of Technological Sleepwalk by ...
Apidays New York 2024 - APIs in 2030: The Risk of Technological Sleepwalk by ...
 
Connector Corner: Accelerate revenue generation using UiPath API-centric busi...
Connector Corner: Accelerate revenue generation using UiPath API-centric busi...Connector Corner: Accelerate revenue generation using UiPath API-centric busi...
Connector Corner: Accelerate revenue generation using UiPath API-centric busi...
 
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
 
TrustArc Webinar - Unlock the Power of AI-Driven Data Discovery
TrustArc Webinar - Unlock the Power of AI-Driven Data DiscoveryTrustArc Webinar - Unlock the Power of AI-Driven Data Discovery
TrustArc Webinar - Unlock the Power of AI-Driven Data Discovery
 
Six Myths about Ontologies: The Basics of Formal Ontology
Six Myths about Ontologies: The Basics of Formal OntologySix Myths about Ontologies: The Basics of Formal Ontology
Six Myths about Ontologies: The Basics of Formal Ontology
 
Finding Java's Hidden Performance Traps @ DevoxxUK 2024
Finding Java's Hidden Performance Traps @ DevoxxUK 2024Finding Java's Hidden Performance Traps @ DevoxxUK 2024
Finding Java's Hidden Performance Traps @ DevoxxUK 2024
 
Exploring Multimodal Embeddings with Milvus
Exploring Multimodal Embeddings with MilvusExploring Multimodal Embeddings with Milvus
Exploring Multimodal Embeddings with Milvus
 
ICT role in 21st century education and its challenges
ICT role in 21st century education and its challengesICT role in 21st century education and its challenges
ICT role in 21st century education and its challenges
 
Boost Fertility New Invention Ups Success Rates.pdf
Boost Fertility New Invention Ups Success Rates.pdfBoost Fertility New Invention Ups Success Rates.pdf
Boost Fertility New Invention Ups Success Rates.pdf
 
AWS Community Day CPH - Three problems of Terraform
AWS Community Day CPH - Three problems of TerraformAWS Community Day CPH - Three problems of Terraform
AWS Community Day CPH - Three problems of Terraform
 
MINDCTI Revenue Release Quarter One 2024
MINDCTI Revenue Release Quarter One 2024MINDCTI Revenue Release Quarter One 2024
MINDCTI Revenue Release Quarter One 2024
 
Understanding the FAA Part 107 License ..
Understanding the FAA Part 107 License ..Understanding the FAA Part 107 License ..
Understanding the FAA Part 107 License ..
 
Web Form Automation for Bonterra Impact Management (fka Social Solutions Apri...
Web Form Automation for Bonterra Impact Management (fka Social Solutions Apri...Web Form Automation for Bonterra Impact Management (fka Social Solutions Apri...
Web Form Automation for Bonterra Impact Management (fka Social Solutions Apri...
 
Apidays New York 2024 - Scaling API-first by Ian Reasor and Radu Cotescu, Adobe
Apidays New York 2024 - Scaling API-first by Ian Reasor and Radu Cotescu, AdobeApidays New York 2024 - Scaling API-first by Ian Reasor and Radu Cotescu, Adobe
Apidays New York 2024 - Scaling API-first by Ian Reasor and Radu Cotescu, Adobe
 
+971581248768>> SAFE AND ORIGINAL ABORTION PILLS FOR SALE IN DUBAI AND ABUDHA...
+971581248768>> SAFE AND ORIGINAL ABORTION PILLS FOR SALE IN DUBAI AND ABUDHA...+971581248768>> SAFE AND ORIGINAL ABORTION PILLS FOR SALE IN DUBAI AND ABUDHA...
+971581248768>> SAFE AND ORIGINAL ABORTION PILLS FOR SALE IN DUBAI AND ABUDHA...
 
DBX First Quarter 2024 Investor Presentation
DBX First Quarter 2024 Investor PresentationDBX First Quarter 2024 Investor Presentation
DBX First Quarter 2024 Investor Presentation
 
CNIC Information System with Pakdata Cf In Pakistan
CNIC Information System with Pakdata Cf In PakistanCNIC Information System with Pakdata Cf In Pakistan
CNIC Information System with Pakdata Cf In Pakistan
 
Apidays New York 2024 - Passkeys: Developing APIs to enable passwordless auth...
Apidays New York 2024 - Passkeys: Developing APIs to enable passwordless auth...Apidays New York 2024 - Passkeys: Developing APIs to enable passwordless auth...
Apidays New York 2024 - Passkeys: Developing APIs to enable passwordless auth...
 

