Performance Evaluations On A Buffering WDM Optical

See
discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/237379134
Performance Evaluations on a Buffering WDM

Optical Packet Switch Equipped with Shared
Wavelength Converters
Article
CITATIONS READS
0 18
1 author:
Jen-Fa Huang
National Cheng Kung University
116 PUBLICATIONS 726 CITATIONS
SEE PROFILE
All content following this page was uploaded by Jen-Fa Huang on 05 May 2017.
The user has requested enhancement of the downloaded file. All in-text references underlined in blue are added to the original document
and are linked to publications on ResearchGate, letting you access and read them immediately.

Performance Evaluations on
a Buffering WDM Optical Packet Switch
Equipped with Shared Wavelength Converters
StudentChia-Hsing Chou
AdvisorJen-Fa Huang
Department of Electrical Engineering,

Tainan, Taiwan, R.O.C.
Thesis for Master of Science
June, 2002
* **
(Wavelength dimension)
(Wavelength Division Multiplexing, WDM)
(Tunable Wavelength Converters, TWCs)
(Packet loss probability)
(survivor function)
(Mean contention time)

(Mean packet delay)
BI/D/B/N G[x]/D/1/B
(Mean total delay)
*
**
Performance Evaluations on
a Buffering WDM Optical Packet Switch
Equipped with Shared Wavelength Converters
Chia-Hsing Chou* Jen-Fa Huang**

Department of Electrical Engineering,
National Cheng Kung University, Tainan, Taiwan, R.O.C.
Abstract
In this thesis, a buffering WDM optical packet switch scheme employing the
wavelength converters for contention resolution is evaluated. The architecture is dedicated
to an individual buffer for each output and is equipped with tunable wavelength converters
(TWCs) shared among the input lines. Combined with the wavelength dimension for
contention resolution, this reduces the size of each output buffer and employs a lesser
number of wavelength converters; therefore, it is cost-effective. Based on the architecture,
lower packet loss probability can be achieved. We also observe that buffering optical
switch allows reductions on the employed number of converters compared with the
bufferless architecture. Besides, we will evaluate survivor function by means of numerical
simulations.
Furthermore, we analyze mean contention time in output contention process and

mean packet delay in output buffer. Our studies indicate that the output contention process
can be modeled by a BI/D/B/N queueing process and the output buffer can be modeled by a
G[x]/D/1/B queueing process. Mean total delay in buffering packet switch equipped with
shared TWCs is evaluated by means of numerical simulations.
* The author
** The advisor
(Ackonwledgments)

Table of Contents
Chapter 1. Introduction ... .1
1.1. SCM Header and Packet Format.........2
1.2. Contention Resolution.......5
1.2.1. Wavelength Dimension Technique.5
1.2.2. Buffering Technique.7
1.2.3. Deflection Routing Technique....9
1.3. The Motivation of our Research.....12
Chapter 2. Previous Architectures of WDM Packet Switch...14
2.1. Wavelength Conversion Technique........14
2.2. Bufferless WDM Switch with Shared TWCs.......17
2.3. Partially Shared Buffering for WDM Packet Switch..19
2.4. Comparisons of Bufferless and Shared Buffer Schemes22
Chapter 3. Shared TWCs for Buffering WDM Packet Switch..25
3.1. WDM Optical Packet Switch......25
3.2. Dimensioning the Number of TWCs..28
3.3. Evaluation of packet loss probability36
3.3.1. Evaluation of Ploss in Full-Conversion Configuration (r = NM)..36

3.3.2. Evaluation of Ploss in No-Conversion Configuration (r = 0)37
3.3.3. Evaluation of Ploss when r TWCs are utilized (0 < r < NM)39
3.4. Numerical Results of Packet Loss Probability41
Chapter 4. Evaluation of mean total delay and survivor function ..45
4.1. Numerical Results of Mean Total Delay........45
4.1.1. Output Contention Process (BI/D/B/N queue).....46
4.1.2. Output Buffer Process (G[x]/D/1/B queue).47
4.2. Numerical Results of Survivor Function.52
Chapter 5. Concluding Remarks......55
References
List of Figures
Figure 1.1. Qualitative power spectrum of the laser modulation current...3
Figure 1.2. Header retrieval in SCM..4
Figure 1.3. One TWC is needed for each input wavelength channel for the WDM packet
switch.6
Figure 1.4. A two-input, two-output routing node in the network.8
Figure 1.5. The shared memory optical packet switch..9
Figure 1.6. The Manhattan Street Network.10
Figure 1.7. The ShuffleNet..11
Figure 2.1. Buffer filling process with and without wavelength converters15
Figure 2.2. (a)Mach-Zehnder interferometer and (b) Michelson interferometer setups using
a pair of SOAs for implementing the cross-phase modulation wavelength-conversion
scheme.16
Figure 2.3. Structure of bufferless WDM optical switch equipped with TWCs..18
Figure 2.4. Wavelengths of arriving packets (a) before and (b) after conversions...19
Figure 2.5. WDM Packet Switches with Partially Shared Buffering (PSB)20
Figure 2.6. Wavelength conversions for recirculated packets from the shared buffer.21
Figure 2.7. (a) Queueing model without TWCs; and (b) PSB model with TWCs...22
Figure 3.1. Structure of the buffering WDM optical switch equipped with tunable
wavelength converters..26
Figure 3.2. The queueing model with wavelength conversions..27
Figure 3.3. A buffer size of B=2 with arriving packet(s) on wavelength 230/31
Figure 3.4. Arriving packets wavelengths (a). before and (b). after conversion processes
(M = 4, k = 2, wk = 4, p = q1 , q2
Ak , p = 9 , B = 2)32
Figure 3.5. Conversion process when the arriving packet is less than BM
(M = 4, k = 2, wk = 3, p = q ,q Ak , p = 6 , B = 2)...34
1 2
Figure 3.6. Packet loss probability of the analytical switch model when N=16, q=0.8, M= 4,
6, and B=1 (no buffer), 2.42
Figure 3.7. Packet loss probability versus number of converters for q=0.6.43
Figure 3.8. Packet loss probability versus number of converters for q=0.8. ..44
Figure 3.9. Packet loss probability versus number of converters for N=16.44
Figure 4.1. State transition diagram of contention process for output buffer k50
Figure 4.2. Mean total delay versus traffic load q51
Figure 4.3. Survivor function versus the number of conversions for M=4 and q=0.8.53
Figure 4.4. Survivor function versus the number of conversions for M=6 and q=0.8.53
Figure 4.5. Survivor function versus the number of conversions for B=154
Figure 4.6. Survivor function versus the number of conversions for B=254
List of Tables
Table I. M = 4, k = 2, B = 2, total number of packets BM.33
Table II. M = 4, k = 2, B = 2, total number of packets < BM...35

Chapter 1. Introduction
Wavelength-division multiplexing (WDM) techniques have rapidly gaining

popularity as technology is able to handle the dramatic bandwidth demand in future
networks. The future transparent optical network is likely to use WDM to obtain high
flexibility and network throughput. Current optical switching technologies allow us to
rapidly deliver the enormous bandwidth of WDM networks. Besides the huge amounts of
bandwidth, all-optical WDM networks also allow high-speed data transmission without
electronic converters at intermediate nodes. The packets are switched from input fibers to
output fibers optically. The networks are transparency with respect to data format to be
achieved. The key system features of WDM are as follows: [1]
Capacity upgrade: The classical application of WDM has been to upgrade the
capacity of existing point-to-point fiber optic transmission links. If each wavelength
supports an independent network signal of perhaps a few gigabits per second, then WDM
can increase the capacity of a fiber network dramatically.
Transparency: An important aspect of WDM is that each optical channel can carry
any transmission format. Thus, using different wavelengths, fast or slow asynchronous and
synchronous digital data and analog information can be sent simultaneously, and
independently, over the same fiber, without the need for a common signal structure.
Wavelength routing: In addition to using multiple wavelengths to increase link

capacity and flexibility, the use of wavelength-sensitive optical routing devices makes it
possible to use wavelength as another dimension, in addition to time and space, in
designing communication networks and switches. Wavelength routed networks use the
actual wavelength of a signal as the intermediate or final address.
Wavelength switching: Whereas wavelength-routed networks are based on a rigid

fiber infrastructure, wavelength-switched architectures allow reconfigurations of the
optical layer. Key components for implementing these networks include optical add/drop
multiplexers, optical cross connects, and wavelength converters.
1
A further reason leading to optical packet switching is its intrinsic flexibility to
cheaply support incremental increases of the transmission bit rate, so that frequent
upgrades of the transmission layer capacity can be assumed to match increasing bandwidth
demand with a minor influence on switching nodes. As a consequence, a promising
guideline for the network evolution could consist in the migration of most of the switching
burden into the optical domain in order to exploit the scalability provides by optical
technology to support the progressive increments of transmission capacity of WDM lines.
This method could result in an effective decoupling between transmission/switching and
routing/forwarding functionality. The former should be handled in the optical domain, in
order to access the huge fiber bandwidth; the latter should be accomplished in the
electronic domain, where the routing/forwarding functions based on packet header
processing would be performed.
1.1 SCM Header and Packet Format
The packet header is transmitted serially with the payload data at the same data rate
in electronic networks (e.g., ATM cells and IP packets). Electronic routers or switches will
process the header information at the same data rate as the payload. In an optical network,
the bandwidth is much larger than their electronic counterparts. A typical wavelength
channel has a line speed of 2.5 Gb/s (OC 48). Although there are various techniques to
detect and recognize packet headers at gigabit/second speed, either electronically or
optically, it is still hard to implement electronic header processors operating at such high
speed to switch packets on the fly at every node. [2]
Among several different proposed solutions, packet switching with subcarrier

multiplexed (SCM) header is attracting increasing interest. In this approach the header and
payload data are multiplexed on the same wavelength (optical carrier). In the current that
modulates the laser transmitter, payload data is encoded at the baseband, while header bits
are encoded on a properly chosen subcarrier frequency at a lower bit rate, as shown in
Figure 1.1. The header information on different wavelengths can be retrieved by detecting
a small fraction of the light in the fiber with just a conventional photodetector, without any
2
type of optical filtering. In the output current of the photodetector various data streams
from different wavelengths jam at baseband, but the subcarrier remains distinct, and the
header can be retrieved by electrically filtering out the desired subcarrier (Figure 1.2).
Header
Power
Payload
(baseband)
Frequency f sc
Figure 1.1. Qualitative power spectrum of the laser modulation current.
A nice feature of the SCM header is that is can be transmitted on top of the payload
data and can take up the whole payload transmission time, since it does not interfere with
the payload. Of course, the header can also be transmitted serially with the payload if so
desired. One potential pitfall of the SCM header is its possible limit on the payload data
rate. If the payload data is increased, the baseband will expand and might eventually
overlap with the subcarrier frequency, which is limited by the microwave electronics.
In many of the routing and switching protocols, packet headers have to be updated
at each node. There have been several approaches proposed on all-optical
same-wavelength header replacement for headers transmitted serially with the payload data
stream. All-optical header replacement could be done by blocking the old header with a
fast optical switch and inserting the new header, generated locally by another laser, at the
proper time. One important issue here is that in WDM networks the new header should be
precisely at the same wavelength as the payload data; otherwise, serious problems could
arise because of dispersion, nonlinearity, or wavelength-sensitive devices in the network.
3
Payload Header
1
Photodetector
Laser
2
1
f1 3
Payload Header
Payload
Laser
2
f2
f1 f2 f3
Payload Header
Laser
3
f3
Figure 1.2. Header retrieval in SCM.
Packet length is another problem of concern for network designers. A greater

