0% found this document useful (0 votes)
6 views13 pages

Mars An 1993

The document compares three meshed topologies for implementing large all-optical wavelength-routing networks, focusing on the K-dimensional bidirectional square lattice, twin shuffle, and de Bruijn graph. It evaluates these topologies based on their connectivity, distance metrics, and the number of wavelengths required for end-to-end transparent channels. The study emphasizes the trade-offs between wavelength and space diversity in designing efficient optical communication systems.
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
6 views13 pages

Mars An 1993

The document compares three meshed topologies for implementing large all-optical wavelength-routing networks, focusing on the K-dimensional bidirectional square lattice, twin shuffle, and de Bruijn graph. It evaluates these topologies based on their connectivity, distance metrics, and the number of wavelengths required for end-to-end transparent channels. The study emphasizes the trade-offs between wavelength and space diversity in designing efficient optical communication systems.
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 13

I I , I

534 IEEEIACM TRANSACTIONS ON NETWORKING, VOL. 1, NO. 5, OCTOBER 1993

Topologies for Wavelength-Routing


All-Optical Networks
M. Ajmone Marsan, Senior Member, IEEE, Andrea Bianco, Emilio Leonardi, and Fabio Neri

Abstruct- Three regular meshed topologies are compared in to implement in large networks with the optical technology
light of their possible use for the implementation of large all-available today.
optical wavelength-routing communication networks (or inter- A different approach is represented by wavelength routing
connection systems). These systems provide all source-destination
pairs with end-to-end transparent channels that are identified systems [2], that provide end-to-end transparent channels’ by
through a wavelength and a physical path. The considered topolo-exploiting both wavelength and space diversity in networks
with meshed topologies. In these systems, the path of the
gies are the K-dimensional bidirectional square lattice, the twin
shuffle, and the de Bruijn graph. The comparison is based on theoptical signal is determined by the source user location, by
maximum and average distance between source and destination the fiber on which the source user transmits, and by the trans-
(number of traversed nodes), on the degree of connectivity for
each node (number of input and output fibers), and on the mitted wavelength. The network nodes perform very simple
minimum number of wavelengths in the WDM comb necessary operations on the transmitted signal, and “route” it toward its
to discriminate all source-destinationpairs. destination. In general, these simple operations may include
wavelength conversion and filtering, and they may depend on
a node state defined by the traffic pattems in the network [5].
I. INTRODUCTION
In this paper we consider the case of nonadaptive wavelength-

A LL-OPTICAL COMMUNICATIONS networks exploit routing systems, where the operations performed in nodes are
photonic technologies for the implementation of both independent of the network load2. Also, we do not consider the
switching and transmission functions [11. This means that data possibility of wavelength conversion. By being so restrictive,
generated in the electronic domain by the source user are it is possible to use state-of-the-art optical components to
converted to the optical domain at the user-network interface, design systems in which an extraordinarily large transport
and transported to the destination with no conversion into an capacity can be provided to users. However, only a small
electronic format before the network-user interface. Such an fraction of this transport capacity can be accessed, due to
approach allows the elimination of the “electronic bottleneck” the limitations of the end-user transmitting and receiving
of communications networks with electronic switching, pro- equipments. Indeed, while the systems that will be described
vided that means are devised for the users to exploit the in this paper provide transparent channels for all possible
enormous speed of optics. It is widely accepted that the end-user pairs, each user normally transmits and receives
wavelength division multiplexing (WDM) approach provides only on a limited number of channels, which depends on the
such means, by partitioning the optical bandwidth into a large number of (tunable) transmitters and receivers he is equipped
number of channels whose speeds match those of the electronic with. If users are equipped with only one transmitter and one
transmission equipments [2]. receiver, the maximum network utilization is obtained when
Several approaches are currently being investigated for the each source transmits to a different destination, thus accessing
implementation of all-optical WDM networks. The two recent only one end-to-end transparent channel; this means that, if M
surveys by Mukherjee [3], [4] report on two possibilities denotes the number of users, a total of M ( M - 1) end-to-end
that are being studied: single-hop, and multihop lightwave transparent channels are provided by the network, but only M
networks. The multihop approach has the disadvantage of of those are used at most, so that the overall efficiency cannot
reintroducing to some extent the electronic bottleneck, because exceed l/(M--1). The low efficiency is the price that has to be
of the very limited processing capabilities of state-of-the-art paid for the ability to implement all-optical systems with the
optical components, which are not sufficient for the imple- presently available photonic technology. Further advantages
mentation of the node functions. The single-hop approach is of wavelength-routing systems over single-hop and multihop
normally based on optical stars, and on protocols that often lightwave networks are in the simpler network control, and
require slotted transmissions, and seem in general difficult in the spatial reuse of wavelengths permitted by meshed
Manuscript received March 30, 1993; revised August 13, 1993; approved
by IEEE/ACM TRANSACTIONS ON NETWORKING Editor Rajiv Ramaswami. This
End-to-end channels are said to be transparent because they can transport
work was supported in part by the European Institute of Technology, the
any sort of user traffic: packets, continuous digital streams, or even analog
Italian National Research Council, and the Chamber of Commerce of the City
signals.
of Turin.
The authors are with Dipartimento di Elettronica, Politecnico di Torino, 2Note that in this case the term “routing” indicates quite a different
10129 Torino, Italy (email: ajmone@polito.it) (e-mail: bianco@polito.it) (e- operation with respect to electronic networks: in that case the routing function
mail: leonardi@polito.it) (e-mail: neri@polito.it). may produce different results at different times. For nonadaptive wavelength-
IEEE Log Number 9214567. routing networks, the path followed by a wavelength never changes.