IEEE paper

  • 1. JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 18, NO. 12, DECEMBER 2000 1657 WRAP: A Medium Access Control Protocol for Wavelength-Routed Passive Optical Networks Michael J. Spencer, Member, IEEE, and Mark A. Summerfield, Member, IEEE Abstract—We describe the WDM Request/Allocation Protocol of the medium access control (MAC) protocols employed [34]. (WRAP), a media-access control protocol for wavelength-routed In broadcast networks, simplifying the node hardware by using passive optical networks (WR-PONs) in which each node has a fixed transmitters or receivers limits the number of nodes to the single fixed optical receiver and a single tunable optical trans- mitter. The protocol does not require a carrier sensing capability, a number of available wavelengths [1]. Furthermore, existing and separate control channel, or any centralized control or scheduling. emerging tunable receiver technologies do not provide the wide Access to transmission channels is regulated by allocations made tuning range and rapid tunability required to implement flexible at destination nodes in response to requests made by source and efficient packet networks. Attempts have been made to nodes. Computer simulation is used to investigate three different work around tuning speed limitations using MAC protocols allocation algorithms, one of which—the Preferential/Random algorithm—is shown to provide significantly better performance which aggregate packets for transmission, and use scheduling than the alternatives. Simulations are presented comparing the and “pipelining” techniques to mask the effects of slow-tuning performance of WRAP to two previously proposed applicable pro- components [12]–[14]. Such protocols mitigate the effects tocols—the Interleaved Time Division Multiple Access (I-TDMA) of tuning time under conditions of heavy network loading, protocol, and the FatMAC protocol. WRAP is shown to provide however they cannot overcome the fundamental limitations of fair and flexible access to the transmission capacity, enabling high network utilization to be achieved under a wide range of traffic slow-tuning devices, and are characterized by high packet delay conditions, while providing a guaranteed minimum bandwidth even under lightly-loaded conditions, and high delay variance between each source–destination pair. We conclude that of the under heavy loading. Finally, many of the proposed protocols three protocols considered here, WRAP is the best-suited to for WDM LANs require the use of an additional control general-purpose data communications applications such as local, channel and/or the provision of a “carrier-sense” capability, campus, and metropolitan area networks. resulting in further cost and complexity in the implementation Index Terms—Media-access control (MAC) protocols, passive of the network nodes [1], [5], [15], [31]–[33]. optical networks, wavelength division multiplexing, wavelength routing. Recent advances in optical component technologies, and in particular the development and commercialization of wavelength-selective components such as fiber-Bragg grat- I. INTRODUCTION ings (FBGs) and arrayed waveguide gratings (AWGs), make I N RECENT years, there has been a great deal of interest in the application of wavelength division multiplexing (WDM) to local area networks (LANs) [1]–[5]. The principal advantage wavelength-routed PONs (WR-PONs) a practical alternative to broadcast topologies. Wavelength-routed star networks can be constructed by using an AWG in place of a wavelength-in- of WDM is that it makes available the enormous bandwidth of dependent coupler—a technique which has been proposed optical fiber while allowing the individual wavelength channels for low-cost optical access networks [16]. Wavelength-routed to operate at bit-rates attainable using affordable electronic ring and bus networks can be built using WDM add-drop components [6]. Most proposed WDM LANs have been based multiplexers (W-ADMs) in place of wavelength-independent upon broadcast passive optical network (PON) architectures splitters [5], [17]. In a wavelength-routed network, the trans- employing either a star topology based on a central passive mitter wavelength acts as a form of addressing, determining coupler [7]–[9], or a ring topology with a simple optical the destination of a transmission over a single optical hop. splitter at each node [10]. These broadcast PON architectures Potential advantages of this approach include improved power are conceptually simple, and can be constructed from optical budget and superior security (since signals are not needlessly components which have been commercially available for over and indiscriminately broadcast to nodes other than the intended a decade. However, the penalty for this simplicity in the optical destination) and the ability to support more nodes than there components is generally either: an increase in the complexity of are wavelengths available through the use of multihop paths the network nodes, requiring, for example, tunable transmitters [18]. Furthermore, recent advances in multiwavelength source and tunable receivers at every node [1], [11]; and/or relatively technology indicate that rapidly-tunable sources with wide low channel utilization and throughput due to the inefficiency tuning ranges will become a commercial reality in the foresee- able future [19]–[22], [35], [36]. Grating assisted codirectional Manuscript received April 3, 2000; revised September 26, 2000. Coupler with rear Sampled grating Reflector (GCSR) tunable The authors are with Australian Photonics Cooperative Research Centre, lasers have achieved tuning times of less than 5 ns between 40 Photonics Research Laboratory, Department of Electrical and Electronic ITU channels [35]. Sampled grating distributed Bragg reflector Engineering, The University of Melbourne, Victoria 3010, Australia (e-mail: m.spencer@ee.mu.oz.au; m.summerfield@ee.mu.oz.au). tunable lasers with a tuning time of approximately 5 ns between Publisher Item Identifier S 0733-8724(00)10886-2. 50 channels [36] and multifrequency lasers with expected 0733–8724/00$10.00 © 2000 IEEE
  • 2. 1658 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 18, NO. 12, DECEMBER 2000 Fig. 1. Generic wavelength-routed passive optical network (WR-PON) architecture. Each node has a single tunable transmitter (TT), and a single fixed receiver (FR). All filtering and optical routing is performed by passive optical elements within the network. tuning times of less than 10 ns between any of 40 channels [22] presented previously in [27]. The present paper extends this pre- have also been reported. vious work, including a more comprehensive description of the Therefore, WDM LANs based on WR-PONs and requiring design and performance issues, more detailed results and com- only tunable transmitters represent a very promising approach to parisons with previously proposed applicable protocols. meeting future high-capacity data network requirements. To re- The structure of this paper is as follows. In Section II, we duce the complexity and cost of the node hardware further, MAC provide an overview of the operation of WRAP, and describe protocols which do not require any additional control channel or the WDM network architectures to which the protocol is appli- carrier-sense capability are highly desirable. cable. Simulations of the protocol demonstrating the impact of In this paper, we consider the design and performance of a different allocation algorithms are described in Section III. In MAC protocol, called simply the WDM Request/Allocation Section IV we present simulation results demonstrating the per- Protocol (WRAP), applicable to WR-PONs with fast-tuning formance of WRAP as a function of the network and protocol transmitters and fixed receivers. Our protocol does not re- parameters, and compare it with alternative applicable proto- quire an additional control channel, performing all necessary cols. signaling in-band. Unlike the previously-proposed FatMAC protocol [24], which also uses in-band signaling, WRAP II. OVERVIEW OF THE PROTOCOL AND NETWORK TOPOLOGIES does not require an optical broadcast facility, and is therefore Fig. 1 illustrates schematically the “generic” wave- applicable to systems using DBR or GCSR tunable semicon- length-routed PON considered in this paper. Each node has ductor lasers which can emit only one wavelength at a time. a single tunable transmitter (TT) and a single fixed optical Furthermore, WRAP delivers similar performance to the ideal receiver (FR). Note that no optical filtering is employed at interleaved time-division multiplexed (I-TDMA) scheme [25] the receivers—all necessary “filtering” is inherently provided under heavy, uniform loading, while offering superior perfor- by the wavelength-routing functionality of the PON. Light mance and capacity sharing under light and/or uneven loads. entering the PON at an input port is routed by passive optical WRAP is robust in the presence of significant propagation components within the network to one and only one output delays in transmission, and in many situations shows improved port. The particular output port selected may depend only performance in the presence of high delay-bandwidth product upon the input wavelength, or it may depend upon both the as a result of inherent pipelining within the network. wavelength and the input port. In either case, a transmitting The initial motivation for the development of WRAP was to node has the ability to select a specific destination node by provide a MAC protocol for use in the MAWSON WDM net- tuning its transmitter to the corresponding wavelength. The work demonstrator currently being developed within the Aus- network has no “master” or “controller” nodes, and allocation tralian Photonics Cooperative Research Centre at the University of network resources is distributed amongst the nodes. The of Melbourne [26]. MAWSON is a ring-based network, however lack of centralized control reduces the system complexity as we discuss in Section II, WRAP is also applicable to other and eliminates the possibility of there being a single critical wavelength-routed topologies such as the AWG-based star. This congestion point, thus making the network simpler to install is in contrast to most previously proposed protocols for WDM and maintain. LANs which are applicable only to star topologies. The basic The MAC protocol described in this paper operates in a design of WRAP, and initial performance simulations have been slotted mode to maximize efficiency. Each slot is divided into
