3.5   WAVELENGTH CONVERSION IN OPTICAL PACKET SWITCHING

The addition of wavelength conversion capability to an optical packet switch essentially
extends the capacity of the switch as it allows a packet to be served using any
of the available wavelengths on the preferred output port instead of just the one it
used to enter the switch. As a result, the packet loss rate in an optical packet
switch employing wavelength converters is dramatically reduced. Wavelength
converters are also used to increase the utilization of FDL buffers. Furthermore,
in architectures with wavelength-selective switches, wavelength conversion is the
only way to access the desired output port because the internal switch operation
depends on the incoming wavelength.

Wavelength conversion is a powerful contention resolution method [26–29]. In
the most typical scenario of contention, two or more incoming packets carried in the
same wavelength wish to simultaneously occupy the same output port. This conflict
can be resolved by converting all but one contending packets to other wavelengths
that are available in the target output port. Resolving contentions by wavelength
conversion has several advantages as it does not cause extra packet latency or
jitter, and does not disturb the order of packets. However, it does require additional
hardware components (wavelength converters), which are not technologically
mature.

We chose to present wavelength conversion separately and not in the section
about contention resolution that follows, because of the fact that wavelength conversion
is more often viewed as an important enabling technology rather than a contention
resolution alternative. As a result, several switch architectures incorporating
wavelength conversion have been proposed. This section presents typical WDM
optical packet switch architectures with wavelength conversion capability, which
are realized with splitters, combiners, multiplexers and demultiplexers, tunable
optical wavelength converters (TOWCs), and semiconductor optical amplifiers
(SOAs) used as optical gates.

In the classic optical packet switch architecture employing wavelength converters
there is a dedicated TOWC for each channel. The control of this switch is
fairly straightforward, but the implementation cost in terms of the number of
required converters is high. Switch architectures in which the converters are
shared are motivated by the following observations [4, 30]:

• Not all input wavelength channels contain packets at a given time instant unless
  we assume a load equal to 1.
• Not all packets directed to a specific output need to be wavelength shifted
  because some are already carried by different wavelengths.
• Any packet directed to an output where all wavelengths are busy will not need
  to be converted because it must be dropped.

In converter-sharing architectures, there is a shared pool (or bank) of TOWCs
that can be accessed by any arriving packet in need of wavelength conversion.
The drawback is the additional complexity in the switch control unit that must
take converter sharing into account. Two sharing techniques are possible [4]:¹

• Shared per node (SPN). In this architecture, the pool of converters is shared
  between all input wavelength channels in all input fibers. Therefore, they can
  be accessed by any packet regardless of its incoming or outgoing port.
• Shared per link (SPL). In this architecture, there is a separate pool of converters
  for each output fiber that can be accessed only by packets destined for that
  output port. Obviously, the number of converters in each converter pool is less
  than the number of channels (otherwise, there would not be any sharing).

These three switch architectures (one with dedicated converters and two with
converter sharing) are described and illustrated in the following paragraphs and
compared in terms of control complexity and component count. All switches
have N input and output fibers with each fiber carrying W wavelengths. A channel
is identified by a tuple (i,λ_j), where i is the input or output port and λ_j is the
wavelength, with i Î [ {1, . . . , N) and j Î [ {1, . . . ,W}.

The scheme that requires the maximum number of converters is the classic
OPS architecture in which there is a dedicated converter for every channel. This
architecture, also referred to as single per channel (SPC) architecture, is depicted
in Figure 3.7. In this switch, the traffic destined for each output port is handled by
a separate output module, which means that there is a total of W output modules.
The optical signal carried in an incoming fiber is split into sets of wavelengths
with a common destination and directed to the output module responsible for
that destination. There, the signal is split into individual wavelengths and each

wavelength is fed into a tunable filter. The role of the optical filters is to select the
wavelength that will undergo conversion in the case of contention. Each tunable
filter consists of a demultiplexer, a multiplexer, and a bank of W SOAs. Each
SOA is able to select a specific wavelength. Following wavelength conversion, all
wavelengths are again multiplexed and sent to the corresponding output fiber.
Up to W packets are allowed to exit the switch through a given output port.
Obviously, if more than W packets wish to exit using the same fiber, the excess
packets will have to be discarded. The total number of TOWCs in this architecture
is equal to N × W.

In the share per node (SPN) architecture illustrated in Figure 3.8, a pool of r
TOWCs is shared among all input wavelength channels. In other words r converters
are shared by N × W channels. The switch control unit (described later), which
is not shown in the figure, determines which packets can be switched without
wavelength conversion and which ones need to be wavelength shifted. A packet
that does not require wavelength conversion is directed to the appropriate output
module where the tunable filter is configured to select the wavelength that carries it.

Assume for example that a packet arriving at input #1 on wavelength λ_W wishes to
exit at output #N, where this wavelength is available. The SOAs that must be turned
on in order to realize this connection are the ones circled in Figure 3.8.

A packet that needs to be converted to another wavelength is first directed to the
pool of TOWCs by turning on the appropriate SOA gates. Once the packet has been
wavelength shifted, it is sent to the output fiber through an additional SOA gate.
The dotted circles in the figure indicate the gates that need to be turned on in
order for a packet arriving at input #N on wavelength λ_W to exit at output #1 on
wavelength λ₁.

The sharing of TOWCs in the SPN architecture increases the complexity of the
switching matrix in terms of the number of SOA gates. An additional set of SOA
gates is needed to allow the packets that require wavelength conversion to reach
the TOWCs and the shifted packets to reach the output fibers. When designing an
optical packet switch with shared converters arranged in the SPN architecture,
both the number of SOA gates and the minimum number of TOWCs needed to
achieve a packet loss rate similar to an SPL architecture must be computed. It
must be noted that the SPN architecture results in the greatest savings in the
number of TOWCs.