10634692/93$03.00 0 1993 IEEE


AJMONE MARSAN et a1 : WAVELENGTH-ROUTING ALLOPTICAL NETWORKS 535

topologies [ 5 ] . The possibilities offered by the combined use With the acronym ETC we denote an end-to-end transpar-
of wavelength and space diversity are illustrated in [6]. ent channel used for the transfer of information in the optical
In this paper we compare three meshed topologies in light domain between a pair of end-users. The total number of
of their possible use for the implementation of large all-optical ETC's in the network is M(M - 1). An ETC is identified
wavelength-routing communications networks or interconnec- by a path and a wavelength. A path is a sequence of point-
tion systems. As was already noted, the considered networks to-point fibers connecting a sequence of nodes that starts with
provide transparent channels for all possible end-user pairs; the origin node, traverses a number of intermediate nodes,
each channel is identified through a wavelength, and a physical and ends in the destination node. The term link will be used
path in the network. Since the allocation of a very large number to indicate the portion of an ETC between two consecutive
of wavelengths in a WDM comb is not possible with the nodes; this means that a link is identified by a fiber and a
current technology, an important measure of the quality of wavelength, and that a fiber is a set of links with different
a topology is the number of wavelengths that are required wavelengths.
to establish the end-to-end transparent channels. Obviously, For every considered topology, the number of point-to-point
the different meshed topologies provide means to trade wave- fibers out of (into) a given node is a constant, denoted by C,
length diversity for space diversity, so that many design that is also called the connectivity of the topology. The C fibers
opportunities exist. Furthermore, optical components are still out of a node are normally arranged in bundles (or cables) that
quite inefficient in terms of power loss, and optical amplifiers group fibers leading to the same adjacent node. The number
operating on very large frequency bands with good linearity of fibers per bundle (F), and the number of bundles ( B ) are
characteristics (e.g., erbium-doped fiber amplifiers) are only also constant, so that the connectivity C is obtained as the
beginning to appear. Thus, topologies that limit the power product C = FB. C is an indicator of the space diversity
loss in the path from the source to the destination are very available in the network. B measures the cabling complexity
desirable. In this paper, the comparison of the different regular of the topology. F defines the required number of fibers per
meshed topologies will be based on their wavelength reuse and bundle. The total number of unidirectional fibers in a topology
power loss characteristics. The wavelength reuse capability of is obviously obtained as MC.
a system is measured in terms of the number of wavelengths In some cases, in order to increase the spatial diversity,
necessary to establish all end-to-end transparent channels. The we also consider the possibility of using P parallel identical
power loss characteristics is measured in terms of the average topologies, thus multiplying the connectivity by P . These
and maximum number of nodes traversed along the path from parallel topologies are totally disjoint, except for the fact that
the source user to the destination user, assuming comparable they are connected to the same set of (origin and destination)
node complexity (hence losses) for the different topologies. users, thus interacting only at the user-network, and network-
In this paper we only concentrate on the design of the user interface. This parallelism can be achieved either by
wavelength routing system, and of algorithms for an efficient multiplying by P the number of fibers per bundle, or by using
information transfer, not considering the problems related P bundles running in parallel. Since we are not dealing with
to the design of transmitters and receivers. In the simplest this level of detail, we keep the same values of F and B as in
case, it may be assumed that each user is equipped with one the base topology, and denote the connectivity as C = FBP.
tunable transmitter and one tunable receiver, and that when a The distance between node i and node j in a given topology
user wants to transmit to a given destination, it continuously is the mini" number of fibers that must be traversed
sends a request message to the destination, waiting for a to reach node j starting from node i, and is denoted by
response. When the response arrives, the data transfer starts. d i l . Obviously, dii = 0. The maximum distance in a given
The destination locks its receiver to the end-to-end transparent topology is denoted by d,, = dij.
channel whose request was confirmed, until the end of the The average distance related to node z in a given topology
data transfer; then the receiver scans all incoming end-to- is denoted by E[d,],and it can be obtained as
end transparent channels until a new request is found. This
mode of operation is similar to the one adopted in the
Rainbow prototype developed at the IBM Laboratories [7].
Other approaches to the design of the user transmission and
reception equipments are possible, including a larger number where m;(d) denotes the number of nodes at distance d from
of transmitters and receivers, as well as some sophisticated node i. Note that the situation where a node transmits to itself
processing schemes to reduce the impact of receiver collisions is ruled out by this definition.
[8]. Further extensions, considering packet-switched multiple The overall average distance in a given topology is denoted
access protocols, or load-dependent circuit switching schemes, by E [ d ,and it can be obtained as
can also be designed on top of the considered topologies,
without losing the properties that are discussed in this paper.

11. PRELIMINARIES The number of wavelengths necessary to guarantee that all


The number of nodes in a given topology will be denoted the ETC's can be simultaneously active without interference
by M. will be denoted by A,; this is the number of wavelengths to
536 IEEEIACM TRANSACTIONS ON NETWORKING, VOL. 1, NO. 5, OCTOBER 1993

be allocated in the WDM comb, which is a function of M , of


C, and of the wavelength routing scheme.
For any given topology, it is possible to define a lower
bound At to the number of wavelengths AT. Indeed, it is
simple to count the total length, in number of fibers, of all the
paths connecting any source to any destination. This quantity
is denoted by H , and can be computed as

a=1 j = 1 i = l d=I

Since the total number of fibers in the network is M C , an ideal


utilization of each fiber would lead to a number of wavelengths
equal to
Fig. 1. A 27-node 3-grid topology.

(4) dimension, and hk the number of parallel bidirectional rings


in the lcth dimension.
In general, it is not possible to guarantee that this lower It can be immediately observed that, as the number of
limit to the number of wavelengths is achievable. Indeed, the dimensions K increases, the number of fibers to which a node
computation of the bound only enforces the existence in the is connected increases, so that the node structure becomes
network of a sufficient number of separable links. However, more complex. In the case of a K-grid topology with hk = h
since one and the same wavelength is assumed to be used for all K dimensions, we have B = 2K, F = h, and P = 1,
along the ETC (wavelength conversions are not considered), so that the connectivity is C = 2hK. This is the case that we
sequences of links using the same wavelength are necessary, consider throughout the paper.
while the derivation above does not guarantee the possibility A special case of a K-grid is the “hypersquare” network,
of connecting sequences of links with the same wavelength to in which the number of nodes in all dimensions is equal to
form an ETC. N , so that M = N K .
A necessary (but not sufficient) condition for the bound to be
reachable when M C is a divisor of H in (4) is that each fiber in A . Wavelength Routing
the network is used by the same number of source-destination
In a K-grid, nodes are labeled with a K-tuple of integers
pairs.
( u l , a2, . . . , a K ) , with 0 5 Uk < Nk indicating the position
of the node in the grid along the kth dimension, with respect
111. THE K-GRIDTOPOLOGY to a given reference node.
The K-grid is a direct extension to K dimensions of the A dimension, say dimension 1, is chosen as the one along
toroidal bidirectional square lattice topology, already proposed which all optical signals are transmitted. The information sent
for the implementation of high-speed networks with electronic from node ( u l , u2, . . . , a ~ to) node ( b l , 62, . . . ,b ~ travels
)
technology. Such networks were named Manhattan street along one of the hl bidirectional rings in dimension 1 (a
networks (MSN’s), and studied in [9] in the unidirectional routing function is necessary inside the origin node to select
case, and in [lo] in the bidirectional case. The all-optical one of the hl rings, as well as the appropriate circulation
version of MSN’s was proposed in [ 111-[ 131. direction along the ring, but this first routing choice can
The MSN topology is made of a number of horizontal and be implemented by the transmission electronics) until node
vertical rings that can either be unidirectional, or bidirectional. (bl, a2, . . . , a ~ is) reached. This node belongs to a ( K -
In the bidirectional case, the MSN topology is identical to the 1)-grid orthogonal to dimension 1 whose dimensions are
2-grid topology; the number of bundles out of each node is labeled 2,3, . . . ,K , that contains the destination node, but not
B = 4, and if each bundle only comprises one fiber ( F = l), necessarily the source. In node ( b l , a2, . . . , U K ) the signal is
the connectivity of the topology is C = 4, in the case of no extracted from the ring in dimension 1, and routed to one
parallelism ( P = 1). among the available 2hz unidirectional rings in dimension 2.
The 3-dimensional extension of the MSN topology (3-grid) The routing continues in all dimensions until the destination
is obtained simply by replicating in a 3-D space the planar is reached: in dimension 2, once the node ( b l , b 2 , u3, . . . , U K )
MSN topology, connecting the nodes as shown in Fig. 1 for is reached, the signal is extracted and routed in the (K - 2)-
the case of 27 nodes. Similar extensions can be used to obtain grid with dimensions 3,4, . . . , K , and so on. Data reach their
the K-dimensional topology. destination along one of the 2 h unidirectional
~ rings of the
As a further generalization, several parallel rings may be 1-grid (set of parallel rings) in dimension K.
used in the topology (this implies either F > 1, or P > 1). As K-grids are equipped with bidirectional rings, a mini-
We call Nk, k = 1,.. . , K the number of nodes in the kth mum path routing [ 121 can be obtained. Every routing choice is
AJMONE MARSAN et al.: WAVELENGTH-ROUTING ALL-OFTICAL NETWORKS 531

made so as to minimize the length of the data path. Along the to the number of wavelengths given in (4) becomes (see [ 141)
kth dimension, node ( b l , . ,b k - 1 , U k , . . ’ ,a ~ is ) traversed
by h k bidirectional rings Rt2k,j = 1,2, . . ,h k , each one of
which is made of two unidirectional rings RtiT’ and Rti;’.
The signal travels on the two unidirectional rings RPiT’ and
Rfi;) in opposite directions, called progressive and regressive
w.r.t. node numbering. Node ( b l , . . . , b k - l , a k , . . . , u K ) se-
lects one of the h k rings ti:) i~ Ibk - U k l N h < +, whereas,
if Ibk - @ I N k > +, one of the h k rings ~2;;) is chosen. For the “hypersquare” K-grid topology, the formulas reduce
When N k is even, it may happen that ( b k -
~n this case any one of the two rings ~t:’ ~2;’;
or
=
can be
+. to

selected to provide a shortest path routing; we chose to use the


progressive ring when a h is even, and the regressive ring when
a k is odd, in order to balance the utilization of the two rings.