  • 3. SPENCER AND SUMMERFIELD: MAC PROTOCOL FOR WAVELENGTH-ROUTED PONS 1659 Fig. 2. WR-PON queuing model. a number of shorter transmission periods known as minislots, followed immediately by the transmission of the corresponding some of which are used for signaling, and others for data data (e.g., [29]); and those in which the reservations must be transmission, as described below. The protocol is applicable to received by all nodes before implicit slot allocations are made a range of wavelength-routed optical network topologies, the based on a common algorithm (e.g., [11], [24]). WRAP is only requirement being that a mechanism for synchronizing based on a three-stage scheme in which a node wishing to send transmissions at the slot boundaries must be provided. The data first requests an allocation from the destination node. The method of synchronization, and the effective point of synchro- destination receives these requests from all source nodes, and nization, depends upon the details of the physical topology. For then sends back specific allocations to each source (hence the some topologies this may have significant implementation and name WDM Request/Allocation Protocol). Finally, the source performance implications, as we show later when discussing nodes send their data using the allocated slots. the application of WRAP in a ring topology. In this section we The WRAP scheme has a number of advantages over previ- first present an overview of the operation of WRAP, followed ously-proposed protocols: by further details of its implementation in star and ring topolo- • It does not require reservation requests from all source gies. This paper is concerned only with the performance of the nodes to be distributed globally within the network, nor protocol in the simplest case of a single-hop topology in which does it require a centralized scheduler to handle requests. the number of nodes is equal to the number of wavelengths This means that it does not require a shared control , i.e., . Other, more complex, implementations are channel, or a broadcast facility. possible, including multihop topologies, and variations on the • It is inherently fair, under any well-defined fairness crite- protocol in which wavelengths may be shared (i.e., ); rion, since access to any specified destination can be reg- however, these are beyond the scope of the present paper. ulated completely by the allocation algorithm used by the destination node. A. Queuing Model and Protocol Description • With a suitable choice of allocation algorithm, perfor- In the absence of any carrier-sense capability there are two mance can be made equivalent to that of I-TDMA under basic classes of MAC protocol: those that allow collisions such conditions of uniform heavy loading. as ALOHA and slotted ALOHA [28]; and those that avoid all The queuing model of the network is shown in Fig. 2. Each collisions by the careful scheduling of transmissions. Protocols node maintains a set of queues, one for each of the dis- which allow collisions are inherently inefficient and suffer tinct destinations. The tunable transmitter at node can access from poor throughput as a result of the high levels of data loss the set of wavelengths which, in the present work, is as- which occur under moderate to heavy loading [28]. A simple sumed to be sufficient to reach all other nodes in the network. scheme based on the static preallocation of all data transmission The transmitter can only tune to one wavelength in at a slots such as the I-TDMA protocol [25] provides a fair and time. The receiver at node receives signals on the set of wave- optimal use of network resources for uniform network loading lengths but makes no distinction between these wave- and does not suffer any data collisions. Thus it is desirable lengths. If two wavelengths in are active simultaneously, as the uniformly loaded limit for any collision-free protocol. or the same wavelength is active via two sources simultaneously, Dynamic sharing of transmission resources requires the use of a collision occurs at node . The objective of the WRAP protocol some kind of reservation scheme, usually based on a separate is to avoid all such collisions. control channel for exchanging the reservation information. To implement WRAP, time is divided into a series of uni- Most previously proposed reservation-based protocols fall into form length slots, of duration . The slot structure is shown one of two basic categories: those in which the reservations are in Fig. 3(a). The nodes are synchronized so that the start of a
  • 4. 1660 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 18, NO. 12, DECEMBER 2000 N N0 M Fig. 3. WDM Request/Allocation Protocol (WRAP) slot structure and WDM transmission schedule. (a) Slot structure as viewed from a single transmitter or 1 Request/Allocation (R/A) minislots, and data minislots. Transmitted minislots are t t receiver in an -node network consists of a synchronization minislot, separated by a guard time , and the slot period is . (b) Sample multichannel transmission schedule in a three-node network. Each row represents transmissions to a single destination node; the numbers in the R/A minslots represent the source node; allocation of data minislots to source nodes is determined by the WRAP protocol. slot occurs at the same time on all the wavelengths in the net- resynchronization. They may also be the result of small vari- work at the synchronization point. As described later, the syn- ations in propagation delays caused by the effects of dispersion chronization points are at the network hub for a star topology on different wavelengths and by temperature changes and other and at the WDM ADMs for a ring topology. As mentioned pre- environmental fluctuations in the fibers. The framing overhead viously, WRAP is based on a three-stage bandwidth allocation enables nodes to perform clock and data recovery on each in- scheme. In the first stage source nodes with data to transmit send coming minislot. This is needed for correct reception of the data requests to the destination nodes. In the second stage the desti- contained within the minislot. The framing is also needed to nation nodes independently make allocations of data minislots byte-align and delineate the data within each minislot. and send them to the source nodes. In the final stage the source The synchronization minislot is followed by re- nodes transmit the data to the destinations nodes in the allocated quest/allocation (R/A) minislots and data minislots. The data minislots. In the simplest and most efficient implementa- R/A minislots are preallocated on each wavelength in an tion, these three procedures should occur in consecutive slots. It interleaved TDMA manner in such a way that: is thus apparent that if the slot period cannot be too short oth- erwise, for example, requests will not arrive in time for alloca- • each source node is allocated one R/A minislot to each tions to be sent out in the next slot. This implies that, regardless destination node in each slot; and of the topology, there is a minimum slot length. As we shall see, • no source node is required to transmit R/A minislots to the ring topology results in even more severe constraints on the two different destinations at the same time. allowable slot length. The constraints on the slot length are dis- This is illustrated for the simple case of a three-node network cussed in more detail in the following sections. in Fig. 3(b). The length of the data minislots is a selectable pa- As shown in Fig. 3(a), each slot is further divided into a se- rameter of the protocol, and is denoted by (bits). ries of minislots, whose lengths depend on their functions. The Once per slot, a node examines the state of its queues, and de- synchronization minislot is used to maintain slot synchroniza- cides how many data minislots to request, up to the maximum tion between nodes. The length of the synchronization minislot of per destination. Thus, a request requires at most is bits. A practical implementation also requires a guard bits. Each node also examines the requests received from all time, , between minislots and an additional framing overhead other nodes during the previous slot, and makes specific, col- of bits per minislot. The guard time accommodates the laser lision-free allocations of data minislots which are effective for tuning time as well as any timing variations. These timing vari- the next slot. A number of different allocation algorithms are ations may be caused by a drift in the node clocks between possible, as discussed in Section III. Allocations therefore re-