The other converter sharing alternative is the share per link architecture illustrated
in Figure 3.9, where the TOWCs are shared per output fiber. There are r_i

TOWCs (iÎ{1, . . . , N)) shared among the packets directed to a specific output
fiber. The packets that need wavelength conversion are sent to the output branches
where there are TOWCs. Packets are directed to wavelength converters and/or
output fibers by SOA gates, which are turned on by the switch control unit.
The SPL architecture facilitates partial sharing of wavelength converters. As
a result, the savings in the number of TOWCs are less compared to the SPN architecture.
Nevertheless, fewer wavelength converters are required compared to the
SPC architecture. Furthermore, the number of SOA gates needed to realize an
SPL architecture is equal to the one needed for SPC but less than the one needed
for SPN.

A simple control algorithm can be applied in the SPL architecture when the
switch operates in an asynchronous manner [4]. An arriving packet is forwarded
to the desired output port if the wavelength that carries it is free at that port.
Otherwise, if a converter is available, the packet is shifted to a free wavelength
that is randomly selected. If all converters are occupied or all wavelengths at the
desired output port are busy, the packet is dropped.

In a synchronous switch architecture, more than one packet on the same wavelength
destined for the same output port may arrive within a given time slot.
In this case, the control algorithm dictates that one of the packets will be chosen
at random and forwarded to the output without wavelength conversion [4]. The
remaining packets will be wavelength shifted assuming there are both available
channels and TOWCs. The channels selected for packet transmission are again
chosen randomly. The switch control unit adopts a simple and fair technique for
the assignment of the TOWCs to incoming packets at each timeslot. The assignment
of TOWCs to packets is performed in three phases: in the first phase, the algorithm
determines the set of packets contending for a specified wavelength at a given
output. In the second phase, the algorithm selects a packet, if one exists, that will
be transmitted without wavelength conversion. In the third phase, the algorithm
attempts to forward the remaining packets to the switch exit by using wavelength
conversion. Packets are handled in a random order. Obviously, a packet can be forwarded
only if there is a free wavelength and an available TOWC. This control
algorithm minimizes the number of required wavelength conversions, because it
ensures that one of the packets that contend for the same output will be forwarded
without wavelength conversion [4].

When comparing the complexity of the three packet switching architectures
described above, both the number of SOA gates and the number of TOWCs
must be taken into account. The number of SOA gates needed to realize these
architectures for N fibers and W wavelengths can be calculated by the following
expressions [4]:

From these expressions, it is obvious that the SPC and SPL architectures require the
same number of SOA gates, and the SPN architecture requires an additional
r(2N + W) gates controlling the access to the wavelength converters.

The evaluation of the number of TOWCs is not as straightforward as that of
SOAs for the SPL and SPN architectures. For the SPC architecture, the number
of required converters is fixed and equal to N × W because there is no sharing of
converters. For the other two architectures, however, a dimensioning procedure
must be carried out in order to compute the minimum number of TOWCs needed
to keep the packet loss rate below a specified threshold or equal to that of an
SPC architecture of the same size. Obviously, this dimensioning depends on both
the system parameters (number of input fibers, number of wavelengths), and the
traffic parameters. In the dimensioning procedure described in ref. 4 it is assumed
that the packet arrivals in a given channel and each timeslot are independent and
occur with probability a₀. Because the number of TOWCs needed at a given timeslot
depends only on the number of packets arriving at that slot, no assumptions regarding
the arrivals of packets at different timeslots are necessary. Furthermore because
the switch architectures under study are bufferless, the number of packets lost at
each timeslot depends only on the traffic intensity (a₀). Hence, the results obtained
by the dimensioning procedure for a given performance level hold for any type of
input traffic.

In the analysis performed in ref. 4, it is assumed that traffic is symmetric, that is,
the probability that an arriving packet is destined for output #k (k = 1, . . . , N) is the
same for all k. The symmetric traffic scenario, although not very realistic, was
chosen because it leads to the lowest packet loss probability (in the SPC architecture)
and thus requires the most severe dimensioning of TOWCs. This can also be
explained as follows: in an optical packet switch with wavelength conversion
capability, a packet is lost only if all wavelengths at the desired output port are
busy or when there are no available converters. The assumption of symmetric
traffic leads to the minimum number of packets lost due to unavailable wavelengths,
which means that more packets will require wavelength conversion; hence a
larger number of wavelength converters will have to be utilized in order to
achieve a given performance objective.

Results obtained from simulation and analysis [4] indicate that in both architectures
with shared converters (SPN and SPL) the packet loss rate decreases as
the number of converters increases until it saturates at the value of the packet loss
probability of the SPC architecture. This occurs when the number of TOWCs r
reaches a threshold value r_th, which represents the minimum number of TOWCs
that must be used to guarantee a packet loss probability equal to that of the SPC
architecture. The SPN architecture allows remarkable savings in the number of
TOWCs both for synchronous and asynchronous switch architectures [4], which
comes at the cost of an increase in the number of SOAs (not nearly proportional
to the decrease in TOWCs). The savings in asynchronous architectures, however,
are significantly smaller compared to synchronous ones. This difference is due to
the fact that in synchronous architectures, packets are aligned on a timeslot basis
and all packets that arrive within a given timeslot are switched together; this type
of switching enables the optimum allocation of wavelength converters. On the other
hand, in asynchronous architectures, packets arrive and enter the switch without
being synchronized. In this case each packet is served individually and without
considering other packet arrivals. As a result, wavelength conversion decisions
are not optimum. The ideal decision would be to convert a packet to a wavelength
on which no packets will arrive while it is being forwarded. However, this is not
possible as packets are switched asynchronously and there is no knowledge regarding
the arrivals of other packets within the current timeslot.