It is possible to show that, for “hypersquare” topologies


B. Distances where N is a multiple of 4, the lower limit (7) to the number
Considering a minimum distance routing scheme, the max- of wavelengths can be reached, i.e., that A, = he. A sketch
imum distance along the kth dimension in an ETC is [+I.
of the proof under the constraint that N is a multiple of 4 is
The maximum distance between any two nodes in the network presented in Appendix A for the special case K = 2, h = 1.
can be computed as the sum of the maximum distances in This approach can be generalized to prove the same result for
each dimension: d, = E:=’=, [+I.
For the “hypersquare” other values of K and h.
K-grid topology, the formula reduces to d, =K 141.
the
By further generalizing the proof, it can be shown that
lower bound (4) to the number of wavelengths can also
In order to compute the average distance, we first define a
modified average distance along the kth dimension, E[d’,], be reached for nonhypersquare K-grids that are obtained by
which is computed considering the possibility that a node composing several hypersquare “subgrids”, whose number of
transmits to itself. This definition is very convenient, since nodes in one dimension is a multiple of 4; in the resulting
the overall modified average distance E[d‘] is then simply K-grid, N k / h k must however be an integer constant in all
computed by adding all the E[d’,]referring to the K dimen- dimensions. It is possible that the lower limit (4) be reachable
sions. Then, the transformation from E[d’] to E[dl is trivial, also for more general K-grid topologies.
amounting to the exclusion of the transmission from anv node
IV. THE TWIN SHUFFLE TOPOLOGY
In the kth dimension, in the case of N k even, E[&]= 2, The shuffle topology was first proposed in [15] for the
whereas for N k odd, ~ [ d ’ , ]= - 9 &. possible application in optical networks. Many studies of
As a special case, when the N k are either all even or all the properties of optical shuffle networks followed (see for
odd, we get example [16]-[20]). However, in some of those studies the
routing operations are performed in the electronic domain, so
M that the network is not an all-optical network in the sense
E[dl = E[d‘]-
M-1 we give to that term in this paper. Instead, we consider here
[ K
the use of the shuffle topology for all-optical wavelength-
routing systems. Since two parallel shuffles will be considered
as the basic topology, so as to simplify the wavelength routing
algorithms, we refer to this case with the name twin shuffle
topology.
In a (single) shuffle topology, nodes are arranged in k
More general combinations are easily derived by appropriately columns with p k positions each. Hence, the total number of
selecting the elements of the summation. nodes is M = k p k . The k nodes in the same position of the
For the “hypersquare” K-grid topology, the formulas reduce different columns are said to belong to the same “row” of the
to E[dl = -in the
) : - !$ Of even, whereas for odd shuffle. The node in position i of column j is connected to
we obtain E[dl = &(N - h). + +
the nodes in positions lpilpk, Ipi l l p k , . . . , Ip(i 1 ) - llpk
+
of column Ij 1 fibers are folded back from the kth column
C . Number of Wavelengths to the first, i.e., the topology ideally covers the lateral surface
of a cylinder. The connectivity of the topology is C = p ,
For K-grids with h k = h in all dimensions, in the special
since B = p, and F = P = 1. Single shuffle topologies for
case when the N k are either all even or all odd, the lower limit
all-optical wavelength-routing networks were studied in [141.
3We denote by lzly the remainder of the division of z by y, i.e., z modulo The twin shuffle-topology-is Obtained by using two fibers
31. (rather than one) in each bundle connecting a pair of adjacent
I I

538 IEEEIACM TRANSACTIONS ON NETWORKING, VOL. 1, NO. 5, OCTOBER 1993

nodes ( F = 2), thus doubling the connectivity, that becomes (recall that m ; ( d ) is the number of nodes at distance d from
C = 2p. This allows the use of a simpler routing scheme, that the reference node 2). Indeed, from any node, p nodes can
significantly reduces the number of wavelengths A, necessary be reached directly, p 2 nodes with two links, and so on, until
to connect the nodes with respect to the shuffle topology the p k nodes in the same column in which the origin node is
presented in [ 141. The two fibers belonging to the same bundle comprised are reached. However, since a node needs not reach
are identified with the labels 0 and 1, respectively, so that we itself, only p k - 1 useful nodes are reached with k links. With
have a shuffle 0 (comprising fibers with label 0) and a shuffle 1. more links, the nodes that were not already reached in the first
tour of the shuffle topology can be reached.
The average distance is easily computed from the distribu-
A . Wavelength Routing
tion as follows
The complete description of the routing scheme in an 2k-1
all-optical twin shuffle network goes together with the com-
putation of A,, the number of necessary wavelengths. In this
section we describe only the general idea behind routing in
twin shuffle networks, while details are reported in Appendices
B, C, and D.
The complexity in the routing scheme for (single) shuffle which is a well-known formula [16].
wavelength-routing networks described in [ 141 was due to
the need of guaranteeing diversity to signals transmitted from C . Number of Wavelengths
nodes belonging to one and the same column and sent to nodes To obtain A,, we first compute how many wavelengths are
at distance d > k, i.e., to nodes that can be reached only with required to provide the ETC's connecting the nodes belonging
more than one tour around the cylinder. It could indeed happen to a given column to all other nodes; then we combine
that signals using the same wavelength share the same fiber these wavelengths with column identifiers to discriminate the
after j and k + j links, for some 1 5 j < d - k. In twin shuffle position of the transmitting node.
networks, all the signals traveling on paths of length d > k, The number of wavelengths necessary to connect all the
are transmitted on fibers with label 0 by the first k traversed nodes of one column to all the nodes at distance d is pd-'
nodes (including the source node), while they are transmitted when d < k, and p"' when k 5 d 5 2k - 1. The proofs are
on fibers with label 1 by subsequent nodes (in other words, reported in Appendices B and C .
shuffle 0 is used for the first tour around the cylinder, and The number of wavelengths necessary to separate all the
shuffle 1 for the partial second tour). In this way, the signals signals transmitted from the nodes belonging to one column
reaching the same column of the source node after traveling on to all the other nodes in the network is
fibers with label 0 for k links, proceed on fibers with label 1, k-1 k-l
thus being spatially separated from signals still in the first part
of their path. All the signals reaching the destination node
on paths shorter than k (i.e., with less than one tour of the
A column identifier must be associated with each one of
cylinder), are sent and received on fibers with label 1. All the
signals following paths of length k (exactly one tour of the the Acol wavelengths, in order to allow the discrimination
of all source/destination pairs. This is due to the fact that
cylinder) are routed on fibers with label 0 if the source node
the wavelengths used in one column may interfere with the
belongs to an even column, and on fibers with label 1 if the
same wavelengths used by transmitting nodes belonging to
source node belongs to an odd column.
a different column. The total number of wavelengths can be
This routing scheme allows a solution with a number of
wavelengths not far from the bound (4) to be obtained, as thus obtained by multiplying Aco] by the number of columns
discussed in Section IV-C. k, i.e., A, = kAcol; this number can however be reduced as
shown in Appendix D, reaching a total number of necessary
wavelengths equal to
B . Distances
By inspection, it is possible to observe that the maximum
distance between any two nodes is 2k - 1. Indeed, with k links
it is possible to reach an arbitrary row in the topology (with v. THE MODIFIEDDE BRUIJNGRAPHTOPOLOGY
a first tour around the cylinder), and with k - 1 more links,
A de Bruijn graph can be seen as the state transition diagram
all nodes can be reached (with a partial second tour around
of a d stage p-ary shift register. The number of nodes in a
the cylinder).
de Bruijn graph is thus equal to M = p d ; p is the number
The distribution of the distances from an arbitrary node i
of neighbors to which each node is directly connected. The
is independent of the chosen node, and can be computed as
connectivity of this topology is C = p , since B = p , F = 1,
follows
and P = 1.
With the choice p = 2, d = 4, the interconnection of the
d<k network nodes results in the drawing presented in Fig. 2, a
m ; ( d )= pk - pd-k (8)
k 5 d 5 2k - 1 network with M = 16 nodes.
AJMONE MARSAN et al.: WAVELENGTH-ROUTING ALL-OFTICAL "IWORKS 539