percentage of the bandwidth is given to the header or guard time between time slots, so a
short packet might not give good throughput. On the other hand, a long packet would need
longer optical buffers, and not provide a guarantee that is fine enough. Payload usually
contains many more bits than the header. If the header is updated at every traversed node,
the bits in the payload will have to suffer more physical impairment than the bits in the
header. Another fact is that if SCM is used, the header is usually transmitted at a lower bit
rate than the payload data. All these facts lead to a big advantage of lower bit error rate
(BER) for header bits over the payload bits over the payload bits. Therefore, it is
imperative to optimize the amount of power to be tapped from the packet at each node and
the packet length in order to achieve a balanced packet error rate (PER) for payload and
header at the destination node.
For a header of fixed size, the time taken for demultiplexing and processing the
header is fixed, and the remainder of the packet is buffered optically using a fiber delay
4
line of appropriate length. The processing of the header bits may be done electronically or
optically, depending on the kind of control input required by the switch. Electrically
controlled switches employing the electro-optic effect and fabricated in lithium niobate are
most commonly used in switch-based network experiments today. In this case, the header
processing can be carried out electronically (after the header bits have been demultiplexed
into a parallel stream). The packet destination information from the header is used to
determine the outgoing link from the switch for this packet, using a look-up table. If the
switch has two inputs and outputs, as shown in Figure 1.2, the look-up table determines,
for each input packet, whether the switch should be set in the cross state or bar state. Of
course, this leads to a conflict if both inputs have a packet destined for the same output.
This is one of the reasons for having buffers in the routing node, as explained front.
1.2 Contention Resolution
One of the key problems in optical application of packet switching is the handling
of packet contentions that take place when two or more incoming packets arriving on the
same wavelength are directed to the same output line. In optical WDM packet switch, if
contention occurs, some of the contending packets can be processed to reach the
destination, the others may be lost. How contention is resolved has a great effect on
network performance. Various techniques on overcoming packet contentions have been
examined in the literatures, including wavelength dimensioning [3], input/output buffering
[4, 5], deflection routing [2, 6], and wavelength conversions [7,8,9].
1.2.1 Wavelength Dimension Technique
The wavelength dimension technique [3] uses wavelength dimensions as a logical

buffer in the WDM optical network layer. The requirements to the size of optical buffers
are relaxed by introducing multiple wavelength channels to decrease the traffic load of
each channel. By using TWCs to assign packets to unused wavelengths in the fiber
delay-line buffers we show that a reduction of fiber delay-lines in the WDM packet switch
is obtained. In Figure 1.3, a network solution is proposed that eliminates the need for
5
optical buffers by splitting the traffic load on the wavelength channels by using TWCs, and
one TWC is needed for each input wavelength channel. As an example, it has 16 input and
output fibers, each fiber supports a WDM signal with 16 wavelengths; therefore, the
optical switch needs a high number of converters equal to the number of input fiber
multiplied the number of wavelengths supported on each fiber, that is, 256 TWCs. [8]
The architecture uses the wavelength dimension for storing packets. It uses the
wavelength dimension as a logical buffer in the WDM optical network layer. This is
followed by a brief description of the traffic model that is applied to the switches in the
network. This model is then coupled with the network protocol to illustrate that the buffers
can indeed be eliminated: with 11 WDM channels per fiber that buffers can be discarded if
packet loss probabilities of 10-10 are accepted.
Input Fiber DMUX MUX Output Fiber
1
1
1 1
...
...
1 , 2 ,L, M 1 , 2 ,L, M
M
M Space
...
...
1 switch 1
N N
...
...
1 , 2 ,L, M 1 , 2 ,L, M
M
M
Tunable wavelength converter
Figure 1.3. One TWC is needed for each input wavelength channel for the WDM packet
switch.
As aforementioned, in Figure 1.3, the number of TWCs is assumed to be

proportional to both the number of input fibers and the number of wavelengths. However,
it is to be noted that only a small part of TWCs is simultaneously utilized; this is due to
two main reasons:
6
A. An input channel, identified by the couple input line, wavelength, at a given
instant could not contain packets, whether we suppose to have on the channel a
load less than 100%.
B. Not all of packets contending for the same output line have to be shifted in
wavelength because they are already carried by different wavelengths.
1.2.2 Buffering Technique
In general, a routing node contains buffers to store the packets from the incoming
links before they can be transmitted or forwarded on the outgoing links. Hence the name
store and forward for these networks. In a general store and forward network, electronic or
optical, the buffers may be present at the input only, at the outputs only, or at both the
inputs and the outputs, as shown in Figure 1.4. The buffers may also be integrated within
the switch itself. This option is used quite often in the case of electronic networks where
both the memory and switch fabric are fabricated on the same substrate, say, a
silicon-integrated circuit. This is unlikely to be practical, or even feasible, in photonic
packet-switching networks since the memory typically consists of relatively long lengths
of fiber (20 cm/ns), whereas the switch fabric is made in an integrated-optics substrate
such as lithium niobate.
There are at least three reasons for having to store or buffer a packet before it is
forwarded on its outgoing link. First, the incoming packet must be buffered while the
packet header is processed to determine how the packet must be routed. Second, the
required switch input and/or output port may not be free, causing the packet to be queued
at its input buffer. The switch input may not be free because other packets that arrived on
the same link have to be served earlier. The switch output port may not be free because
packets from other input ports are being switched to it. Third, after the packet has been
switched to the required output port, the outgoing link from this port may be busy
transmitting other packets, thus making this packet wait for its turn. In the case of photonic
packet switching networks, buffers are scarce resources. The best known way to construct
an optical buffer is to use a piece of optical fiber and delay the signal within it. Thus
7
usually very small buffers are used in photonic packet switching networks. Note that unlike
an electronic buffer, a packet cannot be accessed at an arbitrary point of time; it can exit
the buffer only after a fixed time interval after entering it. This is the time taken for the
packet to traverse the fiber length. [10]
Input Output
buffers buffers
Header
recognition
Packets
Switch
Header
recognition
Control
input
Figure 1.4. A two-input, two-output routing node in the network.
The output buffering [4,5] achieves optimal throughput performance and is widely
applied to WDM optical packet switch architectures. In electronic routers, contention is
usually resolved by a store-and-forward technique, which means that the packets in
contention are stored in a queue and sent out one by one. This is possible because of the
available random access memory (RAM). At the current technology, optical buffer can
only be implemented through a bundle of fiber delay lines, with lengths equal to a multiple
of packet duration. If contention occurs, only one of the contending packets may be
allowed to reach the destination, the others must be buffered for later transmission. Here,
wavelength also assists in contention resolution, where contending packets may be
transmitted on different wavelengths, reducing the optical delay-line buffering
requirements.
8
Delay lines of lengths d1 ,L, d m
dm
...
d1
...
1 1
...
...
m m
1 1
Input Output
...
...
n n
( n + m ) ( n + m)
Figure 1.5. The shared memory optical packet switch.
The example is the shared-memory optical packet (SMOP) switch, which belongs
to the single-stage feedback category. It is very straightforward to see how the switch
works from Figure 1.5. The lengths of the delay lines could be 1, 2, 3,, m packet
duration. The (n+m) (n+m) space switch can switch a packet either directly to an output
port or to one of the delay lines, according to how much delay the packet needs. Delay
lines of length greater than one recirculation loops needed, resulting in a reduced need for
amplifiers and therefore less noise. This scheme also allows packet priorities since a
lower-priority packet may be preempted by being sent to another circulation. Since the
number of recirculations a packet is to take is unpredictable, some packets could suffer
more power loss than others, making optical amplification necessary. This will inevitably
introduce additional signal-to- ration degradation into the recirculating packets.
1.2.3 Deflection Routing Technique
9
Deflection routing was invented by Baran in 1964. In the deflection routing scheme
[2,6], contention is resolved as follows: if two or more packets need to use the same output
link to achieve minimum distance routing, only one will be routed along the desired link,
while others are forwarded on paths which may lead to greater than minimum distance
routing. Hence, for each source-destination pair, the number of hops taken by a packet is
no longer fixed. Deflection routing does not necessary exclude the use of optical buffers.
Deflection routing is also sometimes called hot-potato routing.
00 01 02
10 11 12
20 21 22
3 3 Manhattan Street Network
Figure 1.6. The Manhattan Street Network.
Extensive studies on deflection routing have been carried out in regular network
topologies with uniform traffic load. These logical topologies can be built on different
physical topologies (e.g., star, ring, or mesh). Figures 1.6 and 1.7 show the two most
typical logical topologies used for network performance simulation: the Manhattan Street
Network (MSN) and ShuffleNet. Each node in these two topologies has two input ports
and two output ports. A node has to handle both bypassing and locally
10
generated/terminated packets.
0 4 0
1 5 1
2 6 2
3 7 3
8-node ShuffleNet
Figure 1.7. The ShuffleNet.
Studies have been conducted to determine the impact of different routing strategies
on network performance, such as delays, average number of hops (i.e., the number of
switches a packet has to traverse between the source and destination node) for each packet,
and network aggregate capacity (the number of packets a network can process within a
certain period of time). A comparison done on the ShuffleNet topology between
store-and-forward and hot-potato routing shows that the average number of hops for each
packet is larger for hot-potato routing, because not all the packets take the shortest route
toward their destinations. As the number of users (or number of nodes) increases, both the
average number of hops and aggregate capacities increase for both routing strategies. In
multihop networks, where information from a source node to a destination node may be
routed through intermediate nodes, only a portion of the network capacity is used for
11
newly generated traffic.
The deflection routing is simply a multiple path routing technique that allows the
contention problems to be solved and the buffer depth and the number of optical gates to
be reduced with a sensitive saving in hardware volume and cost. The effectiveness of this
technique critically depends on the network topology; as a matter of example, meshed
topologies with a high number of interconnections benefit of the largest gain from
deflection routing whereas minor advantages arises from more simple topologies.
1.3 The Motivation of our Research
Both output buffering and deflection routing schemes have advantages and
disadvantages: output buffering offers better network throughput but involves more
hardware and controls; deflection routing is easier to implement, but cannot offer ideal
network performance. When combined with wavelength conversion, the disadvantages
could be overcame or minimized, therefore giving the network designer more choices and
flexibilities.
A wavelength converter is needed to shift the wavelength of an optical packet. The