  • 5. SPENCER AND SUMMERFIELD: MAC PROTOCOL FOR WAVELENGTH-ROUTED PONS 1661 quire bits (one per data minislot). The minimum number of bits required in each R/A minislot (excluding overheads) is (1) Finally, each node examines the allocations it received during the previous slot (based on requests it made in the slot prior to that), and transmits data from each of its queues during the data minislot periods allocated to it by the corresponding destination nodes. Note that although there is a minimum three-slot delay be- tween the arrival of data in a queue, and its delivery to the des- tination, under heavier loading (when there is usually data in the queues), the protocol is expected to remain efficient due to the effects of pipelining, in which requests, allocations and data transmissions take place continually over long sequences of contiguous slots. Furthermore, as we will demonstrate, in prac- tical implementations the slot length is often comparable to the unavoidable propagation delays within the network, implying that the built-in protocol delay is of the same order as the prop- agation time. For an -node, -wavelength network, the total theoretical Fig. 4. Four-node wavelength-routed ring network. WDM add–drop data transmission capacity of a slot in bits is given by multiplexers (ADMs) drop one wavelength out of the ring at each access point, while allowing any wavelength to be added. the length of a slot is constrained to be a submultiple of the (2) ring round trip propagation delay, , to ensure simple node synchronization, high utilization and collision free operation The first expression gives an upper bound on the transmission [27]. As described in the previous section, it is necessary for capacity obtained by subtracting the protocol and network over- the nodes to be able to send requests, return allocations and heads for the total slot capacity, while the second expression is transmit data in consecutive slots. This is possible so long as the actual capacity achieved given a particular choice of data 1) the slot period is not less than the ring round trip propagation minislot length . Differences between these two quantities delay, and 2) the largest distance from a node to an access point may arise in topologies in which is constrained, resulting is less than half the distance around the ring. Thus the only in a period of “dead time” at the end of each slot. Equation (2) viable choice for a slot length is , i.e. describes the fixed relationships between slot length, number (3) of nodes, and number of data minislots. There are tradeoffs amongst these parameters, which are investigated in more detail Each node must transmit its minislots at such a time that they through simulations in Section IV. Briefly, protocol overheads arrive at the ADM and are inserted into the ring at exactly are more significant for small slots and/or large and , while the right time (to within the accuracy allowed by the guard the choice of the number of data minislots is a tradeoff between times) to avoid collisions with adjacent minislots on the same the impact of protocol overheads for large and the wasted wavelength. This requires that each node be aware of all the capacity of small if the traffic patterns lead to many partially propagation delays within the network, and maintain an internal filled data minislots. Furthermore, as we discuss later, the imple- timer to control transmissions. Global synchronization of the mentation of an efficient allocation algorithm may place further node timers can be maintained using the synchronization min- constraints on . islot provided at the beginning of each slot. We have developed detailed cold- and warm-start procedures for establishing and B. Ring Topology maintaining synchronization which have been verified by sim- The basic ring topology is illustrated for the case of a ulation, and which will be implemented within the MAWSON four-node network in Fig. 4. It consists of a fiber optic ring with demonstrator. A full description of these procedures is beyond a WDM add/drop multiplexer (ADM) located at each access the scope of the present paper. point. In the simplest configuration each ADM drops a single, The constraint [see (3)] on the slot length in the ring topology fixed wavelength that is different from those dropped at other implies that the maximum available slot capacity is dependent access points. The ADM also allows any one wavelength, or upon the ring length, which is not a dynamically-selectable pa- any number of wavelengths simultaneously, to be added. This is rameter. However, the approximate length can be selected when the architecture of the MAWSON network demonstration being the network is installed, and the ring length may be increased developed by the Australian Photonics Cooperative Research by the addition of spooled fiber if the geographical size is found Centre at the University of Melbourne. In the ring topology, to be too small. Equation (2) shows that the maximum capacity
  • 6. 1662 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 18, NO. 12, DECEMBER 2000 Fig. 5. Four-node wavelength routed star network based on an arrayed waveguide grating (AWG) hub router. Signals are routed uniquely from source to destination based on their wavelength, and the port at which they are input to the AWG. per slot increases linearly with the slot length, and decreases within the accuracy allowed by the guard times) to avoid colli- approximately linearly with the number of nodes and/or data sions with adjacent minislots on different wavelengths, destined minislots. For practical values of the network and protocol over- for the same output port. In this way, transmissions on all wave- heads, a rule-of-thumb which can be derived from (2) is that at lengths are fully synchronized at the hub, as shown in Fig. 3(b). 1 Gb/s data rates, a maximum capacity of greater than 90% (i.e., This requires that each node be aware of the propagation delay 900 Mb/s per wavelength) can be maintained if the length of the between itself and the hub within the network. As in the case ring is greater than m (i.e., 1 km per node). Note that of the ring topology, global synchronization of the node timers the maximum capacity tends toward 100% as the physical net- can be maintained using the synchronization minislot provided work size or the bit-rate increases, the key parameter being the at the beginning of each slot. A technique for determining the data-storage capacity of the ring itself. node-hub propagation delay, and achieving global synchroniza- tion, known as Distributed Slot Synchronization (DSS) has been C. Star Topology developed for use in the CORD network testbed at Stanford, which is based on a broadcast star [4]. This technique may be The basic star topology is illustrated for a four-node network readily adapted for use in the wavelength-routed star topology, in Fig. 5. It consists of a passive optical wavelength-routing hub although a full description of the scheme is beyond the scope of (which may be implemented using a single AWG, or equiva- the present paper. lent structure) to which the nodes are connected by fiber pairs. Once again it is necessary for the nodes to be able to send re- Light entering the hub on any given wavelength and input fiber quests, return allocations and transmit data in consecutive slots. is routed to a specific corresponding output port. The routing For the star topology this implies that the slot length must be function is such that light on the same wavelength but arriving more than twice the longest propagation delay between the hub from different input ports must always be routed to different and a node. Other than this requirement, is a freely-se- output ports. Under this condition, the routing function is unique lectable parameter of the protocol, along with the number of data and introduces no optical loss in principle, although any prac- minislots, . These parameters may be established by negoti- tical implementation of the passive hub will exhibit some finite ation among the nodes in the network. Analysis and simulation insertion loss. (For AWG-type devices, the insertion loss is typ- of the network performance is therefore important to assist in ically 3–6 dB, and is independent of the number of ports [30].) the development of formulae and/or heuristics that enable these Fig. 5 illustrates the specific routing properties of a four-port parameters to be optimized during network initialization. AWG. Entering the hub from Node 1, light of wavelength is routed back to Node 1; wavelength is routed to Node 2; III. INVESTIGATION OF ALLOCATION ALGORITHMS and so on. However, wavelength is routed to Node 2 when entering the hub from Node 2; to Node 3 when entering from A node must consider the amount of transmission bandwidth Node 3; and so on. The routing permutations follow the same requested by other nodes once per slot and make appropriate pattern for the other wavelengths. allocations to the requesting nodes. The goal of the algorithms With a star topology a central synchronization point at the considered here is to make this allocation fair, in the sense that, hub of the star is used. Each node transmits its minislots at such all else being equal, a receiving node will on average allocate a time that they arrive at the hub at exactly the right time (to bandwidth to requesting nodes in proportion to their demands,
  • 7. SPENCER AND SUMMERFIELD: MAC PROTOCOL FOR WAVELENGTH-ROUTED PONS 1663 while providing some guaranteed bandwidth to all nodes. The TABLE I node must also consider the independent allocations received SYMMETRIC TRAFFIC LOAD DISTRIBUTION FOR SIMULATIONS OF A THREE-NODE NETWORK from other nodes, as they may clash, and make a choice of which one to use for each data minislot. Three allocation algorithms have been developed and have been compared using simulations of a 3-node symmetrically loaded network. The simulations were performed using event driven simu- lation code written in C. A slot period of 20 s, a data rate of 1 Gb/s, and four data minislots per slot were assumed. Finite buffers of 40 data minislots per node per destination were used. Realistic protocol overheads were accounted for in the simulation as follows. A conservative guard-time of ns was assumed, which allows for the relative drift of node clocks between synchronization events, as well as differences in propagation delays on different wave- access to the network capacity. Furthermore, under symmetric lengths due to fiber dispersion and environmental fluctuations. heavy loading, the cyclic allocation algorithm may result in re- The guard time thus contributes 100 bits overhead per minislot. ceiving nodes making allocations which are identical to the slot The total of all framing and protocol overheads, each of which allocations in the I-TDMA protocol, which is known to be op- is rounded to the next 8-bit multiple (i.e., assuming a byte timal under these loading conditions. is the minimum transmission unit), is 996 bits per slot. This Fig. 6(a)–(d) shows the distribution of delays experienced by comprises 124 bits for the synchronization minislot (100 for the data arriving at the source nodes obtained from a number of guard time, 16 for the framing and 8 for the protocol as a slot simulation runs at total offered loads of 50%, 65%, 80%, and identifier), 156 bits for each R/A minislot (100 for the guard 125%, respectively. At offered loads less than the network ca- time, 16 for framing and 40 for protocol, including 8 each for pacity most data is queued only as a result of the protocol delay, the requests and allocations) and 140 bits for each data minislot thus resulting in short queue lengths, as shown in Fig. 6(a). The (100 for the guard time, 16 for framing and 24 for protocol variability of the data arrivals sometimes results in more instan- overheads). Inclusion of the overheads results