n Since all ETC's follow paths of length d , even if the source


and the destination nodes are physically connected by a shorter
path, the nodes crossed by an ETC are not necessarily all
distinct, which means that also non simply connected paths
are allowed.
For example, in the de Bruijn graph topology with p = 2,
d = 4 of Fig. 2, the ETC between the origin node 0011
and the destination node 1001 follows the path traversing
the intermediates nodes 0111, 1110 and 1100, while the ETC
connecting the origin node 0011 with the destination node
1110 provides an example of a non simply connected path,
since the optical signal traverses node 0111, node 1111, again
node 1111, and finally reaches the destination node 1110.
Each bundle leaving one given node is labeled with an index
b, 0 5 b < p . The value of the index b associated with the
bundle from node nl to node n2 is defined by
b = IWd-i(ni) + Wo(n2)lp (12)
By so doing, the p input bundles for a given node are labeled
with different p-ary symbols; the same holds true for the p
output bundles.
Each individual fiber within a bundle is labeled with an
W index f. 0 5 f < d. The value of index f can be arbitrarily
assigned to the fibers belonging to a bundle.
Fig. 2. A de Bruijn graph topology with p = 2 and d = 4. Each ETC is associated with a p-ary string X, X =
(20, z1,. .zd-1) generated by the sequence of the b symbols
of the fiber bundles followed by the ETC from the source to the
De Bruijn graphs were considered as possible logical topolo- destination. There exists a one-to-one correspondence between
gies for all-optical networks in [21]. We consider an extended the string X and the wavelength associated with the ETC.
de Bruijn graph topology, connecting each node to the p In each node, the routing scheme, given the indices f i n , b,,
adjacent nodes with d parallel fibers, so that B = p , F = d , of the data input fiber and bundle, must compute the indices
and C = pd. Furthermore, we shall also consider the use of fout, bout of the outgoing fiber and bundle. The knowledge of
P = h topologies in parallel, thus bringing the connectivity the wavelength (string X) and of the input fiber is sufficient
to C = pdh. to route the signals.
Signals arriving on an input fiber with index f i n are routed
A. Wavelength Routing
+
to an output fiber with index fout = fin 1. This means
that signals are routed into a fiber with index f = 0 (in
We associate with each node n a p-ary string W ( n ) = the proper bundle b) by the source node to reach the first
[wd-l(n),...,wl(n),w~(n)] through the isomorphism be- intermediate node; then into a fiber with index f = 1 by
tween nodes and states of the p-ary shift register. wd-l(n) the first intermediate node to reach the second intermediate
is the most significant digit (corresponding to the oldest digit node, and finally into a fiber with index f = d - 1 by the
in the shift register), and wo(n) is the least significant digit last intermediate node to reach the destination node. In this
of the string. way, considering the fibers in a generic bundle, the routing
Differently from [2 11. where minimum distance routing scheme is such that only the signals in the fth link of the ETC
is considered, in our case the route followed by an ETC flow in the fth fiber, i.e., signals that have already crossed f
always traverses d links, and d - 1 intermediate nodes intermediate nodes.
n k , k = l , . . . , d - 1. The strings W ( n k ) associated The computation of the output bundle index bout requires
with the intermediate nodes are produced by progressively the knowledge of the string X associated with the ETC. The
shifting the digits of the string W(nd), associated with output bundle index bout is set to z,,where 2, is the cth digit
+
the destination node n d into the least significant positions of string X and c = f i n 1 = f o u t .
of the string W(n0) associated with the source node no. The information on the input bundle index b;, is available to
Thus, the first intermediate node is associated with the the node, but it is not used by the routing scheme as described
string W(n1) = [Wd-2(no),...,wo(no),~d-l(nd)J
second intermediate node is associated with the string
. the so far.
As an example, the ETC between the origin node 001 1 and
W(n2) = [Wd-3(nO),. ' . ,wO(nO),Wd-l(~d),Wd--2(~d)], the destination node 1001 is associated with the p a r y string
and the lcth intermediate node is associated with a string X = 1010; the optical signal is sent over the fiber with f = 0
comprising the d - k least significant digits of W(no),and of the bundle with b = 1by the origin node 0011, over the fiber
the k most significant digits of W(nd). with f = 1 of the bundle with b = 0 by the first intermediate
540 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 1, NO. 5, OCTOBER 1993

TABLE I
NUMERICAL
RESULTS
FOR THE THREE
TOPOLOGIES
IN THE CASE B =4
Topology network parameters dmax E14 A, A”,,,
2-grid M= 64 N= 8 h= 8 8 4.063 8 64
2-grid M=256 N= 16 h= 8 16 8.031 64 512
2-grid M= 1024 N= 32 h= 8 32 16.012 512 4096
2-grid Mz4096 N=64 h= 8 64 32.007 4096 32 768
2-grid M= 16384 N = 128 h= 8 128 64.003 32 768 262 144
twin shuffle M= 32 p= 4 k= 2 h= 4 3 2.258 3 24
twin shuffle Ad= 192 p= 4 k= 3 h= 4 5 3.691 32 256
twin shuffle M= 1024 p=4 k= 4 h= 4 7 5.173 224 1792
twin shuffle M=5120 p=4 k= 5 h= 4 9 6.668 1472 11 776
twin shuffle M=24576 p=4 k= 6 h= 4 11 8.197 8448 67 584
de Bruijn M= 64 p= 4 d= 3 h= 4 3 3 4 48
de Bruijn M= 256 p= 4 d= 4 h= 2 4 4 32 256
de Bruijn M= 1024 p=4 d= 5 h= 2 5 5 128 1280
de Bruijn M=4096 p=4 d= 6 h= 2 6 6 512 6144
de Bruijn A4= 16384 p=4 d= 7 h= 1 7 7 4096 28 612
de Bruijn M=65536 p=4 d= 8 h= 1 8 8 16 384 131 072