switch architecture must be designed in a cost-effective way such that only a minimum
number of wavelength converters are needed. In addition, TWCs are active devices that
create noise. The drawback of the delay-line buffer is that it is spacious and inconvenient.
For example, more than 200m-length fiber is needed to delay an optical packet for 1s.
Also, the delayed optical packet cannot be randomly retrieved. However, at present, there
is no better device to buffer optical packets than the fiber delay line. Therefore, combined
with the use of wavelength dimension for contention resolution, this reduces the size of
each output buffer and employs a lesser number of wavelength converters. Output
buffering technique and wavelength conversion technique each has its advantages and
disadvantages; therefore, we will utilize output buffering technique and wavelength
conversion technique to propose a new architecture based on [3].
In this paper, we evaluate the buffering optical WDM packet switch equipped with
12
shared tunable wavelength converters (TWCs) for contention resolution. An analytical
model is carried out on the required number of converters with regard to number of buffers,
number of wavelengths, buffer size, and traffic loads. We will observe that buffering
optical switch allows reductions on the employed number of converters compared with the
bufferless architecture discussed in [3].
Also, we analyze mean contention time in output contention process and mean
packet delay in output queue. From these two parameters, mean total delay is evaluated.
Our studies indicate that the output contention process can be modeled by a BI/D/B/N
queueing process and the output buffer can be modeled by a G[x]/D/1/B queueing process.
Performances of mean total delay and survivor function in buffering packet switch
equipped with shared TWCs are extensively studied by means of numerical simulations.
13
Chapter 2. Previous Architectures of
WDM Packet Switch
In this chapter, we briefly discuss the bufferless WDM switch with shared TWCs
and the Partially Shared Buffering for WDM Packet Switch. The architecture for a
bufferless optical packet switch making use of the wavelength dimension for contention
resolution of packets directed to the output line. The optical switch architecture is provided
with wavelength converters shared among the output lines. An analytical model, allowing
the dimensioning of the number of wavelength converters according to a required
performance, has been presented. The other architecture for an optical packet switch is
with Partially Shared Buffering (PSB). The architecture allocates a buffer to be shared by
all outputs, in addition to the statistical nature of packet arrivals, more efficient use is made
of the shared buffer resources. Therefore, a lower packet loss probability can be achieved
without the need of either employing wavelength converters.
2.1 Wavelength Conversion Technique
In order to reduce the buffer size, the wavelength conversion techniques [7,8,9]
have been proposed. Packet contentions are handled in the wavelength domain: packets
addressed toward the same output are converted to different wavelengths by means of
tunable wavelength converters (TWCs). The advantage of wavelength conversion is the
ability to shift the wavelength of an optical packet to another wavelength so as to
effectively resolve wavelength contentions. The use of TWCs is recognized as essential for
reducing the complexity of photonic WDM packet switches.
Figure 2.1 shows that buffer filling process with and without wavelength converters.
Two packets, P1 and P2 destined for the same outlet, arrive simultaneously at the same
wavelength, 1. without converters, two fiber delay lines are needed to store the packets,
whereas with converters, only one delay-line is needed.
14
1
1
P2 P1
P1
(a) 1 1
P2
1
1
TWC 2
P1
P1
(b) 1 P2
TWC
P2 2
Figure 2.1. Buffer filling process with and without wavelength converters.
An optical wavelength converter is a device that can directly translate information

contained on an incoming wavelength to a new wavelength without entering the electrical
domain. This is an important component in all-optical networks, since the wavelength of
the incoming signal may already be in use by another information channel residing on the
destined outgoing path. Converting the incoming signal to a new wavelength will allow
both information channels to traverse the same fiber simultaneously. Here, we will describe
two classes of wavelength converters. [1]
The first class is optical-gating wavelength converters. A wide variety of

optical-gating techniques using devices such as semiconductor optical amplifiers,
semiconductor laser, or nonlinear optical-loop mirrors have been investigated to achieve
wavelength conversion. The use of a SOA in a cross-phase modulation (XPM) mode has
been one of the most successful techniques for implementing single-wavelength
conversion. The configurations for implementing this scheme include the Mach-Zehnder or
the Michelson interferometer setups shown in Figure 2.2.
The XPM scheme relies on the dependency of the refractive index on the carrier
density in the active region of the SOA. As depicted in Figure 2.2, the basic concept is that
an incoming information-carrying signal at wavelength s and a continuous-wave (CW)
15
signal at the desired new wavelengthc (called the probe beam) are simultaneously
coupled into the device. The two waves can be either copropagating or counterpropagating.
However, the noise in the latter case is higher. The signal beam modulates the gain of the
SOA by depleting the carriers, which produces a modulation of the refractive index. When
the CW beam encounters the modulated gain and refractive index, its amplitude and phase
are changed, so that it now carries the same information as the input signal.
A limitation of the XPM architecture is that it only converts one wavelength at a

time. In addition, it has limited transparency in terms of the data format. Any information
that is in the form of phase, frequency, or analog amplitude is lost during the wavelength
conversion process. Consequently, this scheme is restricted to converting digital signal
streams.
(a)
CW signal
( c )
SOA1
Input signal ( s )
SOA2
Converted signal (c )
(b)
CW signal
( c )
SOA1
Input signal ( s )
SOA2
Converted signal (c )
Figure 2.2. (a)Mach-Zehnder interferometer and (b) Michelson interferometer setups using
a pair of SOAs for implementing the cross-phase modulation wavelength-conversion
scheme.
16
The second class is wave-mixing wavelength converters. Wavelength conversion
based on nonlinear optical wave mixing offers important advantages compared with other
methods. These include a multiwavelength conversion capability and transparency to the
modulation format. The mixing results from nonlinear interactions among optical waves
traversing a nonlinear material. The outcome is the generation of another wave whose
intensity is proportional to the product of the intensities of the interacting waves. The phase
and frequency of the generated wave are a linear combination of those of the interacting
waves. Therefore, the wave mixing preserves both amplitude and phase information, and,
consequently, is the only wavelength-conversion category that offers strict transparency to
the modulation format.
Two successful schemes are four-wave mixing (FWM) in either passive

waveguides or SOAs, and difference-frequency generation in waveguides. For wavelength
conversion, the FWM scheme employs the mixing of three distinct input waves to generate
a fourth distinct output wave. In this method, an intensity pattern resulting from two input
waves interacting in a nonlinear material forms a grating. Difference-frequency generation
in waveguides is based on the mixing of two input waves. Here, the nonlinear interaction
of the material is with a pump and a signal wave.
2.2 Bufferless WDM Switch with Shared TWCs
The considered optical switch architecture is shown in Figure 2.3. It has N

input/output fibers, each fiber supports a WDM signal with M wavelengths, so an input (or
output) channels is characterized by the couple (i,j) wherein i (i {1,,N}) identifies the
input/output fiber andj, (j {1,,M}) identifies the wavelength. [3]
The switch architecture is equipped with a number r of Tunable Wavelength

Converters (TWCs) shared among the input channels. At each input line, a small portion of
the optical power is tapped to the electronic controller not shown in the Figure 2.3. All
packets have a fixed size and their arrivals on each wavelength are synchronized on a
time-slot basis, and a time slot is the time needed to transmit a signal packet. The switch
control unit detects and reads packet headers and drives the space switch and the TWCs.
17
Incoming packets on each input line are wavelength demultiplexed (DMUX blocks in
Figure 2.3). Electronic control logic, on the basis of the routing information contained in
each packet header, handles packet contentions and decides which packets have to be
wavelength shifted. Packets not requiring wavelength conversion are directly routed
toward the output lines (e.g., the packet arriving at the input line 1 and wavelength 1 and
directed to the output 1, as shown in Figure 2.3); on the contrary, packets requiring
wavelength conversions will be directed to the pool of r TWCs and, after a proper
wavelength conversion, they will reach the output line (e.g., the packet arriving at the input
line N and wavelength1 and directed to the output line 1, as shown in Figure 2.3).
Input Fiber DM UX M UX Output Fiber
1 1
1 1
...
...
1 , 2 , L , M 1 , 2 , L , M
M M
...
...
1
Space 1
N switch N
...
...
1 , 2 , L , M 1 , 2 , L , M
M M
2
...
Shared TW Cs
r
Figure 2.3. Structure of bufferless WDM optical switch equipped with TWCs.
As a remark, it is to be noted that, the use of shared TWCs involves two main
drawbacks to be dealt with: 1) the enlargement of the switching matrix of a factor equal to
the number of used converters r (see Figures 2.3 and 2.4) the introduction of an additional
attenuation of the optical signal caused by the twice crossing of the switching matrix.
18
For example, Figure 2.4 shows that1 on which five packets arrive,2 on which
three packets arrive, and other two wavelengths, 3 and4, on which no packets arrive.
In the1 channel, one packet arrives the output on the1, and others have two packets
after conversion process which arrive the output on3 and4, and others will be lost. In
the2 channel, one packet arrives the output on the2, and others will be lost.
1 1
2 2
(a)
3 3
4 4
Conversion Process MUX Output Fiber
TWC
1
Arriving Packet
2
(b) Shified Arriving Packet
3
Lost Packet
4
Figure 2.4. Wavelengths of arriving packets (a) before and (b) after conversions.
2.3 Partially Shared Buffering for WDM Packet Switch
The considered optical switch architecture is shown in Figure 2.5. Each input and
output fiber consists of M different wavelengths. Each wavelength represents a unique
channel. To resolve the packet contention, an optical buffer, with the capacity C, is
included to each output fiber. These buffers are referred to as prime buffers. In addition, a
separate group of fiber delay lines forming a common buffer, with the capacity of C, is
allocated for all outputs to share. The buffer is referred to as shared buffer. All packets are
of fixed size representing typical ATM traffic. Each wavelength can independently carry
one packet at one time frame.[4]
19
Input Fiber DMUX Prime Buffer Output Fiber
1 1
1 1
...
...
1 , 2 ,L, M C 1 , 2 ,L, M
M
...
...
1
Space
N switch 1 N
...
...
1 , 2 ,L, M C 1 , 2 ,L, M
M
Shared Buffer
1
1
...
...
1 , 2 ,L, M C'
M
Figure 2.5. WDM Packet Switches with Partially Shared Buffering (PSB).
When a packet carried by a wavelengthj, (j {1,,M}) arrives from an input

fiber, it is demultiplexed by the DMUX with its address information in the packet header
processed electrically. This address is used to route the packet from the input to the
destined output. The packet will then be transmitted on the output fiber if the wavelength
j is available; otherwise, it will be queued at the output prime buffer for later transmission.
Generally, if n packets arrive at the same wavelength and are simultaneously destined to
the same output fiber, up to n fiber delay lines are need to store these packets. These
buffered packets will then be transmitted out to the output fiber in the subsequent time
frames.
If the prime buffers are fully occupied, a newly arriving packet will be blocked.
However, with the PSB, even when a prime buffer has been fully occupied, the overflow
packets can be routed to the shared buffer for temporary storage. Thus, a packet will not be
completely blocked until it fails in queuing in the shared buffer. Subsequently, the packets
stored in the shared buffer will recirculate back to an input port for further attempts (this
recirculation may be done more than once). Therefore, from the space switch perspective,
20
the N inputs are used to serve the recirculation packets. To solve the potential packet
sequence disordering problem, we always give the recirculation packets higher priority
against with other newly arriving packets, and let the end user handle sequence correction
if packet sequence disordering occurs.
The PSB without wavelength conversion architecture is shown in Figure 2.5. The
corresponding queuing model is illustrated in Figure 2.7(a). Each plane represents an
individual wavelength channel. Since each prime buffer consists of M logical buffers, each
plane therefore contains one logical buffer associated with a wavelength. Thus the rejected
packets always retain the same wavelength in all their attempts to be transmitted. Let us
assume a packet, carried by a wavelengthj, (j {1,,M}) arrive at an input fiber,
destined to an output fiber i (i {1,,N}). If no packet contention occurs, the packet will
be transmitted immediately, otherwise it queues in the prime buffer of the i-th output for
later transmission. However, if the buffer is full, the packet will be routed to the shared at
the same plane j for temporary storage, unless the shared buffer is also full.
DMUX TWC
1
M
Figure 2.6. Wavelength conversions for recirculated packets from the shared buffer.
The dashed block of Figure 2.5 is replaced by the new block shown in Figure 2.6.
Each demultiplexed wavelength in the dashed block is connected to a wavelength
converter before joining the space switch. Thanks to the input tunable wavelength
converters, the recirculation packets can be carried by any other wavelengths in its
destination output fiber. The queuing model illustrated in Figure 2.7(b). The recirculation
21
packets can be buffered across wavelength planes due to wavelength conversion.
(a)
1 new packets Prime Buffer