in an effective taneous load than the network can handle, and this results in the reduction of capacity, as described by (2). Data arrivals to the excess being queued. As the offered load is increased the queue network were modeled as packets of uniformly distributed lengths increase, as shown in Fig. 6(b) but they are limited by length of 0 to 16 000 bits arriving under a Poisson process. the buffer capacity. When the buffers are full, further data ar- The network loading was varied from 50 to 150% by changing rivals result in loss. Thus, under overloaded conditions the queue the mean inter-arrival time appropriately. Network overload lengths approach the buffer capacity. The average rate at which (greater than 100% loading) was considered to enable us to the queues are served (i.e., the average number of data minis- investigate the limiting behavior of the protocol, however we lots transmitted per second) is proportional to the throughput. recognize that real networks are unlikely to be operated in this Data arriving at the queue must wait for all the data already in regime due to the substantial data losses incurred. Simulation the queue to be served before it will be sent. Thus the mean runs of 100 000 slot periods (2 s) were repeated eight times at queuing delay experienced by data is proportional to the mean each of these values of loading. queue length and inversely proportional to the mean throughput. The network was loaded symmetrically as shown in Table I, Also, the delay is bounded, as the buffers are of finite length with 50% of the traffic at each node sent to and received from the and there is a nonzero worst case network throughput. Protocol other two nodes. Unity (100%) loading of a source/destination and transmission delays are incurred in addition to the queuing route corresponds to a mean arrival rate equal to the raw capacity delay, however pipelining results in the removal of most of the of a wavelength neglecting overheads, i.e., 1 Gb/s. protocol delay for heavy loading conditions. Thus for the net- work conditions simulated the buffer size of 40 minislots and worst case throughput of one minislot per slot bound the delay A. Cyclic Allocation to just above 40 slots. This bound on the delay is clearly visible In the cyclic allocation algorithm, each receiving node con- in Fig. 6(b) and (c). siders the requests from transmitting nodes in a round robin The cyclic allocation algorithm is simple, and allows each fashion [27]. Individual data minislots are allocated sequen- node to make allocations independently of its location in the tially, to nodes that have made requests until either there are network. However, Fig. 6(c) and (d) demonstrate an undesir- no more requests or no more unallocated minislots. This cyclic able property of the algorithm which only became apparent round-robin allocation continues in the next slot, recommencing during these simulation runs. The graphs clearly show the from the node following the last one considered in the previous development of a bimodal delay distribution under conditions slot. When a transmitting node receives more than one alloca- of moderate-to-heavy loading. The observed results are due tion for the same data minislot from different destination nodes, to an undesirable synchronization and lock-up property of the conflict is resolved in a similar way, i.e., by cyclic allocation the cyclic allocation algorithm. Fig. 6(d) shows that under of such minislots in a round-robin fashion. When executed iden- very heavy load, there is a tendency for the protocol to tically at every node, these algorithms result in fair and flexible fall into one of two possible states, indicated by the gray
  • 8. 1664 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 18, NO. 12, DECEMBER 2000 (a) (b) (c) (d) Fig. 6. Distribution of packet delays for the cyclic allocation algorithm in a three-node network under symmetric loading, for average network loads of (a) 50%, (b) 65%, (c) 80%, and (d) 125%. histogram (State 1) and the black histogram (State 2) respec- Fig. 6(c), the protocol falls in and out of the lock-up states in tively. This behavior may be explained as follows. As the response to short-term peaks in demand, and the observed delay network loading increases, the amount of data in the queues distribution is the time-averaged effect of the protocol spending will tend to exceed the maximum capacity of a slot. This periods of time in both states. will result in more outstanding requests every slot than there Fig. 7 shows the average delay and packet loss under the are available data minislots, as a result of which the source cyclic allocation algorithm as a function of the mean network nodes will consistently make identical allocations to the des- loading. The key feature of these results is the “knee” in the tination nodes. The cyclic round robin allocation method is delay characteristic, which occurs at an average load of about completely deterministic, and generates the same allocations 60%. This is preceded by a dramatic rise in the packet loss. The for as long as this overloaded state exists. State 1, which minimum delay under light loads is 110 s (5.5 slots). The delay has lower mean delay, corresponds to the fortuitous circum- results under heavy loading conditions have been averaged over stance in which the allocations received back at each source the length of the simulation runs, and they represent the average node from the two destination nodes do not clash, i.e., both effect of both states. Delay in State 1 limits to 19.1 slots, while destinations have allocated different data minislots, and the State 2 has an average delay of 30.4 slots (a result of the bimodal transmitter is able to use all allocations. State 2, with higher distribution with peaks near 20 and 40 slots delay). mean delay, corresponds to a situation in which many of the allocations clash, and the source node is unable to use both B. Random Allocation since it can only transmit to one destination at a time. In the One strategy for avoiding the undesirable synchronization of 3-node network, this state corresponds to a drop in available the allocations which occurs with the cyclic algorithm is to allo- capacity to 2/3 of the maximum capacity ( in the case cate the minislots randomly instead of sequentially. Similarly, if of an -node network). It therefore corresponds to a form of more than one allocation is received at the source node from two congestion collapse for the cyclic allocation algorithm, and (or more) destination nodes, then a random selection is made. must be avoided. This algorithm is simple, it obviously avoids any deterministic Under overloaded conditions, the protocol tends to fall into lock-up states, and it retains the property that allocations can one state or the other and stay there, and the two histograms in be made independently of a node’s location within the network. Fig. 6(d) show the results of separate simulation runs in which However, it clearly does not limit to the ideal I-TDMA schedule each state was observed. Under moderate loading, shown in under uniform heavy loading.
  • 9. SPENCER AND SUMMERFIELD: MAC PROTOCOL FOR WAVELENGTH-ROUTED PONS 1665 Fig. 7. Average delay and packet loss as a function of offered load for the cyclic allocation algorithm in a three-node network under symmetric loading. (a) (b) (c) (d) Fig. 8. Distribution of packet delays for the random allocation algorithm in a three-node network under symmetric loading, for average network loads of (a) 50%, (b) 65%, (c) 80%, and (d) 125%. Fig. 8(a)–(d) shows the delay distributions for the random al- lots per slot) as a result of the random allocation clashes and thus location algorithm obtained from a number of simulation runs the delay limits to an average value of 520 s (26 slots—slightly at total offered loads of 50%, 65%, 80%, and 125%, respec- less than the theoretical 26-2/3 as the buffers never fill com- tively. The queues evolve in a similar manner to the cyclic allo- pletely). The evolution to this condition is clearly visible in cation algorithm, however in this case the bimodal distribution Fig. 8(c) and (d). Fig. 9 shows average delay and packet loss of does not emerge. As a result, the upper tail of the distribution the random allocation algorithm as a function of mean network at 65% loading shown in Fig. 8(b) is less heavy than the corre- loading. The knee of the delay curve occurs at a slightly higher sponding distribution for the cyclic allocation algorithm. Under loading of about 65% compared with the cyclic allocation algo- heavy loading the peak average capacity is 3/4 (i.e., 1-1/2 minis- rithm, and there is a corresponding improvement in packet loss.
  • 10. 1666 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 18, NO. 12, DECEMBER 2000 Fig. 9. Average delay and packet loss as a function of offered load for the random allocation algorithm in a three-node network under symmetric loading. C. Preferential/Random Allocation minslot will be preferentially allocated. For example, Node 0 will preferentially allocate the first data minislot to Node The cyclic allocation algorithm was designed to converge to 3, the second to Node 2, the third to Node 1, and so on. the I-TDMA schedule under uniform heavy loading, but was The set of preferential allocations actually made is shown in found to suffer from undesirable lock-up states. The random al- Fig. 10(c)—these allocations are deterministic, and are derived location algorithm avoids these lock-up states, resulting in im- directly from Fig. 10(a) and (b). Each node then makes random proved average performance, but still underutilizes the available allocations of any unallocated data minislots to nodes with bandwidth under heavy loading. The ideal allocation algorithm additional requests. One such random allocation is shown in would avoid the lock-up states and converge to the I-TDMA Fig. 10(d). At Node 0, Data Minislots 0 and 3 are not needed schedule under uniform heavy loading. We have developed such for preferential allocations, and in this example, Data Minislot an algorithm, which we call the preferential/random (P/R) algo- 0 has been randomly allocated to Node 2. Likewise, at Node rithm. Under this scheme the fixed I-TDMA schedule is consid- 1, Node 0 has been randomly allocated two minislots chosen ered to be the preferred allocation at the receiving nodes. When from Data Minislots 2, 4, and 5. Fig. 10(e) summarizes the requests arrive from the source nodes, allocations are made ini- allocations received by each node, with a cross representing tially from the I-TDMA schedule. Once this has been done, if an allocation. An example of the allocations which may there are further requests from some source nodes and there are subsequently be used, after resolving any clashes, is given in remaining free data minislots, then these are allocated by con- Fig. 10(f). At Node 0, Data Minislot 1, the random allocation sidering the requests from each source node in a round-robin from Node 3 clashes with the preferential allocation from Node manner (to ensure that the allocation of the remaining capacity 2, thus the preferential allocation is used. The clash which is fair on average), and assigning to each source one of the re- occurs at Node 0, Data Minislot 4, is between two random maining data minislots chosen at random (to avoid any deter- allocations (from Nodes 1 and 3). One of these is chosen at ministic lock-up effects). Similarly, if a source node receives al- random—in this example it is the allocation from Node 1. locations for the same data minislot from two (or more) different Similar choices apply to all the other allocation clashes present. destinations, if one of these is the preferred I-TDMA allocation, The main disadvantage of the P/R algorithm is that it requires that one is used. If all are the result of random allocations, one each node to “know” its place in the I-TDMA schedule, and thus is chosen at random. it is not independent of the node’s location in the network. In An example of the P/R allocation scheme for a 4-node practice, the labeling scheme used for nodes and I-TDMA slots network with six data minislots is given in is arbitrary (i.e., the physical ordering of nodes in the network is Fig. 10. Fig. 10(a) shows a table of all the requests made for the not significant), and can easily be assigned as part of the cold- slot under consideration. The numbers in the table indicate the and warm-start procedures which are required in any case. For number of data minislots requested for transmission by each the preferred I-TDMA schedule to map fairly into each slot, the source node to each destination node. For example, Node 0 is additional constraint is imposed that , where requesting four minislots to Node 1, one minislot to Node 2, is any positive integer. and five minislots to Node 3. With , if Node 0 receives The effect of the P/R algorithm is that each node is guaran- allocations for all 10 requests, it will obviously be able to teed at least of its transmission capacity to every use six at most, however in the absence of any information other node. A node may be able to obtain up to the entire avail- regarding the requests made by other nodes, the R/A protocol able capacity, depending on the requests made by other nodes. considers all requests independently. The preferential source The access to this capacity is shared in a statistically fair way nodes for the data minislots at each of the destination nodes among those requesting it. Under heavy uniform loading the net- are shown in Fig. 10(b), in which the number in each box work limits to the optimal fixed TDMA schedule and does not represents the node number to which the corresponding data suffer the instabilities of the cyclic algorithm. It can be shown,
  • 11. SPENCER AND SUMMERFIELD: MAC PROTOCOL FOR WAVELENGTH-ROUTED PONS 1667 N M Fig. 10. Example of the P/R allocation scheme for a 4-node ( = 4) network with six data minislots ( = 6). (a) Requests made to the nodes. (b) Preferential sources for the data minislots at each node. (c) Preferential allocations made. (d) Possible random allocations made. (e) Summary of the allocations received by each node. (f) Actual allocations used. although space does not permit its inclusion in this paper, that Simulation of the P/R allocation algorithm resulted in the all three allocation algorithms produce the same number of allo- delay distributions shown in Fig. 11(a)–(d). The guaranteed ca- cations, the difference between them being in the data minislots pacity of two minislots per slot and the buffer size of 40 min- chosen. Furthermore, allocation clash resolution by the P/R al- islots per node per destination ensure that the worst case delay gorithm can be shown to result in a mean performance that is is limited to just above 20 slots. The protocol delay is always always better than the random algorithm’s, and that under cer- present, but is masked by pipelining at high loading. tain loads (including uniform) the P/R algorithm achieves the Fig. 12 shows the average delay and packet loss under the maximum possible minislot usage. P/R allocation algorithm as a function of the mean network
  • 12. 1668 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 18, NO. 12, DECEMBER 2000 (a) (b) (c) (d) Fig. 11. Distribution of packet delays for the P/R allocation algorithm in a three-node network under symmetric loading, for average network loads of (a) 50%, (b) 65%, (c) 80%, and (d) 125%. Fig. 12. Average delay and packet loss as a function of offered load for the P/R allocation algorithm in a three-node network under symmetric loading. loading. The maximum average delay under overload conditions considered. Therefore, in the remainder of the paper only the is 360 s (18 slots), comparable to the desirable behavior of the P/R algorithm is considered. cyclic scheme. The knee of the delay characteristic occurs at a mean network loading level of 80%. At high loading there is full IV. PERFORMANCE SIMULATIONS utilization of the available network capacity under the I-TDMA schedule. A. Scaling Performance of the Protocol The results in this section illustrate the clear superiority of By choosing a value for the slot period in conjunction the P/R allocation algorithm. It is stable under heavy loading with the number of nodes a tradeoff can be made between and overload conditions, and exhibits the lowest delay, delay the minimum delay at low loads, and the cost of the per-slot variance and highest average throughput of the three algorithms overheads. As discussed previously, in practice, is a free
  • 13. SPENCER AND SUMMERFIELD: MAC PROTOCOL FOR WAVELENGTH-ROUTED PONS 1669 Fig. 13. Average packet delay for the WRAP protocol under symmetric loading, as a function of total offered network load, for a channel capacity of B = 1 Gb/s per wavelength and a fixed slot duration of 15 s. Results are shown for three-node (squares), five-node (diamonds), seven-node (triangles), and nine-node (circles) networks. Fig. 14. Total throughput for the WRAP protocol under symmetric loading, as a function of total offered network load, for a channel capacity of B = 1 Gb/s per wavelength and a fixed slot duration of 15 s. Results are shown for three-node (squares), five-node (diamonds), seven-node (triangles), and nine-node (circles) networks. parameter in star networks but may be constrained by the the overall network capacity grows as but the number of fiber length in ring networks. Simulations were carried out source/destination pair routes grows as . These obser- for , 5, 7, and 9 and , 25, 35, and 45 s vations are confirmed in the throughput results in Fig. 14, which with symmetric loading to provide insight into this tradeoff. show clearly that the fractional utilization of the raw bandwidth Symmetric loading was used as it fully utilizes the network is reduced for higher numbers of nodes. In this sense, the uni- capacity under high loads, and is therefore a case of particular form symmetric loading example represents a worst-case for the interest in which the results are relatively easy to interpret. protocol. The number of minislots was chosen to be The relative effect of the increasing overheads with increasing and the total buffer capacity was fixed at 1 Mbyte per node, number of nodes can be mitigated by making corresponding divided equally amongst all possible destinations. All network changes to the slot period , at the expense of some addi- and protocol overheads were again included, using the same tional delay under light loading. The effect of changing both the parameters discussed previously. slot length and the number of nodes, while keeping them propor- In the first set of simulations the number of nodes was in- tional was simulated. The delay as a function of the total offered creased while keeping constant at 15 s. The average delay load to the network is shown in Fig. 15, and the corresponding as a function of the total offered load to the network is shown throughput is shown in Fig. 16. The overheads remain at about in Fig. 13, and the corresponding total throughput is shown in the same fraction of the network capacity and as a result the Fig. 14. The worsening delay performance, with the knee of the knees of the delay curves occur at roughly the same fractional curve occurring at a relatively small fraction of the total net- loading in all cases. Similarly, the fractional utilization under work capacity as the number of nodes increases, is due mainly overload conditions is approximately the same in all cases. The to the effect of increasing overheads. The increase in the max- slight increase in delay under light loading is difficult to see in imum delay under overload is due to the reduced capacity be- Fig. 15, as it is small compared to the queuing delays which tween each source/destination pair, resulting from the fact that occur under moderate to heavy loads.
  • 14. 1670 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 18, NO. 12, DECEMBER 2000 Fig. 15. Average packet delay for the WRAP protocol under symmetric loading, as a function of total offered network load, for a channel capacity of B = 1 Gb/s per wavelength and a slot duration which is increased in proportion to the number of nodes. Results are shown for three-nodes and t = 15 s (squares), five-nodes and t = 25 s (diamonds), seven-nodes and t = 35 s (triangles), and nine-nodes and t = 45 s (circles). Fig. 16. Total throughput for the WRAP protocol under symmetric loading, as a function of total offered network load, for a channel capacity of B = 1 Gb/s per wavelength and a slot duration which is increased in proportion to the number of nodes. Results are shown for three-nodes and t = 15 s (squares), five-nodes and t = 25 s (diamonds), seven-nodes and t = 35 s (triangles), and nine-nodes and t = 45 s (circles). We conclude that the WRAP protocol scales reasonably well topology. In addition, FatMAC has been specifically optimized with increasing numbers of nodes, so long as the slot period for multiprocessor interconnection, and as such tends to trade is similarly scaled. This scaling property is the result of the off bandwidth efficiency in order to obtain lower minimum fact that each node brings with it additional capacity, in the delay. The FatMAC protocol is a time-slotted protocol with form of an added wavelength. The additional delay caused by two phases. In the first, control slots containing the destination the longer slot period under lightly loaded conditions must be of the packet at the head of the queue are broadcast by each considered. If the geographical size of the network increases as node. In the second phase the packet from the head of each nodes are added then the minimum delay remains proportional node’s queue is sent, with the order of transmissions to a to the inherent propagation delays. Since propagation delays are destination determined by the number of control slots with that unavoidable, the penalty due to the protocol delays in such a destination’s address in the previous phase. This two phase case is minimal. This additional delay for light loads is small procedure is then repeated. A simulation comparison between in comparison with queuing delays and it is deterministic, and WRAP, I-TDMA, and FatMAC has been carried out in order to thus unlikely to adversely affect multimedia traffic. The effect assess the relative performance of our protocol. of this delay on applications using the network is an important In the first set of simulations, all protocols were simulated topic, but is beyond the scope of the present paper. for the symmetrically-loaded 3-node case, and for loads ranging from 10 to 150%. In order to make the comparisons as fair as B. Comparison with Other Applicable Protocols possible, the I-TDMA and FatMAC slot lengths were set equal to the WRAP minislot length. To accommodate FatMAC, the Two previously-proposed MAC protocols which could be network topology is assumed to be a star, the optical transmit- applied to WR-PONs of the type considered here are I-TDMA ters are assumed to have a WDM broadcast facility, and propa- [25] and FatMAC [24]. The former protocol is applicable to gation delays in the network are assumed to be negligible, i.e., both stars and rings, while the latter is only practical in a star less than the transmission time of the smallest data or protocol
  • 15. SPENCER AND SUMMERFIELD: MAC PROTOCOL FOR WAVELENGTH-ROUTED PONS 1671 (a) (b) Fig. 17. (a) Comparison of the average packet delay for the WRAP (squares), I-TDMA (diamonds), and FatMAC (triangles) protocols for three-node networks under symmetric loading, as a function of the offered load. (b) Corresponding comparison of packet loss. unit. While both WRAP and I-TDMA can accommodate prop- TABLE II ASYMMETRIC TRAFFIC LOAD DISTRIBUTIONS FOR SIMULATIONS OF AN agation delays within their slot structures, FatMAC cannot tol- EIGHT-NODE NETWORK erate propagation delays which exceed its small control slot du- ration. The delay and loss of these protocols as a function of the fractional offered load to the network are shown in Fig. 17(a) and (b) respectively. As expected, FatMAC gives low delays under light load, but has poor performance as the loading is in- creased. I-TDMA has similar but better performance to WRAP for the symmetric loading, as it does not incur as much protocol delay and has lower protocol overheads. While I-TDMA is op- timal for symmetric loading with small traffic arrival variations, it is not flexible, in that the maximum available capacity for any source/destination route is limited to of that obtain- able under WRAP. To illustrate this, an eight-node network with traffic loads as shown in Table II was simulated. Equivalent slot/minislot lengths were used for the different protocols, and buffer size per source/destination route chosen to be 40 times the capacity of a slot/minislot. Loading from 10 to 150% of the tabulated load was applied, this factor being called the this throughput for an example node, Node 4, into the traffic “Traffic Matrix Scaling Factor” in the figures. The nodes sent to the four destinations making up the total (i.e., Nodes have been numbered 0 through 7. For each protocol we have 0, 3, 5, and 6); and 3) the corresponding average delays plotted: 1) the throughput per node, measured as the total experienced by traffic sent from Node 4 to each of the four output to all other nodes in the network; 2) a breakdown of destinations. Fig. 18 shows the throughput per node for a)
  • 16. 1672 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 18, NO. 12, DECEMBER 2000 (a) (b) (c) Fig. 18. Total throughput achieved by each node in an eight-node network under the asymmetric traffic pattern shown in Table II as a function of the factor by which the traffic matrix is scaled. Simulation results are shown for (a) I-TDMA, (b) FatMAC, and (c) WRAP. I-TDMA, b) FatMAC, and c) WRAP; Fig. 19(a)–(c) shows tination capacity is sought. Only Node 0 is able to obtain all the breakdown for traffic sent from node 4; and Fig. 20(a)–(c) of the desired throughput as the loading is increased, as indi- shows the corresponding delays. cated by the linear growth of throughput as a function of loading It can be seen from Fig. 18(a), Fig. 19(a), and Fig. 20(a) that in Fig. 18(a). This is because the total output from Node 0 is the fixed bandwidth allocations of I-TDMA lead to low limiting uniformly distributed to all seven remote nodes, a circumstance throughput and high delays if more than the available per-des- which is perfectly accommodated by the I-TDMA schedule. The
  • 17. SPENCER AND SUMMERFIELD: MAC PROTOCOL FOR WAVELENGTH-ROUTED PONS 1673 (a) (b) (c) Fig. 19. Breakdown of the throughput achieved by Node 4 to each of the four destination nodes, Node 0, Node 3, Node 5, and Node 6, corresponding to the total throughput shown in Fig. 18. Simulation results are shown for (a) I-TDMA, (b) FatMAC, and (c) WRAP. throughput of all other nodes is severely limited due to exhaus- exhausts the capacity of the slots corresponding to those desti- tion of the capacity of the per-destination slots in the transmis- nations in the schedule. The onset of this capacity exhaustion is sion schedule. This is demonstrated by the flattening out of the accompanied by the large increases in queuing delays observed throughput curves in Fig. 18(a). For the particular case of Node in Fig. 20(a). 4, shown in Fig. 19(a) and 20(a), the small proportion of traffic Fig. 18(b) shows that FatMAC obtains better overall destined to Nodes 0 and 3 is able to be transmitted. However the throughput than I-TDMA. Fig. 20 also shows that it has the majority of traffic, which is destined for Nodes 5 and 6, rapidly lowest delays of all three protocols under light loading and
  • 18. 1674 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 18, NO. 12, DECEMBER 2000 (a) (b) (c) Fig. 20. Breakdown of the average delay experienced by packets sent from Node 4 to each of the four destination nodes, Node 0, Node 3, Node 5, and Node 6, corresponding to the throughput per destination shown in Fig. 19. Simulation results are shown for (a) I-TDMA, (b) FatMAC, and (c) WRAP. overload conditions. However, Fig. 19(b) demonstrates that to Nodes 0, 3, and 6 could be improved, at the expense of a although Node 4 obtains higher overall throughput under reduction in throughput to Node 5. An additional problem FatMAC than I-TDMA, this is the result of higher throughput with FatMAC is that it suffers from head-of-line blocking as it to Node 5, with traffic to Nodes 0, 3, and 6 fairing somewhat only uses one queue per node. The loading levels at which the worse. This is the result of a fairness problem in FatMAC throughput to each destination begins to limit is more clearly which we have not addressed, although a possible solution observed from the locations of the knees in the delay curves of is discussed in [24]. If this problem were solved, throughput Fig. 20(b).
  • 19. SPENCER AND SUMMERFIELD: MAC PROTOCOL FOR WAVELENGTH-ROUTED PONS 1675 Fig. 18(c) shows that the WRAP protocol achieves the highest the lowest delay and delay variance, and the highest average maximum throughput for all nodes (note that the vertical scale is throughput of the three algorithms considered. twice that of the other two graphs in the Figure). Furthermore, it The scaling behavior of the protocol for increasing number shares the available capacity most efficiently amongst the nodes, of nodes was investigated by simulation, and found to be ac- as evidenced by the fact that the throughput increases linearly ceptable provided that the slot length is increased in proportion with offered load for all nodes to a scaling factor of over 50%. the number of nodes. This maintains the same relative protocol When the throughput does level out, it is due to the fact that overheads but does so at the expense of an increase in the in- the total offered traffic from all sources to a particular desti- trinsic protocol delay. nation exceeds the total capacity available to that destination. Simulations have shown that WRAP provides both fair For example, the total traffic offered to the network destined for and flexible access to the network. It allows bandwidth to be Node 6 is the highest, so nodes transmitting to that node (i.e., allocated according to demand, while providing a guaranteed Nodes 4 and 7) are the first to be affected. Although Node 0 is minimum bandwidth allocation between each source-destina- also transmitting to Node 6, it is not affected in this example, as tion pair. Comparisons were performed with two alternative its demand never exceeds the guaranteed minimum provided by protocols, I-TDMA and FatMAC. For symmetric network the preferential I-TDMA allocations. This is further illustrated loading I-TDMA performs optimally, and has no additional in Fig. 19(c) which shows that the throughput to Node 6 is the protocol overheads. Under these conditions, WRAP limits first to level out, at a scaling factor of about 60%. As the total to the I-TDMA behavior under heavy loads, but has higher network