node 0111, over the fiber with f = 2 of the bundle with b = 1 such that
by the second intermediate node 1110, and finally over the
fiber with f = 3 of the bundle with b = 0 by the third and gi = J~i+l -xiJP 0 5 i < d (13)
last intermediate node 1100, finally reaching the destination
The knowledge of the string and of the input fiber still
node 1001.
allows the node to perform the routing function: the index
fout is obtained by incrementing f i n by 1, and bin defines the
position of the digit in the string Y that allows the computation
B. Distances +
of bout as bout = lyb,,, binlp. With the differential encoding
Due to the particular routing scheme that was chosen, each chosen for the string Y , also the information of the input fiber
node is at the same distance d from all other nodes, so that bundle is used for routing.
d, = E[d]= d. It is necessary to guarantee that two ETC’s originating in
different nodes, and associated with the same string Y cannot
interfere. We already showed that, when the two ETC’s are
C . Number of Wavelengths associated with the same string X , they cannot interfere, since
As we previously noted, wavelengths correspond to the p- the associated paths cannot share a fiber. When the two ETC’s
are associated with two different strings X whose differential
ary strings X of length d associated with ETC’s. An allocation
encodings are the same, we know that all the possible strings
of different wavelengths to different strings X guarantees that
two ETC’s associated with the same wavelength cannot share X generating the same string Y are composed of completely
a fiber. This is due to the fact that the labeling rule (12) different symbols, so that the paths are completely disjoint,
guarantees that two different paths sharing one or more fibers and interference cannot occur.
are associated with two different strings X , since at least one The number of wavelengths necessary to guarantee the
different symbol is present in the two strings at the point were separation of all ETC’s is thus equal to the number of p-ary
strings Y of length d - 1, that is
the two paths merge or split. Each intermediate node is thus
able to uniquely identify the source and the sink nodes by A, = pd-l =
using the information of the wavelength (string X ) and of the
input fiber.
VI. NUMERICALRESULTS
However, the routing operation implemented within nodes
does not use all the informations encoded in the string X. In this section we compare the three considered topologies,
Since the first routing decision (the one implemented by the keeping constant the number B of fiber bundles at each node,
source node) is made by the node electronics, and amounts to from the point of’view of
the selection of the proper fiber, it is not necessary to include ,d, the maximum distance between nodes; this quantity
in the string X the information on the first choice. Only d - 1 provides an indication on the maximum power loss that
digits of the string X are thus necessary, but it is not possible optical signals suffer in the network;
to simply discard the information about the first choice, i.e., the E[d],the average distance between nodes; this quantity
first digit, because the knowledge of the new string X’ (equal provides an indication on the average power loss that
to string X , less ZO), and of the input fiber would not allow to optical signals suffer in the network;
uniquely identify the source node in each intermediate node, A,., the number of wavelengths necessary to discriminate
thus possibly leading to wavelength interferences. A possible all source/destination pairs according to the wavelength
solution is to build a “differential” string Y of length d - 1, routing schemes proposed in the previous sections; this
,, , II I
I I

AJMONE MARSAN et al.: WAVELENGTH-ROUTING ALL-OPTICAL NFIWORKS 54 1

TABLE II
NUMERICAL
RESULTS FOR THE ~ sTHE CASEB = 8
THRFET o m m ~ IN
Topology network parameters amsx E[d Anor,
4-grid M= 256 N= 4 h= 4 8 4.015 32 128
4-grid M= 1296 N=6 h= 4 12 6.004 243 972
4-grid M=4096 N=8 h= 4 16 8.001 1024 4096
4-grid M=20736 N= 12 h= 4 24 12.001 1716 31 104
twin shuffle M= 128 p= 8 k= 2 h= 2 3 2.31 12 48
twin shuffle M=1536 p=8 k= 3 h= 2 5 3.860 256 1024
twin shuffle M= 16384 p=8 k= 4 h= 2 7 5.358 3584 14 336
de Bruijn M= 64 p= 8 d= 2 h= 2 2 2 4 16
de Bruijn M= 512 p= 8 d= 3 h= 2 3 3 32 192
de Bruijn M=4096 p=8 d= 4 h= 1 4 4 512 2048
de Bruijn M=32768 p=8 d= 5 h= 1 5 5 4096 20 480

......
......

."lo1 lo2 10' 10' los 10' lo2 103 104 los
Number of nodes Number of nodes
Fig. 3. Distances for networks with B = 4. Solid lines plot d,,,, while Fig. 4. Normalized number of wavelengths Anor, for networkswith B = 4
dashed lines plot E[dj.

very close in small networks, but twin shuffles become largely


quantity gives the number of frequencies to be allocated advantageous for large systems.
in the WDM comb; In the case B = 8, the qualitative behavior of distance
Anorm,the normalized number of wavelengths, obtained results are very similar, except for the fact that 4-grids now
as the product of A,. times F, the number of fibers per exhibit a growth with the number of nodes comparable to
bundle, times P, the number of parallel topologies; this that of the other two topologies; if larger numbers of di-
quantity gives the number of wavelengths for each input mensions were considered, K-grids would further improve
and output bundle at any node in the topology. their performances with respect to the other two topologies,
Tables I and I1 show results for networks with B equal to 4, as far as distances are concerned, but the nodes complexity
and 8, respectively, and with comparable values of C. Figs. 3, would become prohibitive. The differences in the number
4, 5, and 6 plot the same results versus the number of nodes of wavelengths required by the three topologies are now
in the network. less pronounced than in the previous case: the de Bruijn
In the case B = 4, the average distance in the twin graph topologies are still superior with respect to both twin
shuffle topologies is close to the distance in the de Bruijn shuffles and 4-grid topologies. Twin shuffles require slightly
graph topologies, whereas the average distance for 2-grids is less wavelengths than 4-grids.
much higher. The fact that in the de Bruijn graph topologies
the average and maximum distances are the same, makes
the advantage of these topologies with respect to maximum VII. CONCLUSIONS
distances evident. It must also be noted that distances in 2- The K-dimensional bidirectional square lattice (K-grid)
grids grow much faster with the number of nodes than for the topology, the twin shuffle topology, and the de Bruijn graph
other two topologies. Also for what regards wavelengths, the topology were compared in light of their possible use for
de Bruijn graph topologies are advantageous with respect to the implementation of large all-optical wavelength-routing
2-grids, and twin shuffles; the performances of these two are communication networks.
I I

542 IEEEIACM TRANSACTIONS ON NETWORKING, VOL. 1, NO. 5, OCTOBER 1993

.i...;...j.i.i.isi ......i....i..i.i&ij .......i...&. ;.i.i.iC......i....i..


. . + ......&.. ...:.+*
.:...+.+.:.:.t. . .):+. ......
. :... . . .j.,...... ......
. ..+:.:.
+ bnid ..:.iij.lij .......:...j..:.i.i.LJT ...........

Number of nodes

Fig. 5. Distances for networks with B = 8. Solid lines plot d, while Fig. 6. Normalized number of wavelengths A,,,, for networks with B = 8
dashed lines plot E [ d .

to the minimum number of wavelengths for the whole network;