Prime Buffer 1
M M
Prime Buffer
new packets 2
1
2
N
overflow Shared Buffer

N
packets
packets
packets
(b)

2 new packets Prime
Prime Buffer
Buffer 1
M M
Prime Buffer
2
1
2
N
wavelength overflow
N
Shared Buffer
conversion packets Shared Buffer
Shared Buffer
M M
Figure 2.7. (a) Queueing model without TWCs; and (b) PSB model with TWCs.
2.4 Comparisons of Bufferless and Shared Buffer Schemes
A wavelength converter is needed to shift the wavelength of an optical packet. The

switch architecture must be designed in a cost-effective way such that only a minimum
number of wavelength converters are need. In addition, TWCs are active devices that
create noise. A switch control algorithm is required to reduce unnecessary TWCs.
22
The architecture for a bufferless optical packet switch making use of the
wavelength dimension for contention resolution of packets directed to the output line. The
optical switch architecture is provided with wavelength converters shared among the
output lines. An analytical model, allowing the dimensioning of the number of wavelengths
converters according to a required performance, has been presented. The results of the
analytical model fit very accurately with simulation ones. The model has been applied to
carry out a sensitivity analysis of the required number of converters as a function of the
main system parameters and traffic values. The proposed architecture allows us to save in
terms of employed number of converters with respect to the other architectures proposed in
literature [7].
The drawback of the delay-line buffer is that it is spacious and inconvenient. For
example, more than 200m fiber is needed to delay an optical packet for 1s. Also, the
delayed optical packet cannot be randomly retrieved. However, at present, there is no better
device to buffer optical packets than the fiber delay line.
The other architecture for an optical packet switch is with Partially Shared
Buffering. The architecture allocates a buffer to be shared by all outputs, in addition to the
statistical nature of packet arrivals, more efficient use is made of the shared buffer
resources. Therefore, a lower packet loss probability can be achieved without the need of
either employing wavelength converters. It only requires a very simple control algorithm,
and needs fewer recirculation loops (less than 2). Simulation studies demonstrate the
effectiveness of the PSB. The PSB with a limited number of wavelength converters that
simulation studies have demonstrated that, with a fewer number of wavelength converters
equipped in front of packet recirculation input, both the switch throughput and mean
packet delay can be further improved.
This chapter introduces two architectures, one is bufferless optical WDM switch
employing shared Tunable Wavelength Converters and the other is optical WDM switch
with Partially Shared Buffering. The optical WDM switch employing shared Tunable
Wavelength Converters discusses how to reduce packet loss probability and the number of
wavelength converters. As an example, it has 16 input and output fibers, each fiber
supports a WDM signal with 16 wavelengths, and the total load offered to each input line
is equal to 0.8; the architecture equips only 10 converters instead of 256 that is the number
23
of converters used in the main architecture proposed in the literature in which one
converter is dedicated to each input wavelength channel.
The optical WDM switch with PSB combines the advantages of the low packet loss
probability and optimal throughput delay performance. If a small number of converters are
added to the recirculation delay line, the performance can be improved further. Simulation
study demonstrates that the PSB can achieve a higher throughput without the need of either
dramatically increasing the size of each prime output buffer or heavily employing
wavelength converters.
24
Chapter 3. Shared TWCs for Buffering
WDM Packet Switch
In this chapter, a buffering WDM optical packet switch scheme employing the
wavelength converters for contention resolution is evaluated. The architecture is dedicated
to an individual buffer for each output and is equipped with tunable wavelength converters
(TWCs) shared among the input lines. Combined with the wavelength dimension for
contention resolution, this reduces the size of each output buffer and employs a lesser
number of wavelength converters; therefore, it is cost-effective. Based on the architecture,
lower packet loss probability can be achieved.
3.1 WDM Optical Packet Switch
The WDM optical switch architecture is shown in Figure 3.1. It has N input and
output fibers, each input/output fiber consists of M different wavelengths, 1, 2, , M, so
an input (or output) channel is characterized by the couple (i, j), where i identifies the
input/output fiber, i {1, 2, , N}, and j identifies the wavelength, j {1, 2, , M}.
Each wavelength represents a unique channel. We assume that all packets are of fixed size
representing typical ATM traffic. Each wavelength can independently carry one packet at
one time frame. Hereafter, the duration T of a time slot is assumed as the time unit.
To resolve the packet contention, an optical buffer with capacity B is included to

each output fiber. In addition, the switch architecture is equipped with a number r of TWCs
which are shared among the input channels. Therefore, the switch performs (N+r) (N+r)
switching. Without using TWCs, two fiber delay-lines are needed to store the two packets.
With using tunable converter, one of the packets can be converted to another wavelength
and stored in the same fiber. Therefore, the more wavelengths the more packets can be
stored in each delay-line buffer. Generally, if n packets arrive at the same wavelength and
are simultaneously destined to the same output fiber, up to n fiber delay lines are needed to
25
store these packets.
Input Fiber DMUX Buffer Output Fiber
1
1 1
...
...
1 , 2 , , M
... 1 , 2 , , M
...
M
...
...
1
Space
N switch N
...
...
1 , 2 , , M
... 1 , 2 , , M
M
...
2
...
Shared TWCs
r
Figure 3.1. Structure of the buffering WDM optical switch equipped with tunable
wavelength converters.
When a packet carried by wavelength j arrives from an input fiber, it is

demultiplexed by the DMUX (shown in Figure 3.1) with its address information in the
packet header processed electrically. The switch control unit detects and reads packet
headers and drives the nonblocking space switch matrix and the TWCs. This address is
used to route the packet from the input to the destined output. An electronic control logic
device, on the basis of the routing information contained in each packet header, handles
packet contentions and decides which packets have to be wavelength shifted.
In a conventional architecture, packets not requiring wavelength conversion are

directly routed toward the output k if the buffer in output k is not fully occupied by packets
arrival on the wavelength j. If the buffer in output k is fully occupied by packets arrival on
the wavelength j, a newly arriving packet will be blocked. On the contrary, as depicted in
26
Figure 3.2, the packets requiring wavelength conversions will be directed to the pool of r
TWCs and, after a proper wavelength conversion, they will reach the same output buffer. If
no available wavelength, packets will be lost. The buffered packets will then be transmitted
out to the output fiber in the subsequent frames. In this paper, we will analyze packet loss
probability on the discussed optical switch architecture. Also, the mean contention time in
output contention process, and the mean packet delay in output buffer will be analyzed.
1 new packets
. .. 2 Prime
Prime Buffer
Buffer 1
M Prime Buffer
2
... 1
...
2
N
1
2 overflow
N
M packets Shared Buffer
Shared Buffer
... ...
M 2 1
Shared TWCs
r
Figure 3.2. The queueing model with wavelength conversions.
The mathematical formula from analyzing the architecture will use MATLAB to
simulate the results. According to the conditions which we need, we give the parameter
value. For example, traffic load, the number of wavelengths, the number of inputs/outputs,
buffer size. We firstly evaluate the conditional probability (3.6); from the later analysis,
we will know that the usable number of conversions is form 0 to (BM-B). Using (3.3) and
(3.6), we will get (3.5). From previously, we will also know that h > (BM-B), Pwk(h) will
be 0. We use the Pwk(h) to further evaluate survivor function and packet loss
probability Ploss
( 3)
, and we will use diagram to express the outcomes. Among them, we will
27
use the number of converters as a major variable to observe packet loss probability Ploss
( 3)
; we
will discover that the number of converters is about the same value, the packet loss
probability will tend towards to a fixed value. Because that Pwk(h) will affect the usage
about the number of converters, and unnecessary converters which can not be used will
result in packet loss probability Ploss

( 3)
can not be further improved. Under different
parameters, we will get other different outcomes. Survivor function is direct relation to the
number of conversions, and therefore its conditions will be simpler.
3.2 Dimensioning the Number of TWCs
Denote the random variable Ak,j as the number of packet arrivals on wavelength j
and addressed to the output line k at a time-slot. At each time slot, a packet arrives on a
given channel with probability a0 and is addressed to each output line with probability 1/N,
the probability p Ak , j ( x) that x packets arriving on wavelength j and are directed to the
output line k, therefore, has the binomial probabilities
p Ak , j ( x) = Pr{Ak , j = x}
x N x x = 0, 1, , N. (3.1)
N a a
= o 1 o
x N N
On the basis of Eq. (3.1), we will calculate the probability function Pwk(h) of the
number wk of wavelength conversions required by a single output line k. The evaluation of
Pwk(h) is an intermediate step needed to have an expression for the packet loss probability.
Evaluations of the packet loss probability are examined in Section 4.
Let k denote the number of wavelengths on which no packets arrive at the output
line k in a given time-slot (For example, we have k equal to 2 in Figure 3.3, because there
are two wavelengths, 3 and 4, on which no packet arrive). Therefore, k can be expressed
as
28
M
k = (Ak ,i ) (3.2)
i =1
where (x) is the discrete impulse. The probability function of k can be easily evaluated
by means of the expression
M
Pr{ k = j} = (1 bk ,i ) (bk ,i )
Mj j
(3.3)
j
where
N
a
bk ,i = Pr{ ( Ak ,i ) 0} = 1 o (3.4)
N
Let Pwk(h) be the probability function of the number wk of the employed TWCs.
Note that wk is the number of conversions required by a single output buffer k. By applying
the total probability theorem we get that
M
Pwk (h) = Pr{wk = h | k = j}Pr{ k = j} (3.5)
j =0
The conditional probability in Eq. (3.5) can be evaluated to be
Pr{wk = h | k = j}
0 {Ik < h}