load is increased, throughput to Node 6 from Node 4 overheads, and larger intrinsic protocol delays. FatMAC has actually declines, due to increased traffic demands from other the lowest intrinsic delay, however this is obtained at the source nodes, however it can never fall below the guaranteed expense of network utilization. Under asymmetric traffic minimum. Once again, the knees in the delay curves shown in conditions WRAP provides the best throughput performance Fig. 20(c) are the clearest indicator of the loading level at which in comparison to FatMAC and I-TDMA, offering the I-TDMA the throughput reaches a maximum. capacity as a guaranteed minimum, and maintaining low delays Based on the simulation results in this section, we conclude for comparatively higher offered loads. We thus conclude that that I-TDMA is only appropriate in an environment where the of the three protocols considered here, WRAP is the best-suited total traffic load on the network is distributed evenly amongst to general-purpose data communications applications such as all source-destination pairs. This is unsurprising, as it utilizes local, campus and metropolitan area networks. a fixed TDMA schedule. FatMAC exhibits the lowest delays at low network loading, and it is therefore most appropriate for ap- ACKNOWLEDGMENT plications such as the multiprocessor interconnect for which it was designed, where low latency is more important than high The authors gratefully acknowledge M. Roughan from the utilization. The WRAP protocol represents a compromise in Department of Electrical and Electronic Engineering at the Uni- which latency at low network loads is traded off for higher band- versity of Melbourne, for helpful discussions and comments on width utilization and greater flexibility, while retaining the max- this work. imum throughput of I-TDMA as a minimum guaranteed perfor- mance. REFERENCES [1] V. W. S. Chan, K. L. Hall, E. Modiano, and K. A. Rauschenbach, “Ar- chitectures and technologies for high-speed optical data networks,” J. V. CONCLUSION Lightwave Technol., vol. 16, pp. 2146–2168, Dec. 1998. [2] J. M. Senior, M. R. Handley, and M. S. Leeson, “Developments in wave- This paper has presented WRAP—a collision-free media ac- length division multiple access networking,” IEEE Commun. Mag., pp. 28–36, Dec. 1998. cess protocol for wavelength routed passive optical networks. [3] L. G. Kazovsky, T. Fong, and T. Hofmeister, “Optical local area network We have discussed the application of this protocol to ring and technologies,” IEEE Commun. Mag., pp. 50–54, Dec. 1994. star topologies. Each node requires a single fixed receiver and [4] R. T. Hofmeister, L. G. Kazovsky, C. L. Lu, P. Poggliolini, and F. Yang, “CORD: Optical packet-switched network testbed,” Fiber Integr. Opt., a single wavelength tunable transmitter. WRAP is a slotted pro- vol. 16, pp. 199–219, 1997. tocol that utilizes in-band signaling to send requests to other [5] M. A. Marsan, A. Bianco, E. Leonardi, M. Meo, and F. Neri, “MAC pro- nodes and to return the allocations made independently by them. tocols and fairness control in WDM multirings with tunable transmitters and fixed receivers,” J. Lightwave Technol., vol. 14, pp. 1230–1243, June As a consequence the protocol does not require any additional 1996. control wavelength or network, nor does it require a centralized [6] “Special issue on multiwavelength technology and networks,” J. Light- controller. Several possible allocation algorithms have been de- wave Technol., vol. 14, pp. PAGE NOS?–, June 1996. [7] M. S. Goodman, H. Kobrinski, M. P. Vecchi, R. M. Bulley, and J. L. scribed and their performance compared using computer simu- Gimlett, “The LAMBDANET multiwavelength network: Architecture, lation. It was found that a cyclic round-robin algorithm exhibits applications and demonstrations,” IEEE J. Sel. Areas Commun., vol. 8, lock-up states under heavy loads, leading to instability and a pp. 995–1003, Aug. 1990. [8] T. K. Chiang, S. K. Agrawal, D. T. Mayweather, D. Sadot, C. F. Barry, M. form of congestion collapse. The lock-up states can be avoided Hickey, and L. G. Kazovsky, “Implementation of STARNET: A WDM by using a random allocation algorithm, however this was not computer communications network,” IEEE J. Sel. Areas Commun., vol. found to result in a significant improvements in utilization. A 14, pp. 824–839, June 1996. [9] N. R. Dono, P. E. Green Jr., K. Liu, R. Ramaswami, and F. F.-K. Tong, hybrid scheme, which we have called the preferential/random “A wavelength division multiple access network for computer commu- allocation algorithm, was developed. This algorithm exhibits nication,” IEEE J. Sel. Areas Commun., vol. 8, pp. 983–993, Aug. 1990.
  • 20. 1676 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 18, NO. 12, DECEMBER 2000 [10] I. Chlamtac, A. Fumagalli, L. G. Kazovsky, and P. T. Poggiolini, “A [25] K. Bogineni, K. M. Sivalingam, and P. Dowd, “Low-complexity mul- contention/collision free WDM ring network for multi gigabit packet tiple access protocols for wavelength-division multiplexed photonic net- switched communication,” J. High Speed Networks, vol. 4, pp. 201–219, works,” IEEE J. Select. Areas Commun., vol. 11, pp. 590–604, May 1995. 1990. [11] D. A. Levine and F. Akyildiz, “PROTON: A media access control pro- [26] M. A Summerfield, “Wavelength division multiplexing for local area tocol for optical networks with star topology,” IEEE/ACM Trans. Net- networks,” in Proc. CLEO/Pacific Rim ’99, Seoul, Korea, August working, vol. 3, no. 2, pp. 158–167, April 1995. 30–September 3, 1999, Paper FN4, pp. 1215–1216. [12] M. S. Borella and B. Mukherjee, “Efficient scheduling of nonuniform [27] J. Fransson, M. Johansson, M. Roughan, L. Andrew, and M. A. packet traffic in a WDM/TDM local lightwave network with arbitrary Summerfield, “Design of a medium access control protocol for transceiver tuning latencies,” IEEE J. Select. Areas Commun., vol. 14, a WDMA/TDMA photonic ring network,” in Proc. IEEE Global pp. 923–934, June 1996. Telecommunications Conf. (GLOBECOM’98), Sydney, Australia, [13] H. Choi, H. A. Choi, and M. Azizoglu, “Efficient scheduling of trans- November 8–12, 1998, Paper S89.6. mission in optical broadcast networks,” IEEE/ACM Trans. Networking, [28] L. Kleinrock, Queuing Systems, Volume I: Theory. New York: Wiley, vol. 4, pp. 913–920, Dec. 1996. 1976. [14] S. Trindandapani, J. S. Meditch, and A. K. Somani, “The MaTPi pro- [29] M.-S. Chen, N. R. Dono, and R. Ramaswami, “A media-access protocol tocol: Masking tuning times through pipelining in WDM optical net- for packet-switched wavelength division multiaccess metropolitan area works,” in Proc. INFOCOM ’94, June 1994, pp. 1528–1535. networks,” IEEE J. Select. Areas Commun., vol. 8, pp. 1048–1057, Aug. [15] N. F. Maxemchuk, “Twelve random access strategies for fiber optic net- 1990. works,” IEEE Trans. Commun., vol. 36, pp. 942–950, Aug. 1988. [30] O. Katsunari, “Planar waveguide devices for WDM communications,” [16] S. L. Woodward, P. P. Iannone, K. C. Reichmann, and N. J. Frigo, “A in Proc. CLEO/Pacific Rim ’99, Seoul, Korea, August 30–September 3, spectrally sliced PON employing Fabry–Perot lasers,” IEEE Photon. 1999, Tutorial TuTA, pp. 112–113. Technol. Lett., vol. 10, pp. 1337–1339, Sept. 1998. [31] A. C. Kam, K.-Y. Siu, R. A. Barry, and E. A. Swanson, “A cell switching [17] R. D. T. Lauder, J. M. Badcock, W. T. Holloway, and D. D. Sampson, WDM broadcast LAN with bandwidth guarantee and fair access,” J. “WDM ring network employing a shared multiwavelength incoherent Lightwave Technol., vol. 16, pp. 2265–2280, Dec. 1998. source,” IEEE Photon. Technol. Lett., vol. 10, pp. 294–296, Feb. 1998. [32] S. F. Su and R. Olshansky, “Performance of multiple access WDM [18] D. Guo and A. S. Acampora, “Scalable multihop WDM passive ring networks with subcarrier multiplexed control channels,” J. Lightwave with optimal wavelength assignment and adaptive wavelength routing,” Technol., vol. 11, pp. 1028–1033, May/June 1993. J. Lightwave Technol., vol. 14, pp. 1264–1277, June 1996. [33] K. Kitayama, “Carrier sense using subcarrier-multiplexed signaling for [19] B. Glance, U. Koren, C. A. Burrus, and J. D. Evankow, “Discretely-tuned optical frequency division multiple access,” J. Lightwave Technol., vol. N -frequency laser for packet switching applications based on WDM,” 15, pp. 758–765, May 1997. Electron. Lett., vol. 27, no. 15, pp. 1381–1383, July 18, 1991. [34] J. P. Jue, M. S. Borella, and B. Mukherjee, “Performance analysis of [20] T.-P. Lee, C. E. Zah, R. Bhat, W. C. Young, B. Pathak, F. Favire, P. S. the rainbow WDM optical network prototype,” IEEE J. Select. Areas D. Lin, N. C. Andreadakis, C. Caneau, A. W. Rahjel, M. Koza, J. K. Commun., vol. 14, pp. 945–951, June 1996. Gamelin, L. Curtis, D. D. Mahoney, and A. Lepore, “Multiwavelength [35] O. A. Lavrova, G. Rossi, and D. J. Blumenthal, “Rapid tunable trans- DFB laser array transmitters for ONTC reconfigurable optical network mitter with large number of ITU channels accessible in less than 5 ns,” testbed,” J. Lightwave Technol., vol. 14, pp. 967–976, June 1996. in Proc. ECOC 2000, vol. 2, Sept. 2000, pp. 169–170. [21] A. A. Saavedra, P.-J. Rigole, E. Goobar, R. Schatz, and S. Nilsson, [36] B. Mason, G. A. Fish, S. P. DenBaars, and L. A. Coldren, “Widely tun- “Amplitude and frequency modulation characteristics of widely tunable able sampled grating DBR laser with integrated electroabsorption mod- GCSR lasers,” IEEE Photon. Technol. Lett., vol. 10, pp. 1383–1385, ulator,” IEEE Photon. Technol. Lett., vol. 11, pp. 638–640, June 1999. Oct. 1998. [22] C. R. Doerr, C. H. Joyner, and L. W. Stulz, “40-wavelength rapidly digi- tally tunable laser,” IEEE Photon. Technol. Lett., vol. 11, pp. 1348–1350, Nov. 1999. Michael J. Spencer (M’95) photograph and biography not available for the time [23] D. J. Blumenthal, A. Carena, L. Rau, V. Curri, and S. Humphries, “All- of publication. optical label swapping with wavelength conversion for WDM-IP net- works with subcarrier multiplexed addressing,” IEEE Photon. Technol. Lett., vol. 11, pp. 1497–1502, Nov. 1999. [24] K. Sivalingam and P. W. Dowd, “A multilevel WDM access protocol for an optically interconnected multiprocessor system,” J. Lightwave Mark A. Summerfield (M’96) photograph and biography not available for the Technol., vol. 13, pp. 2152–2167, Nov. 1995. time of publication.