The de Bruijn graph topology was found to be advantageous (iii) it is possible to allocate links on fibers without conflicts
from the point of view of the average and maximum distance using the wavelengths allocated according to the previous
between nodes, and from the point of view of the number of step.
wavelengths necessary to separate the signals relating to all The same three steps can also form the basis of a similar
source-destination pairs. It can be shown (see [22]) that the proof for more complex topologies.
functional structure of the optical node for the implementation
of the wavelength routing algorithm in the de Bruijn topology
is comparable to those of the other two considered topologies, A. Even Subdivision of the Traflc on the Fibers
and it appears to be feasible with state-of-the-art optical
components. Consider an N x N 2-grid; if N is a multiple of 4, from
A further advantage of the proposed wavelength-routing de (7), At = N 3 / 8 . We call “horizontal” the first dimension,
Bruijn topologies is in the constant distance for every source- where transmissions occur, and “vertical” the second dimen-
destination pair, which reduces the signal dynamics at the sion.
receiver, thus simplifying the receiver implementation. To evenly share the number of wavelengths on the two
The proposed all-optical wavelength-routing systems may horizontal rings of one row, we let the nodes with an even
provide extraordinary high aggregate data rate values with horizontal coordinate transmit to N N / 2 = N 2 / 2 destination
available optical components: a 1024-node network with end- nodes using the progressive (+) ring, and to N ( N / 2 - 1)nodes
user data rate of 1 Gbit/s offers a transport capability in using the regressive (-) ring; the N - 1 nodes belonging to
excess of 1 Pbit/s (one million Gbit/s) with a WDM scheme the same vertical ring of the source node are reached without
comprising just 128 frequencies. Of course, no more than traversing the horizontal rings. Nodes with an odd horizontal
about one Tbit/s (one thousand Gbit/s) can be accessed with coordinate will transmit to N 2 / 2 destination nodes using the
just one transmitter per node, but innovative designs for the regressive (-) ring, and to N ( N / 2 - 1) nodes using the
end user transmitter and receiver equipments can significantly progressive (+) ring.
extend this limit. Let’s focus on the fiber between nodes j and Ij l l on ~ +
one (progressive) horizontal ring. If j is odd, the number of
wavelengths traversing the fiber (number of links) is given
APPENDIX A by
OPTIMAL ALLOCATION OF WAVELENGTHS
FOR 2-GRIDS WITH h = 1
(N/2- l ) N wavelengths sent by node j
In the case of 2-grids with N a multiple of 4, and h = 1, it
is possible to prove that the lower limit (7) to the number
( N / 2 - 1)N wavelengths sent by node Ij - 1 IN
( N / 2 - 3)N wavelengths sent by node Ij - 2 ) ~
of necessary wavelengths can be reached. For the sake of (N/2- 3)N wavelengths sent by node Ij - 3 1 ~
conciseness, here we only outline the key elements of the
proof, that goes through three basic steps: (i) it is possible
to evenly share the wavelengths among all the fibers; (ii) on N wavelengths sent by node ( j - N / 2 +2 ( ~
every fiber it is possible to use a number of wavelengths equal N wavelengths sent by node Ij - N / 2 +1 1 ~
~

1 1 I I
1 1

AJMONE MARSAN er al.: WAVELENGTH-ROUTING ALL-OITICAL. NETWORKS 543

0 1 2 3 4 5 6 7 0 0 1 2 3 4 5 6 0

01 1
02 *
N 2
03 3
04 4
12 5
13 6
14 7
8
1
2
3
4
5
6
7
8

Fig. 8. A possible allocation scheme to use the minimum number of


wavelengths in one dimension.

also be applied in the vertical dimension, leading again to the


result of having he wavelengths flowing on each vertical fiber.
N This shows that it is possible to have exactly At wavelengths
+ on each fiber of the N x N network, thus meeting the necessary
condition for the minimum (7) to the number of wavelengths
N to be reached.
4
-
B. Allocation of Wavelengths in One Dimension
Fig. 7. Groups of wavelengths in one direction. We now consider the number of wavelengths necessary in
one dimension. Fig. 7 shows, in the particular case N = 8,
how many wavelengths are necessary on a fiber of a horizontal
Adding these terms, when N/4 is an integer, we obtain progressive ring. An m o w leaving node i and ending in node
j represents a group of N wavelengths used by node i to

2N
N l 4 - 1 ,-

i=O
;1_ _ - (2i+1) = N 3 / 8 = & (14)
reach the N nodes in column j. Every group of wavelengths
is labeled with the source and the destination column4.
Fig. 8 shows a possible optimal allocation of wavelengths,
always for the case N = 8. In this particular case, the
A similar result can be obtained for nodes in an even minimum number of wavelengths is N 3 / 8 = 64, i.e., eight
position of the horizontal progressive ring, and for all the groups comprising eight wavelengths each. It is in general true
fibers of the regressive ring. that this limit can be reached if the group of wavelengths used
It is thus possible, on the horizontal rings, to evenly sub- to establish the ETC’s between node i and the nodes in the
divide in space the wavelengths, in such a way that on each column of node j is used again to establish the ETC’s between
link exactly he wavelengths are used. We now verify that it +
node j and the nodes in the column of node k = li N / 2 1 ~ ;
is possible to do the same on vertical rings. the same wavelength must be used again to establish the
Every node of the network deflects from one horizontal ETC’s between node IC and the nodes in the column of the
rings to one vertical rings the signals transmitted by a node intermediate node 1, and to establish the ETC’s between node
belonging to its row, and directed to one of the N - 1 +
1 and the nodes in the column of node Ik N/21N.
nodes (excluding the node itself) of its column. The minimum The arrows in Fig. 8 represent groups of wavelengths,
distance routing algorithm splits the receivers belonging to and are labeled with the transmitting and the destination
the column of the deflecting node into two sets. These two set column; on the left side an index for the corresponding group
are not perfectly symmetric, due to the fact that the deflecting of wavelengths is shown. Eight groups of wavelengths are
node belongs to one of the two sets, so that N 2 / 2 wavelengths sufficient, as mentioned above.
are routed toward one vertical direction, and ( N / 2 l ) N 4The nodes in even positions use one group of wavelengths more than
wavelengths toward the other. It is however easy to see that nodes in odd positions, since we are considering a progressive ring. On a
the approach used for routing signals on horizontal rings can regressive ring the behavior of nodes in even and odd positions is reversed.
I I