Pr ( Ak , p B) jB | Ak , p 1 p = q1, q2 ,..., qM j {Ik = h} and ( j M ) (3.6)
p=q1 ,q2 ,...,qM j
=
Pr
{Ik (h +1)} and { j (M 1)}
+

p=q ,q ,...,q ( A k , p B) = h | Ak , p 1 p = q1, q2 ,..., qM j

1 2 M j
(h) { j = M}
where B is buffer size. The term I k , p = (A k,p

p = q1 , q1 ,L, q M j
B) denotes the total acceptably
29
incoming number of packet arrival on the wavelengths that no full of output buffer k had
packet(s) and can accept incoming number of packet arrival on the same wavelength. The
symbol ( . )- denotes the smaller value of 0 or the argument inside the parentheses.
Figure 3.3(a) shows the definition of optical buffering on the arriving wavelength
packets. The buffer k has packets arriving on the wavelength 1 and 2 and has no packets
arriving on the wavelength 3 and 4. Because the buffer has no packets on the wavelength
3 and 4, the conditions will not be discussed further. The buffer k has 5-packets arriving
on wavelength 1 and is over the buffer size; therefore, the buffer will have no vacancies.
The buffer k has 1-packet arriving on the wavelength 2 and the buffer size is 2; therefore,
the buffer will have one vacancy. Totally, the buffer k has one vacancy according to the
above definition.
Figure 3.3(b) shows the definition of optical buffering on the arriving wavelength
packets. The buffer k has packets arriving on the wavelength 1 and 2 and has no packets
arriving on the wavelength 3 and 4. Because the buffer has no packets on the wavelength
3 and 4, the conditions will not be discussed further. The buffer k has 5-packets arriving
on wavelength 1 and is over the buffer size; therefore, the buffer will have no vacancies.
The buffer k has 4-packet arriving on the wavelength 2 and the buffer size is 2; therefore,
the buffer will have zero vacancy. Totally, the buffer k has zero vacancy according to the
above definition.
(a)
(A

Ik,p = B) = 1
k,p
p = q1 , q2 ,... q M j
Logic Buffer with capacity
B=2
1 1
2 2
k
3 3
4 4
Conversion Process Output Fiber
30
(b)
(A

Ik, p = k,p B) = 0 Logic Buffer with capacity
p =q1 , q2 ,... qM j
B=2
1 1
2 2
k
3 3
4 4
Figure 3.3. A buffer size of B=2 with arriving packet(s) on wavelength 2.
The term I k = jB (A k,p

p = q1 , q1 ,L, q M j
B) denotes the total acceptably incoming
number of packet in output buffer k. To understand the meaning of Eq. (3.6), notice that
when k is equal to j, there are at most I k wavelengths on which the packets can be
shifted and hence the number wk of conversions cannot be greater than I k . In order to have
a number wk of conversions equal to I k , the number of packets arriving on the M-j
remaining wavelengths q1, q2, , q(M-j) must be greater than BM. The term
p = q1 , q 2 ,L, q M j
Ak , p is the total number of packets in the output buffer k.
Figure 3.4 and Table I illustrate the above case with M = 4, k = 2, wk = 4,
p = q1 , q2
Ak , p = 9 , and B = 2. In Figure 3.4, the buffer k has 5-packets arriving on the
wavelength 1 and 4-packets arriving on the wavelength 2 and no packets arriving on the
wavelength 3 and 4. This illustrates the condition of I k , p = 0 . The 5-packets on
wavelength 1 and the 4-packets on wavelength 2 are both over the buffer size. Therefore,
1 and 2 will respectively have 2-packets to enter the buffer k without using converters.
The other 3-packets arriving on wavelength 1 will randomly have 2-packets converted to
wavelength 3 and 1-packet converted to wavelength 4, totally with the usage of three
converters. One of the other 2-packets arriving on wavelength 2 will be randomly
31
converted to wavelength 4 with the usage of one converter and the other packet will be
lost.

B=2
1 1
2
(a) 2 k
3 3
4 4
TWC
1
Arriving Packet
2
(b) Shified Arriving Packet k
3
Lost Packet
4
Figure 3.4. Arriving packets wavelengths (a). before and (b). after conversion processes.
(M = 4, k = 2, wk = 4, p = q1 , q2
Ak , p = 9 , B = 2).
Table I shows another example. The buffer k has 5-packets arriving on the
wavelength 1 is over the buffer size, and the 1-packet on wavelength 2 is under the buffer
size. Therefore, 1 will have 2-packets and 2 will have 1-packet to enter the buffer k
without using converters. Five of the other 6-packets arriving on wavelength 1 will
randomly have 1-packet converted to wavelength 2, 2-packets converted to wavelength 3,
and 2-packets converted to wavelength 4, totally with the usage of five converters. The
remaining 1-packet will be lost.
32
Table I: M = 4, k = 2, B = 2, total number of packets BM.
1 2 A k,p
p = q1 , q 2 ,...q M j
Ik,p Ik wk Packet
Lost
number
7 3 7+3=10 0+0=0 4+0=4 4 2
8 2 8+2=10 0+0=0 4+0=4 4 2
9 1 9+1=10 0+1=1 4+1=5 5 2
5 4 5+4=9 0+0=0 4+0=4 4 1
7 2 7+2=9 0+0=0 4+0=4 4 1
8 1 8+1=9 0+1=1 4+1=5 5 1
4 4 4+4=8 0+0=0 4+0=4 4 0
6 2 6+2=8 0+0=0 4+0=4 4 0
7 1 7+1=8 0+1=1 4+1=5 5 0
In order to have a number wk of conversions less than I k , ( j M ) , it is needed

that (M j )B + I k , p + h packets arrive on wavelength q1, q2, , q(M-j). Among these,
A k,p
p = q1 , q2 ,..., q M j
(A k, p
p = q1 , q 2 ,..., q M j
B) + of wavelengths will be forwarded without conversions on
the same wavelengths, the remaining h will be shifted on h of the I k available.
Figure 3.5 and Table II illustrate the above case with M = 4, k = 2, wk = 3,
p = q1 , q2
Ak , p = 6 , and B = 2. In Figure 3.5, the buffer k has 5-packets arriving on the
wavelength 1 and 1-packet arriving on the wavelength 2 and no packets arriving on the
wavelength 1 is over the buffer size and the 1-packet on wavelength 2 is under the buffer
size. Therefore, 1 will have 2-packets and 2 will have 1-packet to enter the buffer k
without using converters. The other 3-packets arriving on wavelength 1 will randomly
33
have 1-packet converted to wavelength 2 and 2-packets converted to wavelength 3,
totally with the usage of three converters. The above wavelength conversions scheme
assures that no packets will be lost.

B=2
1 1
2
(a) 2 k
3 3
4 4
TWC
1
Arriving Packet
2
(b) Shified Arriving Packet k
3
Lost Packet
4
no lost packet
Figure 3.5. Conversion process when the arriving packet is less than BM.
(M = 4, k = 2, wk = 3, p = q ,q Ak , p = 6 , B = 2).
1 2
Table II shows another example. The buffer k has 4-packets arriving on the
wavelength 1 is over the buffer size, and the 3-packets on wavelength 2 is under the
buffer size. Therefore, 1 and 2 will respectively have 2-packets to enter the buffer k
without using converters. The other 2-packets arriving on wavelength 1 are converted to
wavelength 3 with two converters. The other 1-packet arriving on wavelength 2 will be
randomly converted to wavelength 4 with one converter. In this way, the wavelength
34
conversions scheme assures that no packets will be lost.
Table II: M = 4, k = 2, B = 2, total number of packets < BM.

1 2 A k,p
p = q1 , q 2 ,...q M j
Ik,p (A k,p
p = q1 , q2 ,..., q M j
B) + wk Packet
Lost
number
4 3 4+3=7 0+0=0 2+1=3 3 0
5 2 5+2=7 0+0=0 3+0=3 3 0
6 1 6+1=7 0+1=1 4+0=4 4 0
3 3 3+3=6 0+0=0 1+1=2 2 0
4 2 4+2=6 0+0=0 2+0=2 2 0
5 1 5+1=6 0+1=1 3+0=3 3 0
2 3 2+3=5 0+0=0 0+1=1 1 0
3 2 3+2=5 0+0=0 1+0=1 1 0
4 1 4+1=5 0+1=1 2+0=2 2 0
In the above description of Figures 18 and 19, together with Tables I and II, we
have emphasized the condition of j M . For the condition of j = M, there is no packet
arrival, and the needed number of conversions is zero.
With the chaining substitutions of Eqs. (3.3), (3.4), and (3.6) into Eq. (3.5), we can
express the probability function Pwk(h) of the number wk of conversions required by an
output buffer k into the following:
M h
Pwk (h) = bk,i (1bk,i )M(hIk, j ) / B Pr (Ak, p B) jB| Ak, p 1 p = q1,q2,...,qM j

(h Ik, j )/ B p=q1,q2 ,...,qM j
(3.7)
M1
M
+ bkj,i (1bk,i )M j Pr (Ak, p B)+ = h | Ak, p 1 p = q1,q2,...,qM j +(h) (1bk,i )M
j=(h+1Ik, j ) / B j p=q1,q2 ,...,qM j
35
3.3 Evaluation of packet loss probability
We will evaluate the packet loss probability Ploss of the buffering optical switch
with respect to the employed number of TWCs. Different configurations of wavelength
conversions, including full-conversion, no-conversion, and medium-conversion, are all
under our examinations. Numerical results measured by mean packet loss probability
versus buffer size, traffic load, number of converters, number of inputs/outputs, and
number of wavelengths, are then illustrated for the performance evaluations.
3.3.1 Evaluation of Ploss in Full-Conversion Configuration (r = NM):
The packet loss probability Ploss

(1)
can be expressed as
E[ N lk ]
Ploss
(1)
= (3.8)
E[ N ok ]
where E[x] denotes the expected value of the random variable x.
Here, N ok is the received number of packets offered to the output buffer k :
E[ N ok ] = Ma0 , (3.9)
and N lk is the lost number of packets offered to the output buffer k:
NM
E[ N lk ] = E[ N lk | Ak = i ] Pr{Ak = i} . (3.10)
i =0
In Eq. (3.10), Ak = j =1 Ak , j is the total number of packets arriving at output line

M
k, and it has the following probability function:
36
M M
Pr{Ak = i} = Pr Ak , j = i = p Ak , j ( x) x =i , (3.11)
j =1 j =1
M
where denotes the convolution operator applied M times and p Ak , j ( x) is given by
j =1
(3.1).
The term E[ N lk | Ak = i ] appeared in Eq. (3.10) is given by
0 if 0 i BM
E[ N lk | Ak = i ] = (3.12)
i BM if BM < i NM
Expression (3.12) is explained by taking into account that, according to the conversion
process, the output buffer k is able to send arriving packets BM per time-slot. Hence a
packet-loss happens just when Ak > BM, and the number of packet losses is equal to Ak -
BM.
From (3.8) ~ (3.12), we obtain the following expression for the packet loss
probability Ploss
(1)
with full wavelength conversions:
1 NM
x N x
M N
ao ao
P = (i BM ) 1 x =i
(1)
loss (3.13)
Mao i = BM +1 j =1 x
N N
3.3.2 Evaluation of Ploss in No-Conversion Configuration (r = 0):
As no conversions are accomplished and because we have assumed that the traffic
is uniform on the various wavelengths; therefore, it is sufficient to calculate the Ploss of the
packets addressed to a generic output buffer (k, j). According to the received and the lost
number of packets offered to the output buffer k, we can write
37
E[ N lk , j ]
Ploss
( 2)
= . (3.14)
E[ N ok , j ]
Here, N ok , j is the received number of packets offered to the output buffer k with
wavelength j,
E[ N ok , j ] = a 0 ; (3.15)
and N lk , j is the lost number of packets offered to the output buffer k,
E[ N lk , j ] = E[ N lk , j | Ak , j = i ] Pr{Ak , j = i}.
N
(3.16)
i =0
At each time-slot, only one packet can be sent to the output buffer k with
wavelength j and hence
0 if 0 i B
E[ N lk , j | Ak , j = i ] = (3.17)
i B if B < i
Expression (3.17) is explained by taking into account that, according to the conversion
process, the output buffer k is able to send arriving packets B on the wavelength j per
time-slot. Hence a packet-loss phenomenon happens just when Ak,j > B, and the number of
packet losses is equal to Ak,jB.
From (3.14) ~ (3.17), together with (3.1), we obtain the following expression for
the packet loss probability Ploss

( 2)
with none of wavelength conversions:
1 N N ao ao
x N i
P = (i B) 1 ,
( 2)
loss (3.18)
ao i = B +1 x N N
38
3.3.3 Evaluation of Ploss when r TWCs are utilized (0 < r < NM):
In general, with a number r of TWCs, 0 < r < NM, shared among NM input/output
wavelength channels, the packet loss probability can be expressed as
E[ N lk ]
Ploss
( 3)
= . (3.19)
E[ N ok ]
Here, N ok is the received number of packets offered to the output buffer k, as given by
(3.9), and N lk is the lost number of packets offered to the output buffer k:
r
E[ N lk ] = E[ N lk | C k = i ] Pr{C k = i} . (3.20)
i =0
In Eq. (3.20), note that Ck is the available number of TWCs to shift the packets addressed
to the output buffer k:
N
Pr{C k = i} = Pr{C k = i | Dk = j}Pr{Dk = j} , (3.21)
j =1
where Dk is the random variable indicating the selection order of the output buffer k by the
manager of the pool of shared TWCs.
If we form the hypothesis that the selection is accomplished in accordance with

uniform distributions, we get
Pr{Dk = j} = 1/N, j = 1, 2, , N. (3.22)
The conditioned probability appeared in (3.21) is given by
r
Pr{C k = i | Dk = j} = Pr{C k = i | Dk = j , S = p}Pr{S = p | Dk = j} (3.23)
p =0
39
where the random variables conditioned to the event Dk = j are the number of TWCs
utilized by the first j selected output buffers whose probability function is easily evaluated
by means of probability function Pwk(h) of the number of conversions wk, k = z1, z2, , zj.
Note that k = z1, z2, , zj represent the indexes of the pool of shared TWCs, before the
output buffer k is considered. By assuming independence of the random variables wks, we
can get
pwh ( x) x = p if p < r
h = z1 ,Kz j

Pr{S = p | Dk = j} = (3.24)
j( BM B )
pwh ( x) x = p if p = r
p = r h = z1 ,Kz j
Obviously, when S is equal to p, the number of TWCs that can be employed by the
output buffer k is equal to the difference between r and p when r > p. On the contrary, if r <
p, no TWC is available for conversions for the packets arriving to the considered output
buffer k. Hence we have
Pr{Ck = i | Dk = j, S = p} = [i - max(0, r-p)] (3.25)
In order to evaluate the conditioned expected value of N lk appeared in (3.20), we
express it as follows:
E[ N lk | C k = i ]
M
= E[ N lk | C k = i, k = j ] Pr{ k = j | C k = i} (3.26)
j =0
M
= E[ N lk | C k = i, k = j ] Pr{ k = j}
j =0
The last equation follows from the fact that independent variables Ck and k are assumed.
40
The term E[ N lk / C k = i, k = j ] can be evaluated according to the remark done in
Section 3 as shown in (3.27). Based on the chaining evaluations of Eqs. (3.20) ~ (3.26), the
packet loss probability of Eq. (3.19) can show in (3.28).
(M j)N
E[ N | C k = i, k = j ] =
l
k
p =0

Pr Ak , q = p | Ak ,q 1 q = h1 , K , h M j
q = h1 ,K, hM j
max( p [ B ( M j ) I k , q ] min([ jB + I k ,q ], i ), 0 )
(M j) N
= ( p [B(M j) I
p =[ B ( M j ) I k , q ]+ min([ jB + I k , q ], i ) =1
k ,q ] min([ jB + I k ,q ], i )) (3.27)

Pr Ak ,q = p | Ak , q 1 q = h1 , K , h M j
q = h1 ,K, hM j
r M
Pr { = u}
1
Ploss
(3)
= k
NMa o i=0 u =0
(M j)N
( v [ B ( M u ) I k , q ] min([ uB + I k , q ], i ))
p = [ B ( M u ) I k , q ] + min([ uB + I k , q ], i ) =1
(3.28)

Pr A k , q = v | A k , q 1 q = h1 , K , h M u
q = h1 ,K , h M u
N
r 1 j ( BM B )

(i max( 0 , r t ) ) pw h ( x ) x =t + (i ) pw h ( x ) x=t
j = 1 t =1
h = z1 ,K , z j h = z1 ,K , z j
t=r
3.4 Numerical Results of Packet Loss Probability
In the following, we evaluate the performance of buffering WDM packet switch

architecture by numerical analyses. We assume that the total load offered to each input
buffer is equal to q and hence, if M wavelength channels are used, as the traffic is
uniformly distributed on each wavelength channel, we have offered load per wavelength
channel a0 = q/M. The performances are measured by mean packet loss probability versus
buffer size, traffic load, number of converters, number of inputs/outputs and number of
wavelengths.
41
Figure 3.6 illustrates the packet loss probability versus number of converters. The
traffic load q to each input fiber is set to 0.8. The number of inputs/outputs is N = 16 and
the number of wavelengths is respectively M = 4 and 6. The buffer size is chosen to be B =
1 and 2, respectively. Note that the buffer size of B=1 corresponds to the bufferless
architecture described in [3].
number of converters (r)

1.0E+00
1.0E-01 0 1 2 3 4 5 6 7 8 9 10
1.0E-02
1.0E-03
packet loss probability
1.0E-04
1.0E-05
1.0E-06
1.0E-07
B=1, M=4
1.0E-08
B=2, M=4
1.0E-09
B=1, M=6
1.0E-10
B=2, M=6
1.0E-11
1.0E-12
Figure 3.6. Packet loss probability of the analytical switch model

when N=16, q=0.8, M= 4, 6, and B=1 (no buffer), 2.
From Figure 3.6, it can be seen that the packet loss probability can be reduced
dramatically with optical buffer. The larger the buffer size, the lower the packet loss
probability. For example, B=2, M=4, N=16 uses no converter and its packet loss probability
is 0.00504381; B=1, M=4, N=16 uses three converters and its packet loss probability is
0.00559564. Therefore, by using optical buffering we can get a better packet loss
probability.
42
For B=1, M=6, N=16, the packet loss probability is Ploss
(1)
= 2.39x10-5. For B=2,
M=6, N=16, it uses two converters and the packet loss probability is Ploss
( 3)
= 3.50x10-6. We
therefore use buffer size 2 and 2 converters for B=2, M=6, and N=16 that the packet loss
probability is lower than Ploss

(1)
for B=1, M=6, and N=16. Ploss
( 3)
for B=1, M=6, N=16, using
7 converters, the same with Ploss

(1)
for B=1, M=6, N=16. Hence we can reduce the number
of wavelength converters from 7 to 2. This reduction saves the system cost. Besides, it has
been shown that fiber delay lines are more effective in solving contention than wavelength
conversion.
Figures 3.7 and 3.8 show the packet loss probabilities versus the traffic load when
the number of input/output fibers is increased from N = 16 to 24. The traffic load q to each
input fiber is set to 0.8 and 0.6, respectively. It is seen that the packet loss probability Ploss
(1)
is similar for the cases of N=16 and N=24. Figure 3.9 shows that when traffic load is lower,
the packet loss probability is also lower for N=16.

1.E+00
0 1 2 3 4 5 6 7 8 9 10
1.E-01
1.E-02
1.E-03
1.E-04
1.E-05
B=1, M=4, N=16
1.E-06 B=1, M=6, N=16
B=2, M=4, N=16
1.E-07 B=1, M=4, N=24
B=2, M=4, N=24
1.E-08 B=1, M=6, N=24
Figure 3.7. Packet loss probability versus number of converters for q=0.6.
43
number of converter (r)
1.E+00
0 1 2 3 4 5 6 7 8 9 10
1.E-01
1.E-02
1.E-03
1.E-04
B=1, M=4, N=16

1.E-05
B=1, M=6, N=16
B=2, M=4, N=16
1.E-06 B=1, M=4, N=24
B=2, M=4, N=24
B=1, M=6, N=24
1.E-07
Figure 3.8. Packet loss probability versus number of converters for q=0.8.

1.E+00
0 1 2 3 4 5 6 7 8 9 10
1.E-01
1.E-02
1.E-03
1.E-04
1.E-05
1.E-06
1.E-07
1.E-08
q=0.8, B=1, M=4
1.E-09 q=0.8, B=1, M=6
1.E-10 q=0.8, B=2, M=4
q=0.8, B=2, M=6
1.E-11 q=0.6, B=1, M=4
1.E-12 q=0.6, B=1, M=6
q=0.6, B=2, M=4
1.E-13 q=0.6, B=2, M=6
1.E-14
Figure 3.9. Packet loss probability versus number of converters for N=16.
44
Chapter 4. Evaluation of mean total delay
and survivor function
In this chapter, we will evaluate mean total delay and survivor function of the
buffering optical packet switch equipped with shared TWCs. Output contention and output
buffer processes are both under our consideration [11, 12]. Studies indicate that the output
contention process can be modeled by a BI/D/B/N queueing process and the output buffer
can be modeled by a G[x]/D/1/B queueing process. Numerical results measured with
respect to buffer size, traffic load, and number of converters, are finally illustrated for the
performance evaluations.
4.1 Numerical Results of Mean Total Delay
Mean total delay is the sum of mean contention time in output contention process
and mean packet delay in output buffer queue. Our discussion of mean total delay is,
therefore, firstly on the output contention and the output buffer queueing processes. System
parameter of traffic loads is then numerically used to evaluate the mean total delay.
The mathematical formula from analyzing the architecture will use MATLAB to
simulate the results. According to the conditions which we need, we give the parameter
value. For example, traffic load, the number of wavelengths, the number of inputs/outputs,
buffer size. The mean contention time in output contention process and the mean packet
delay in output buffer process can be modeled by BI/D/B/N and G[x]/D/1/B individually.
We will get the state transition matrix of this Markov chain individually; we know that the
state probability vector has attained its limiting value and the state probabilities sum to one
over all states; therefore, the state probability can be solved numerically individually. We
will use the state probability to evaluate the mean contention time and the mean packet
delay.
45
4.1.1 Output Contention Process (BI/D/B/N queue)
Consider a tagged packet newly joining head of line (HOL). Now it joins the
contention for the transmission to output buffer k, with other packets destined to the same
output, found in HOL. This contention arises whenever the number of such packets in HOL
is more than B. Hence, the contention process to output k is characterized by
C k' , j = (C k , j B) + + Ak , j (4.1)
where the symbol ( . )+ denotes the larger value of 0 or the argument inside the parentheses.
Other terms are defined as follows:
Ak,j : the number of packets arrival on the wavelength j newly joining the
contention at the current time slot, which will complete for the next time slot.
Ck,j : the number of contending HOL packets arrival on the wavelength j destined to
output buffer k in the current time slot.
C k' , j : the number of packets arrival on the wavelength j destined to output buffer k
for the next time slot.