544 IEEEIACM TRANSACTIONS ON NETWORKING, VOL. 1, NO. 5, OCTOBER 1993

column, so that two nodes in different columns, but on the


same row, have identical labels. Let node in position i be
labeled with the word W ( i )representing the number i in p -
ary notation. Fibers exist between the node labeled W in one
column and the p nodes of the next column labeled with a left
shift of W and with any p-ary symbol as the least significant
digit.
Let’s label every fiber leaving the node in position i in
one column toward the node in position j in the next column
with one p-ary symbol ~ ( i jobtained, ) in a way similar to
+
relation (12), as ~ ( i j=) I W k - l ( z ) wo(j)l,, where W k - l ( z )
is the most significant symbol of W ( i ) ,and wo(j) is the least
significant symbol of W ( j ) .By so doing, it is easy to see that
the p fibers out of, and into, a given node are labeled with
different p-ary symbols.
Fig. 9. A possible set of paths where the same wavelength is used.
Each path of length d from a source to a destination node
C . Bidimensional Problem can thus be associated with a string Xd comprising d p-
ary symbols, whose elements are the labels of the fibers
Every wavelength is used four times in the two directions, orderly forming the path (the most significant symbol of
over paths whose lengths in the first dimension are 111, N / 2 - string Xd is the label of the first fiber of the path). The
111,112 and N/2-112, with 111 and 112 E {0,1,2,3,.. . , N / 2 } ,
string Xd = ( z o , x ~ , z ~ , . . . , can z ~ -be~ seen
) to enjoy the
and in the second dimension are 121, N / 2 - 121, 122 and following properties:
N / 2 - 122, with 121 and 122 E { 0 , 1 , 2 , 3 ,.~. , N / 2 } . Note that,
1) the p d strings of length d identify exactly pd disjoint
when a node is transmitting to a receiver belonging to its same
paths leaving one given node, say node in position
column, signals need not traverse the horizontal rings, and can
be directly transmitted in the vertical dimension; in this case
i of column j ; the string X I , = 12W(i)l, of length
111 and/or 121 are equal to 0 or to N / 2 , and the wavelength is
+
k , whose elements are I w l ( i ) w l ( i ) ( , = 12w~(Z)l,,
1 = 0,1, . . . ,k - 1, identifies a path connecting a node
used only once in one semi-row.
Considering a particular wavelength, let’s assume that it with itself, and thus is not interesting;
2) the same set of p d strings of length d can be used by any
is used in row i to reach columns a, b = ( a 111I~, +
+
c = la N / 2 1 ~and +
, d = IC 1 1 2 1 ~(see Fig. 9). To allow node of one column to reach all the nodes of the column
at distance d; indeed the set of p k paths identified by the
the minimum (7) to be reached, with no loss of generality,
we impose that the paths over which one given wavelength same string Xd leaving all the nodes belonging to the
is used in the two (vertical and horizontal) dimensions must same column can be easily seen to be not overlapping:
have the same lengths in all semi-rows and all semi-columns, they may cross in one node, but do not share a fiber;
i.e., 111 = 112 in every row, and 121 = 122 in every column. 3) each string contains all the information necessary to
Fig. 9 shows in the case of an 8 x 8 2-grid the paths used route signals at any intermediate node;
by a particular wavelength assuming 111 = 112 = 3, and 4) each string contains the information to allow the receiver
121 = 122 = 2. to identify the transmitter.
When N / 2 is even (i.e., when N is a multiple of 4), the It can be observed that the information in the string Xd is
set of ETC’s built as above can be easily seen to cover all redundant, since the routing choice in the source node can
the fibers in the network when suitably translated in the two be done electronically, before entering the optical domain.
dimensions. By so doing, every wavelength is used exactly This permits one symbol to be saved. Similarly to (13), it
once over all the fibers in the network. The case where one is thus possible to label a path with a “differential” p-ary
of the l,, equals N / 2 must be handled in a slightly different string Y d - 1 = (YO1Yl,Y2,...1Yd-2),where Y i = IZi+l -GI,.
way, but it leads to similar results. The strings Y d - 1 enjoy the properties 3) and 4) mentioned
A map similar to the one shown in Fig. 9 can be drawn above for strings Xd. Note that the receiving node can
for every pair of distances 111 and 121; these maps can again rebuild the path followed by a signal, i.e., the string Xd,
cover the whole network when suitably translated. This shows and the position of the transmitting node, by combining the
that the proposed wavelenght allocation rule allows the limit information contained in the string Y d - 1 with the knowledge
(7) to be reached. of the fiber upon which such string is received.
The number of p-ary strings necessary for the identification
APPENDIXB of the paths connecting any given node of one column to
NUMBEROF WAVELENGTHS NECESSARYTO all the nodes of the column at distance d 5 k is computed
ESTABLISHETCS BETWEENNODESAT from the previous discussion as pd-’. A different wavelength
d 5 k IN A TWINSHUFFLENETWORK
DISTANCE must be associated with every string. Hence, the number of
Every node in a (twin) shuffle can be labeled with a word wavelengths needed to provide frequency diversity to the
comprising k p-ary symbols identifying its position in the information transmitted by all the nodes belonging to one
AJMONE MARSAN et al.: WAVELENGTH-ROUTING ALL-OF‘TICAL NETWORKS 545

column, and directed to all the nodes at distance d 5 k is Xkwk-l(y) - - . w k - ~ ( y ) ,wherewk-l(y) isrepeatedd-k
pd-l times.
It is possible to verify that all the paths reaching nodes at
distance d > k from a transmitting node do not overlap. This
is true also when the same set of paths is used for source nodes
AI’PENDIX c in different positions of the same column. The choice of the
NUMBEROF WAVELENGTHS NECESSARYTO
path whose last d - k labels equal wk-l(y) is arbitrary at
ESTABLISHETCS BETWEENNODESAT
this point (any p-ary symbol could have been chosen as such
DISTANCEd > k IN A TWINSHUFFLE
NETWORK
label); it will be useful for the introduction of the “differential”
Consider a path of length k between a source node in coding Y in the sequel.
position y and a destination node in position z of the same The discussion above tells us that all the information
column; the string Xk defmed in Appendix 8 can be built required to properly route signals transmitted from the node
with k sums modulo p between the symbols in omologous in position y and directed to all the nodes at distance d > k is
positions of the W representations of nodes y and z. This can contained in the k most significant symbols of the s h g s x d ,
be written as X k = W(y) @ W ( z ) ,where the symbol @ here as the d - k least significant symbols are uniquely determined
means a component-wise sum modulo p. Note that the role by the position of the transmitting node. We call XI, these
of nodes y and z can be reversed, hence the same string x k substrings.
labels the two paths of length k connecting y to z, and z to Also in this case, part of the information contained in
y. We can also write W ( z )= x k 8 W(y), which can be read the strings X k is redundant, and the “differential” coding is
stating that the representation of the destination node citn be possible, leading to strings Yk- 1. The number of these strings
obtained as the component-wise difference modulo p between ~ k - 1 is p“’.
the representation of the source node and the string labeling The passage from the string Xk to the string Yk-1 is possible
the path. only if a suitable routing of the signals is done after the kth
Suppose that a source node in position y of one column link of the path. Indeed the node traversed after k links, syz,
transmits information to all the reachable nodes at distance d belonging to the same column of the transmitting node, must
along all the possible paths. The transmitted optical signals be able to route the signals (received from shuffle 0) directed
traverse the shuffle topology, and reach, after k links, a to downstream nodes (in shuffle l), and the receiving node
node belonging to the same column of the transmitting node, must be able to identify the transmitting node. According to
where they are switched, according to the routing algorithm our proposal, the routing strategy chooses for the last d - k
of Section IV-A, from fibers with label 0 to fibers with label links of the path those fibers whose label equals the most
1 (i.e., from shuffle 0 to shuffle 1). It is easy to verify significant symbol Wk-I(y) of the W representation of the
that the set of nodes belonging to the same column of the transmitting node. By so doing, the switching node syz can
transmitting node, that are traversed by signals directed to derive the position y of the transmitting node [hence wk-l(y)]
nodes at distance d > k , comprises all the nodes of the column from the knowledge of the string Yk - 1 and of the fiber where
except those whose W representation differs only in the d - k the signal is received. Moreover, the receiving node can trace
most significant symbols from the representation W(y) of the back the last d - k fibers of the path simply by replicating
transmitting node. the label of the fiber upon which the signal is received, thus
Call Syd the set of nodes belonging to the same column of identifying the switching node syz; the knowledge of syz and
the transmitting node that are traversed by signals directed to of the string Yk-1 allows the receiving node to identify the
nodes at distance d > k. transmitting node.
Given the source node in position y, a string x k labels a
path reaching a node in the set Syd, if and only if at least one of
the k - (d - k) = 2k - d least significant symbols of XI, differs APPENDIXD
from twice modulo p the corresponding symbol of W(y). The h THE NUMBEROF
REDUCTION
number of strings X k connecting the node in position y with WAVELENGTHS IN A TWIN SHUFFLE
the nodes in the set Syd is pk - pd-k. The wavelengths used to identify ETC’s of length d 2 k
We propose EO route the signals between nodes at distance can be reused €or ETC’s of length d < k. Indeed, it is easy
d > k as follows. to observe that the signals transmitted from nodes belonging
Given a source node in position y, the nodes belonging to column j to nodes belonging to column h on a path of
to the set Syd are identified. length d > k, travel on all label 0 fibers, but only on part
For every destination node in position x, at distance of label 1 fibers: the signals are present only on the fibers
d from the transmitting node, the node in position syz connecting the nodes of column i to those of column i 1, +
belonging to S; is found by choosing the particular path with li - j l k < d - k. It is therefore possible to reuse the
of length d starting from y and ending in 2 whose last same wavelengths used to transmit from the nodes belonging
d - k fibers are all labeled with a p-ary symbol equal to to column j Eo those of column h (on paths of length d > k),
the most significant symbol Wk-l(y) of W(y). to cany the signals from the nodes of column h to the nodes of
The signal transmitted by y and directed to 5 follows column j on paths of length 2k - d, if we send these signals
the path y + syz + 2, labeled with the string xd = on fibers with label 1.
546 IEEEIACM TRANSACTIONS ON NETWORKING, VOL. 1, NO. 5 , OCTOBER 1993