B: the switch fabric randomly selects up to B such packets arrival on the wavelength
j destined to output buffer k.
Let ak,j,x = Pr{Ak,j = x}. The contention process is modeled by a BI/D/B/N queue,
where BI represents Bernoulli process and at most N packets can arrive to the optical
switch. We will get the state transition matrix of this Markov chain form the BI/D/B/N
queue. The state transition matrix Rk,j of this Markov chain
46
N 1

0 a k , j ,0 a k , j ,1 ak , j,2 a k , j ,3 L a k , j , N 1 1 a
k , j,x
x=0

M M M M M O M M

N 1
B 1 a k , j ,0

a k , j ,1 ak , j,2 a k , j ,3 L a k , j , N 1 1
x=0
ak , j,x

N 1
B a k , j ,0

a k , j ,1 ak , j,2 a k , j ,3 L a k , j , N 1 1 ak , j,x

Rk , j =
x=0
N 2
B +1 0 a k , j ,0 a k , j ,1 ak , j,2 L ak , j,N 2 1 a k , j,x
x=0
N 3

B + 2 0 0 a k , j ,0 a k , j ,1 L ak , j,N 3 1 a k , j,x
x=0

M M M M M O M M
B 1
N 0 0 0 0 L a k , j , B 1 1 ak , j,x
( N + 1 ) ( N + 1 )
x=0 (4.2)
We know that the state probability vector has attained its limiting value, and the
state probabilities sum to one over all states; therefore, the state probability can be solved
numerically. Let pk,j,x = Pr{Ck,j = x}, x = 0, 1, 2, , N, or in a row vector form
Pk,j = (pk,j,0, pk,j,1, pk,j,2, , pk,j,N). (4.3)
which can be derived from
Pk,j = Pk,jRk,j, Pk,jeT = 1 (4.4)
e T is a column unit vector. From this equation system, Pk,j can be solved numerically. The
mean contention time in the contention process (the packets arrival on the wavelength j
destined to output buffer k) is therefore
N
E [Wc ] = x p k , j , x / a o (4.5)
x =1
where Wc is the contention time in the contention process.
4.1.2 Output Buffer Process (G[x]/D/1/B queue)
The output buffer process is characterized by a discrete-time G[x]/D/1/B queue with

batch arrival [13]. The output queue process obeys
47
Qk' , j = (Qk , j 1) + + Ok , j . (4.6)
where the symbol ( . )+ denotes the larger value of 0 or the argument inside the parentheses.
Other terms are defined as follows:
Ok,j: the number of packets arrival on the wavelength j up to B at the current time
slot, which will complete for the next time slot.
Qk,j: the number of packets arrival on the wavelength j in the output buffer k
including the one being served at the current time slot.
Qk' , j
: the number of packets arrival on the wavelength j in the output buffer k for the
next time slot.
We will get the state transition matrix of this Markov chain form the G[x]/D/1/B
queue. Denote by Ek,j the state transition matrix of this Markov chain, which has the
following upper triangular form
0 p
B 1
k , j ,0 pk , j ,1 pk , j , 2 pk , j ,3 L pk , j , B 1 1 pk , j , x
x=0
B 1

1 p pk , j ,1 pk , j , 2 pk , j ,3 L pk , j , B 1 1 pk , j , x
k , j ,0 x=0

B2
2 0 pk , j ,0 pk , j ,1 pk , j , 2 L pk , j , B 2 1 pk , j , x (4.7)
x=0
B 3
Ek , j = 3 0 0 pk , j ,0 pk , j ,1 L pk , j , B 3 1 pk , j , x
x =0

M M M M M M M M

B 1 0
1
0 0 0 pk , j ,0 pk , j ,1 1 pk , j , x
x=0
B 0 0 0 0 0 pk , j , 0

1 pk , j , 0
( B +1) ( B +1)
We know that the state probability vector has attained its limiting value, and the
state probabilities sum to one over all states; therefore, the state probability can be solved
numerically. Let vk,j,x = Pr{Qk,j = x}, x = 0, 1, 2, , B, or in a row vector form
48
Vk,j = (vk,j,0, vk,j,1, vk,j,2, , vk,j,B). (4.8)
which can be derived from
Vk,j = Vk,jEk,j, Vk,jeT = 1 (4.9)
Here we may write a numerical solution

v k , j ,0 = (1 ao )

(1 pk , j ,0 )
v k , j ,1 = vk , j ,0 (4.10)
p
(1 pk , j ,1 )
k , j , 0
p k , j , x 1 x 1 p
vk , j ,0
k , j, y
v k , j , x = v k , j , x 1 vk , j , x y for B x 2
p k , j ,0 p k , j ,0 y =2 p k , j ,0
Once in the queue process, there are x+1 queueing packets (including the served
packet itself) while in state x, the possibility for the packet to leave the buffer is given by
(x+1)-1. Therefore, the delay will have the distribution of transition time in a transient
Markov chain beginning at the initial state until absorption. The state of this transient
Markov chain is Qk,j.
Figure 4.1 gives the state transition diagram of contention process for output buffer
k. The initial state of this transient Markov chain is assigned according to the probability
vector Vk,j, representing the current state of Qk,j when the served packet moves to output k.
While in state x, the possibility of absorption from this chain is (x+1)-1, and the possibility
of moving to state l is given by [1-(x+1)-1] pk,j,l-x+1. We can use the transition diagram of
Figure 4.1 for this calculation as long as we never try to calculate the expected delay at an
absorption state. We can get the transient Markov chain Hk,j that delete all rows and
columns corresponding to absorption states. Therefore, the state transition matrix for this
transient Markov chain H k , j is expressed by
49
0 0 0 0 0 L 0 0

1 1 1 1 1 1 B1
1 2 pk, j,0 pk, j,1 pk, j,2 pk, j,3 L pk , j,B1 1 pk, j, x
2 2 2 2 2 x=0
2 2 2 2 2 B2
2 0 pk, j,0 pk, j,1 pk , j,2 L pk, j, B2 1 pk , j, x (4.11)
3 3 3 3 3 x=0

[ 1
]
Hk, j = diag1 (l +1) Ek, j = 3 0 0
3
4
pk, j,0
3
4
pk, j,1 L
3
4
pk, j,B3
3 B3
4 x=0

1 pk , j, x

M M M M M M M M

B 1 B 1 B 1 1
B 1 0 0 0 0 pk, j,0 pk, j,1 1 pk, j, x
B B B x=0

B 0 0 0 0 0
B
pk, j,0
B
(1 pk, j,0 )
B +1 B +1 ( B+1)( B+1)
diag[1-(l+1)-1] denotes a diagonal matrix with its lth diagonal element equal to 1 (l + 1) 1
for l = 0,1,2,,B.
1
pk , j , 4
2
1 2
1 pk , j , 3 pk , j , 3
pk , j ,1 2 3
2 1 3 3
pk , j ,1 pk , j ,1 pk , j , 2
3 4 4
1 1 2
0 pk , j , 0 1 pk , j , 2 2 pk , j , 2 3
2 2 3
1 1 2 1 3 1 4
pk , j , 0 pk , j , 0 pk , j , 0
2 3 3 4 4 5
absorption
Figure 4.1. State transition diagram of contention process for output buffer k.
With reference to [14], the mean delay time spent in the transient process is related
with the starting probability vector Vk,j and the transmission matrix [I - Hk,j]-1 of the
transition diagram. The row vector Vk,j specifies the starting probability in each state of the
process. The mean packet delay (the packets arrival on the wavelength j in the output
buffer k) is therefore
E[Wb] = Vk,j [I - Hk,j]-1eT , (4.12)
50
where Wb is the packet delay in the output buffer.
From mean contention time (4.5) and mean packet delay (4.12), we can get mean
total delay. In the following, numerical results measured by mean total delay are illustrated
for the performance evaluations. We assume that the total load offered to each input buffer
is equal to q and hence, if M wavelength channels are used, as the traffic is uniformly
distributed on each wavelength channel, we have offered load per wavelength channel a0 =
q/M. For the sake of simplicity, we ignore the wavelength conversion delay and
propagation delay. We only focus on the mean contention time and the mean packet delay.
Figure 4.2 shows that mean total delay versus traffic load q. From Figure 4.2, the
lower traffic load has the lower mean total delay. The number of inputs/outputs increase
form N = 16 to 24 for the same traffic load, they have similar outcome; therefore, the
number of inputs/outputs is not the major cause for affecting the mean total delay. The
number of wavelengths reduces from M = 6 to 4 under the same traffic load, and the mean
total delay is increasing. It is seen that the lower buffer size will have the longer mean total
delay.
1.11
N=16 M=6 B=2
1.1 N=16 M=4 B=2
N=16 M=6 B=4
1.09 N=16 M=4 B=4
mean total delay
N=24 M=6 B=2

1.08 N=24 M=4 B=2
N=24 M=6 B=4
1.07
N=24 M=4 B=4
1.06
1.05
1.04
1.03
0.4 0.6 0.8
traffic load (q )
Figure 4.2. Mean total delay versus traffic load q.
51
4.2 Numerical Results of Survivor Function
We will evaluate the survivor function of the total number of conversions
W = k =1 wk by assuming the variables wks (k = 1, 2, , N) to be statistically

N
independent. By survivor function, we mean the probability of conversions that are

survived in the shared TWCs. The analytical value of survivor function fW(w) was defined
in [3], and we quote it here for our further examination:
f W
( w ) = Pr {W > w}
N (4.13)
= Pr w k > w
k =1
N
=
i = w +1
Pw k ( x )
k =1
x =i
where Pwk(h), as given in Eq. (3.7), is the probability function of the number wk of the
employed TWCs, and w is the number of conversions used by the architecture.
Figures 4.3 and 4.4 show that survivor function versus the number of wavelength
conversions for using buffers is lower than that not using buffers. The number of
wavelengths is set to 4 and 6, respectively. Besides, we see that the higher number of
inputs/outputs has the higher survivor function under the same number of conversions. The
higher number of conversions has the lower survivor function under the same number of
inputs/outputs.
Figures 4.5 and 4.6 show the survivor functions versus the number of wavelength
conversions for buffer sizes of B = 1 and 2, respectively. It is seen that lower traffic load
will have lower survivor function. Figure 4.6 show that the survivor function for q = 0.8, M
= 6, N = 24 is similar for q = 0.6, M = 4, N = 24 and q = 0.8, M = 6, N = 16 is similar for q
= 0.6, M = 4, N = 16 under using buffer size B = 2.
52
number of conversions (w )
1.00E+00
1.00E-01 0 1 2 3 4 5 6 7 8 9 10
1.00E-02
survivor function of number of conversions
1.00E-03
1.00E-04
1.00E-05
1.00E-06
1.00E-07
1.00E-08
B=1 N=16
1.00E-09
1.00E-10 B=1 N=24
1.00E-11 B=2 N=16
1.00E-12 B=2 N=24
1.00E-13
1.00E-14
1.00E-15
Figure 4.3. Survivor function versus the number of conversions for M=4 and q=0.8.
1.E+00
1.E-01 0 1 2 3 4 5 6 7 8 9 10
1.E-02
1.E-03
survivor function of number of conversions
1.E-04
1.E-05
1.E-06
1.E-07
1.E-08
1.E-09
B=1 N=16
1.E-10
1.E-11 B=1 N=24
1.E-12
B=2 N=16
1.E-13
1.E-14 B=2 N=24
1.E-15
1.E-16
1.E-17
Figure 4.4. Survivor function versus the number of conversions for M=6 and q=0.8.
53
1.E+00
0 1 2 3 4 5 6 7 8 9 10
1.E-01
survivor function of the number of conversions
1.E-02
1.E-03
1.E-04
q=0.8 M=4 N=16
1.E-05
q=0.8 M=4 N=24
1.E-06 q=0.8 M=6 N=16
1.E-07
q=0.8 M=6 N=24
q=0.6 M=4 N=16
1.E-08 q=0.6 M=4 N=24
1.E-09 q=0.6 M=6 N=16
q=0.6 M=6 N=24
1.E-10
1.E-11
Figure 4.5. Survivor function versus the number of conversions for B=1.
1.E+00
1.E-01 0 1 2 3 4 5 6 7 8 9 10
1.E-02
survivor function of the number of conversion
1.E-03
1.E-04
1.E-05
1.E-06
1.E-07
1.E-08
q=0.8 M=4 N=16
1.E-09
1.E-10 q=0.8 M=6 N=16
1.E-11 q=0.8 M=4 N=24
1.E-12 q=0.8 M=6 N=24
1.E-13
q=0.6 M=4 N=16
1.E-14
q=0.6 M=6 N=16
1.E-15
1.E-16 q=0.6 M=4 N=24
1.E-17 q=0.6 M=6 N=24
1.E-18
1.E-19
1.E-20
Figure 4.6. Survivor function versus the number of conversions for B=2.
54
Chapter 5. Concluding Remarks
This thesis evaluates buffering optical WDM packet switch equipped with shared
tunable wavelength converters (TWCs) for contention resolution. An analytical model,
allowing dimensioning of the number of wavelength converters, has been presented to
achieve a required performance. Combined with the wavelength dimension for contention
resolution, this reduces the size of each output buffer and employs a lesser number of
wavelength converters. The analytical model has been carried out on the required number
of converters versus the number of input/output buffers N, the number of wavelengths M,
the buffer size B, and the traffic loads. We observe that buffering optical switch allows
reductions on the employed number of converters compared with the bufferless
architecture. In addition, a lower packet loss probability can be achieved. As a matter of
fact, buffering optical switch using 2 converters for B = 2, M = 6, and N = 16 can provide a
lower packet loss probability than bufferless optical switch using 7 converters for B = 1, M
= 6, and N = 16.
Also, we analyze mean contention time in output contention process and mean
packet delay in output queue. We use system parameters, like the number of input/output
buffers, the number of wavelengths, the buffer size, and the traffic loads, for evaluating
mean total delay. Our studies indicate that the output contention process can be modeled by
a BI/D/B/N queueing process and the output buffer can be modeled by a G[x]/D/1/B
queueing process. Performance of survivor function in buffering packet switch equipped
with shared TWCs is extensively studied by means of numerical simulations. It is seen that
lower traffic load will have lower survivor function. The survivor function versus the
number of wavelength conversions for using buffer is seen to be lower than that not using
buffer. Converters are used frequently for using buffer because utility conversion rate is
larger than that not using buffer. Therefore, survivor function is lower for using buffer than
that not using buffer.
Furthermore, performance of mean total delay in buffering packet switch equipped

with shared TWCs is evaluated by means of numerical simulations. From the functional
diagram between mean total delay and traffic load, we see that lower traffic load will result
55
in a lower mean total delay. When the number of input/output fibers increases under the
same traffic load, similar outcome results are obtained; therefore, the number of
inputs/outputs is not the major cause for the mean total delay. When the number of
wavelengths reduces under the same traffic load, the mean total delay will increase; the
lower buffer size will have the longer mean total delay for the buffering optical switches.
56
References
[1]. Gerd Keiser, Optical Fiber Communications, New York: McGraw-Hill, 2000.
[2]. Shun Yao, B. Mukherjee, and S. Dixit, Advances in photonic packet switching:
an overview, IEEE Communications Magazine, vol. 38, no. 2, pp. 8494, Feb.
2000.
[3]. Eramo, V. and M. Listanti, Packet loss in a bufferless optical WDM switch
employing shared tunable wavelength converters, Journal of Lightwave
Technology, vol. 18, no. 12, pp. 18181833, Dec. 2000.
[4]. Juan Diao and P.L. Chu, Analysis of partially shared buffering for WDM
optical packet switching, Journal of Lightwave Technology, vol. 17, no. 12,
pp. 24612469, Dec. 1999.
[5]. Ribeiro, M.R.N. and M.J. OMahony, Improvements on performance of

photonic packet switching nodes by priority assignment and buffer sharing,
2000 IEEE International Conference on Communications, vol. 3, pp.
17381742, 2000.
[6]. Hunter, D.K., M.H.M. Nizam, M.C. Chia, I. Andonovic, K.M. Guild, A.
Tzanakaki, M.J. O'Mahony, L.D. Bainbridge, M.F.C. Stephens, R.V. Penty,
and I.H. White, WASPNET: a wavelength switched packet network, IEEE
Communications Magazine, vol. 37, no. 3, pp. 120129, March 1999.
[7]. Kuo-Chun Lee and V.O.K. Li, Optimization of a WDM optical packet switch
with wavelength converters, INFOCOM '95, the 14-th Annual Joint
Conference of the IEEE Computer and Communications Societies,
Proceedings IEEE, vol. 2, pp. 423430, 1995.
57
[8]. Danielsen, S.L., C. Joergensen, B. Mikkelsen, and K.E. Stubkjaer, Optical
packet switched network layer without optical buffers, IEEE Photonics
Technology Letters, vol. 10, no. 6, pp. 896898, June 1998.
[9]. Danielsen, S.L., B. Mikkelsen, C. Joergensen, T. Durhuus, and K.E. Stubkjaer,

WDM packet switch architectures and analysis of the influence of tunable
wavelength converters on the performance, Journal of Lightwave Technology,
vol. 15, no. 2, pp. 219227, Feb. 1997.
[10]. Rajiv Ramaswami and Kumar N. Sivarajan, Optical Networks, San

Francisco: Morgan Kaufmann, 1998.
[11]. Li, S.-Q., Nonuniform traffic analysis on a nonblocking space-division packet

switch, IEEE Transactions on Communications, vol. 38, no. 7, pp. 1085
-1096, July 1990.
[12]. Lee, M.J. and D.S. Ahn, Packet loss analysis of nonblocking ATM switches
with nonuniform traffic and performance improvement by output buffer
sharing, Technical Program Conference Record, GLOBECOM '93, IEEE in
Houston, vol. 2, pp. 1079 -1084, 1993.
[13]. Guogen Zhang, W.G. Bulgren, and V.L. Wallace, A performance model of
space- division ATM switches with input and output queueing, Proceedings of
the 29-th Hawaii International Conference on System Sciences, vol. 1, pp. 72
-79, 1996.
[14]. R.A. Howard, Dynamic Probabilistic Systems, Volume I: Markov Models,

New York: Wiley, 1971.
58
(About the Author)
English Name : Chia-Hsing Chou

67 2 19
Birthday19th, February, 1978.
Education
(1984~1990)
(1990~1993)
(1993~1996)
(1996~2000)
BS, Department of Electrical Engineering,

(2000~2002)
Graduate Student, Department of Electrical
Engineering, National Cheng Kung University
View publication stats

Performance Evaluations On A Buffering WDM Optical

Diunggah oleh

Informasi Dokumen

Judul Asli

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

Performance Evaluations On A Buffering WDM Optical

Diunggah oleh

Hak Cipta:

Format Tersedia

See

Performance Evaluations on a Buffering WDM

Department of Electrical Engineering,

(Tunable Wavelength Converters, TWCs)

(Packet loss probability)

(Mean contention time)

(Mean total delay)

a Buffering WDM Optical Packet Switch

Equipped with Shared Wavelength Converters

Chia-Hsing Chou* Jen-Fa Huang**

Furthermore, we analyze mean contention time in output contention process and

Chapter 1. Introduction ... .1

1.1. SCM Header and Packet Format.........2

1.2. Contention Resolution.......5

1.2.1. Wavelength Dimension Technique.5

1.2.2. Buffering Technique.7

1.2.3. Deflection Routing Technique....9

1.3. The Motivation of our Research.....12

Chapter 2. Previous Architectures of WDM Packet Switch...14

2.1. Wavelength Conversion Technique........14

2.2. Bufferless WDM Switch with Shared TWCs.......17

2.3. Partially Shared Buffering for WDM Packet Switch..19

2.4. Comparisons of Bufferless and Shared Buffer Schemes22

Chapter 3. Shared TWCs for Buffering WDM Packet Switch..25

3.1. WDM Optical Packet Switch......25

3.2. Dimensioning the Number of TWCs..28

3.3. Evaluation of packet loss probability36

3.3.1. Evaluation of Ploss in Full-Conversion Configuration (r = NM)..36

3.4. Numerical Results of Packet Loss Probability41

Chapter 4. Evaluation of mean total delay and survivor function ..45

4.1. Numerical Results of Mean Total Delay........45

4.1.1. Output Contention Process (BI/D/B/N queue).....46

4.1.2. Output Buffer Process (G[x]/D/1/B queue).47

4.2. Numerical Results of Survivor Function.52

Chapter 5. Concluding Remarks......55

Figure 1.1. Qualitative power spectrum of the laser modulation current...3

Figure 1.2. Header retrieval in SCM..4

Figure 1.4. A two-input, two-output routing node in the network.8

Figure 1.5. The shared memory optical packet switch..9

Figure 1.6. The Manhattan Street Network.10

Figure 1.7. The ShuffleNet..11

Figure 4.2. Mean total delay versus traffic load q51

Table I. M = 4, k = 2, B = 2, total number of packets BM.33

Table II. M = 4, k = 2, B = 2, total number of packets < BM...35

Wavelength-division multiplexing (WDM) techniques have rapidly gaining

Wavelength routing: In addition to using multiple wavelengths to increase link

Wavelength switching: Whereas wavelength-routed networks are based on a rigid

1.1 SCM Header and Packet Format

Among several different proposed solutions, packet switching with subcarrier

Figure 1.1. Qualitative power spectrum of the laser modulation current.

Figure 1.2. Header retrieval in SCM.

Packet length is another problem of concern for network designers. A greater

1.2 Contention Resolution

1.2.1 Wavelength Dimension Technique

The wavelength dimension technique [3] uses wavelength dimensions as a logical

Input Fiber DMUX MUX Output Fiber

Tunable wavelength converter

As aforementioned, in Figure 1.3, the number of TWCs is assumed to be

1.2.2 Buffering Technique

Figure 1.4. A two-input, two-output routing node in the network.

Figure 1.5. The shared memory optical packet switch.

1.2.3 Deflection Routing Technique

3 3 Manhattan Street Network