The total number of different wavelengths eauals the num-


v
1221 M. Aimone Marsan. A. Bianco. E. Leonardi. and F. Neri. “De Bruiin
~1

ber of wavelengths to reach nodes at distance d 2 k from topologies for self-routing all-optical networks;” in Proc. SPIE Int. Symi.
Optical Applied Science and Engineering, San Diego, CA, July 1993.
a given column (p”’), times the number of such distances
( k ) , times the number of columns ( k again, to discriminate the
column of the transmitting node). Thus A, = k k pL-’ = m.
As a final simplification, if the signals sent on paths of lenith Marco Ajmone Marsan (SM’86) received the Dr.
IC from source nodes belonging to an even index column are Ing. degree in electronic engineering from Politec-
sent on label 0 links, while the signals sent from source nodes nico di Torino, and the M. S. from the University
of Califomia Los Angeles.
belonging to an odd index column are sent on label 1 links, it is From November 1975 to October 1987 he was
possible to use the same wavelengths (corresponding to paths with the Electronics Department of Politecnico di
of length IC) for two adjacent columns, thus saving [?1p”’ Torino, first as a Researcher, then as an Associate
Professor. From November 1987 to October 1990,
wavelengths, so that finally A, = - [yip"'. he was a Full Professor at the Computer Science
Department of the University of Milan, Italy. During
REFERENCES the summers of 1980 and 1981 he was with the
Research in Distributed Processing Group, Computer Science Department,
[I] P. E. Green, “The future of fiber-optic computer networks,” IEEE UCLA. He is currently a Full Professor at the Electronics Department of
Computer, vol. 24, pp. 78-87, Sept. 1991. Politecnico di Torino, Italy. He has co-authored about 150 joumal and
[2] C. A. Brackett, “Dense wavelength division multiplexing networks: conference papers in the areas of communications and computer science,
Principles and applications,” IEEE J. Select. Areas Commun., vol. 8, as well as the book Performance Models of Multiprocessor Systems (MIT
pp. 948-964, Aug. 1990. Press). His current interests are in the fields of performance evaluation of
[3] B. Mukherjee, “WDM-based local lightwave networks part I: Single hop data communication and computer systems, communication networks and
systems,” IEEE Network, vol. 6, pp. 12-27, May 1992. queueing theory.
[4] B. Mukherjee, “WDM-based local lightwave networks part II: Multihop Dr. Ajmone Marsan received the best paper award at the Third Intemational
systems,” IEEE Network, vol. 6 , pp. 20-32, July 1992. Conference on Distributed Computing Systems in Miami, FL, in 1982.
[5] R. A. Bany and P. A. Humblet, “On the number of wavelengths and
switches of all optical networks,” submitted for publication.
[6] J. Bannister, M. Gerla, and M. KovaEeviC, “An all-optical multifiber
tree network,” in Proc. INFOCOM ’93, San Francisco, CA, Mar. 1993.
[7] N. R. Dono, P. E. Green, K. Liu, R. Ramaswami, and F. Fuk-Kay Andrea Bianco received the Dr. Ing. degree in 1986
Tong, “A wavelength division multiple access network for computer from Politecnico di Torino, Italy. He is currently
communication,” IEEE J . Select. Areas Commun., vol. 8, pp. 984-994, a Ph. D. student in the Electronics Department,
Aug. 1990. Politecnico di Torino.
[8] I. Chlamtac and A. Fumagalli, “QUADRO-stars: High-performance In 1993 he spent a one-year period at the Hewlett-
optical WDM star networks,” in Proc. GLOBECOM ’91, Phoenix, AZ, Packard Laboratories in Palo Alto, CA, researching
Dec. 1991. on ATM congestion control issues. His research
[9] N.F. Maxemchuck, “Routing in the manhattan street network,” IEEE interests are in the fields of all-optical networks,
Trans. Commun., vol. COM-35, pp. 503-512, May 1987. ATM congestion control, and formal description
[lo] F. Borgonovo and E. Cadorin, “Routing in the bidirectional Manhattan techniques.
network,” in Proc. Third Int. Conf. Data Communication Systems and
their Performance, Rio de Janeiro, Brasil, June 1987.
1111 M. Ajmone Marsan, G. Albertengo, A. Francese, and F. Neri, “Man-
hattan topologies for passive all-optical networks,” in Proc. EFOCILAN
91, London, United Kingdom, June 1991.
1121 M. Ajmone Marsan, G. Albertengo, A. Francese, E. Leonardi, and F. Emilio Leonardi received the Dr. Ing. degree in
Neri, “All-optical bidirectional Manhattan networks,” in P roc. ICC 92, 1991 from Politecnico di Torino, Italy. He is cur-
Chicago, IL,June 1992. rently a Ph. D. student in the Electronics Department
1131 M. Ajmone Marsan, G. Albertengo, A. Francese, E. Leonardi, and F.
at Politecnico di Tonno.
Neri, “All-optical Manhattan networks with quadrant routing,” in P roc.
His research interests are in the fields of all-
ICCC 92, Genova, Italy, Sept. 1992. optical networks, and performance evaluation.
1141 M. Ajmone Marsan, A. Bianco, E. Leonardi, and F. Neri, “A comparison
of regular topologies for all-optical networks,” in Proc. INFOCOM ’93,
San Francisco, CA, Mar. 1993.
I151 A. S. Acampora, “A multichannel multihop local lightwave network,”
in Proc. GLOEECOM ’87, Tokyo, Japan, Nov. 1987.
[ 161 M. G. Hluchyj and M. J. Karol, “Shufflenet: An application of gen-
eralized perfect shuffles to multihop lightwave networks,” IEEE J .
Lightwave Technol., vol. 9, pp. 1386-1397, Oct. 1991.
1171 A. S. Acampora and S. I. A. Shah, “Multihop ~-lightwave networks:
a comparison of store-and-forward and hot-potato routing,” in Proc.
INFOCOM ’91, Miami, FL, Apr. 1991. Fabio Neri received the Dr. Ing. degree and Ph. D.
I181 A. S. Acampora, S. I. A. Shah, and 2. Zhang, “Performance analysis of in electrical engineering from Politecnico di Torino,
hot-potato routing for multiclass traffic in multihop lightwave networks,” Italy in 1981 and 1987, respectively.
in Proc. INFOCOM ’92, Florence, Italy, May 1992. He is currently an Associate Professor in the
I191 A. Bononi, F. Forghieri, and P. R. Prucnal, “Analysis of one-buffer de- Electronics Department at Politecnico di Torino.
flection routing in ultra-fast optical mesh networks,” in P roc. INFOCOM From 1982 to 1983 he was a visiting scholar at
’93, San Francisco, CA, Mar. 1993. George Washington University in Washington, DC.
1201 A. V. Ramanan, H. F. Jordan, and J. R. Sauer, “Space-time switching in From 1991 to 1992 he was with the Information En-
fiber optic packet-switched networks,” in Proc. SPIE Int. Symp. Optical gineering Department at University of Parma, Italy,
Applied Science and Engineering, San Diego, CA, July 1993. as an Associate Professor. His research interests are
1211 K. Sivaraian and R. Ramaswami, “Multihop lightwave networks based in the fields of discrete event simulation, queuing
- -
on de Bkijn graphs,” in Proc. INFOCOM ’9I,-Miami, FL,Apr. 1991. theory, and performance evaluation of communication networks.

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy