Sie sind auf Seite 1von 42

20

Photonic Switching in Space

‘‘And God said, Let there be light: and there was light. And God saw
the light, that it was good.’’ — Genesis 1:4-5a, from the King James
translation of the Bible
This chapter and the next one describe how data in optical form can be switched without
ever converting the data out of optical form. This chapter describes photonic switching in
the space division and the next chapter extends the discussion to include the time and
wavelength divisions. Unlike optical computing, photonic switching (particularly in the
space division) is not some technology that we can’t see until the distant future. To em-
phasize the near-term possibility of photonic switching, this chapter deliberately de-
scribes some relatively old and proven devices and some sensible architectures that use
these devices. Newer research devices are mentioned and one is discussed in a little de-
tail in Section 21.5.1.
We’ll see in these two chapters that photonic technology lends itself to call switch-
ing, capacity switching, protection switching, and provisioning — the right half of Figure
17.14. However, while photonic protection switches are commercially available, photon-
ic circuit switches have not been commercially successful. Because photonic technology
does not lend itself well to packet switching, almost all advances in circuit switching have
been put on hold until we see how ubiquitous packet switching is going to become. This
delay to reach commercialization should not be interpreted to mean that the technology
and the architectures are not ready to be commercialized. Section 22.7 proposes using
photonic technology to build a national capacity-switched all-optical network that would
support circuits and packets.
This chapter describes circuit switching, in the space division, of optical signals us-
ing optical devices. The reader is assumed to be slightly familiar with the fundamental
physics of optical fiber and of elementary devices like LEDs, lasers, and photodetectors.
Section 20.1 is a long introduction — discussing background, motivation, synergy (Sec-
tion 13.1), and definitions. A natural dichotomy of photonic switching devices is dis-
cussed in detail in Section 20.2. It develops some reasons for concentrating on certain
kinds of devices and excluding others, and raises the importance of a specific data format.
Section 20.3 describes several photonic switching devices — but only of one type. Sec-
tion 20.4 is a lengthy description of many different architectures for fabrics that perform

683
684 Telephone Switching Systems

photonic switching in the space division only. Several photonic switching architectures
are described, optimized for different photonic switching devices. Section 20.5 discusses
photonic packet switching. Photonic switching in time and in wavelength and switching
in multiple divisions are described in the next chapter.

20.1 INTRODUCTION
This lengthy introduction discusses some relevant background of optical communications
and photonic switching — the characteristics and limitations, architectural impact,
contemporary utilization, and the almost unbounded future. Very carefully, this section
makes a strong case for circuit switching of optical signals in the space division, using
optical devices. Photonics and optoelectronics are carefully distinguished.
20.1.1 Background
Corning Glass patented optical fiber in the 1970s. Its potential for telecommunications
was immediately recognized and a worldwide effort followed that developed improved
fiber, manufacturing techniques, necessary related electronics, connectors and splicing
technology, practicality, applications, and education. During the 1980s, research
laboratories (most notably NTT Labs in Japan and Bell Labs in the United States)
competed vigorously to break one another’s world record for the optical fiber
transmission rate × distance product. Demand was so great and competition was so
intense that the research state of the art was typically getting to the marketplace in only
about 3 years.
It is remarkable that within 10 years of its invention, optical fiber was practical
enough that it was being deployed by many telephone and CATV companies and many
users. Within 20 years of its invention, optical fiber became the transmission medium of
choice, worldwide. The real potential of optical fiber, and the justification for investing in
its deployment, is its enormous and, even now, relatively unused bandwidth. While we’re
pushing the limits of the bandwidth of moving electrons along copper, it seems hard to
imagine running out of information-carrying capacity over optical fiber.
Compare information moving along several different transmission media to water
flowing through a pipe. Since a pipe with larger diameter allows more water to flow, pipe
diameter is analogous to bandwidth in the transmission channel. If a conventional analog
telephone channel is analogous to conventional household 3/8-inch copper tubing (0.25i
id), then coaxial cable carrying 70 television signals is analogous to a 6-foot diameter
sewer pipe. (This is not a comment about the content.) Admittedly, there’s a bit of an
apple/orange comparison, but the ratio of areas is equal to the ratio of the capacities of a
4-kHz telephone channel to 70 television channels at 4 MHz each. On this same scale,
optical fiber’s ultimate capacity is bigger than the Grand Canyon — not just the Colorado
River at the bottom, but the entire canyon, filled to the brim with moving water. While
we’re not even close to being able to use all this capacity, the capacity we can use is
higher than the competing media, we’re progressing toward using more and more of it
every year, and the end is not in sight (like it is with copper). Some claim that the optical
fiber currently in the ground is future-proof.
Photonic Switching in Space 685

20.1.2 Motivation
While optical fiber has several significant advantages, the most significant is its ability to
carry data at very high rates. Fiber provides a relatively inexpensive medium for
broadband channels. Fiber is also used for conventional digital voice channels, an
application in which the channels are highly multiplexed. We see that transmission is
migrating to optical fiber, that data rates are increasing and will continue to increase, and
that the architecture of the PSTN is migrating toward having a reduced number of nodes
that are interconnected by comparatively few physical channels, which are highly
multiplexed. While this trend has an adverse effect on network reliability, as the United
States witnessed in the early 1990s, this chapter describes architectures that support the
trend.
Recall from Section 13.1 that the migration to digital switching was largely
motivated by synergy, the conservation of A-to-D converters on the boundary between
switching and transmission. Thus, one is tempted to justify the migration to photonic
switching by a similar synergy, the conservation of E-to-O converters on this same
boundary. Photonics technology, however, may not be advanced enough that we can
realistically expect long-haul fiber to terminate directly on photonic equipment, without
intermediate electronic equipment.
• Repeaters. While optical amplifiers, especially the Erbium-doped fiber amplifiers
(EDFAs), are in practical usage, they are analog amplifiers: they amplify a digital
signal; they can not regenerate it. While nonlinear photonic devices exist that
might perform digital regeneration of an optical signal, they are far from practical
in this application. So, at least some of an optical line’s amplifiers might have to be
electronic because they may have to do more than just amplify. Current thinking is
that every third or fourth amplifier in a transmission line will have to be an
electronic digital repeater. It seems logical to place such an electronic digital
repeater at the ends of the optical links inside the network nodes.
• Polarization alignment. Many of the photonic technologies are sensitive to the
polarization of the received optical signal. However, the polarization of a signal
coming off an optical fiber is not necessarily 100% linear and its alignment is
certainly not predictable. While polarization can be linearized and aligned
mechanically, without electronics, it may be cheaper and more reliable to do the
job with an optical repeater.
• Wavelength demultiplexing. We should expect, more and more, that an optical fiber
will carry channels on more than one wavelength, multiplexed together. While
wavelengths are demultiplexed optically, typical demultiplexors have so much
optical loss that we need repeaters, or at least amplifiers, after demultiplexing.
• Format conversion. If line coding formats have to be converted (for example, from
Manchester to RZ), this will require electronics. Converting frame formats (for
example, North American to European DS1) also requires electronics. Facility
switching and add-drop multiplexing with conventional ADH or SDH frame
formats also require electronics.
686 Telephone Switching Systems

• Packet processing. Computing cyclic redundancy code (CRC) checks, reading


headers for routing, and other packet header computations all require electronics.
• Synchronization. We don’t have an optical equivalent to the phase-locked loop and
we certainly can’t do frame synchronization and all the other types of
synchronization that are required with complicated digital frame formats. If we
need to synchronize a signal, we need to do it in electronics.
If any of these functions listed above, or any other functions that may have been omitted
from this list, requires an electronic implementation, then they would probably all be
implemented together in electronics. They would probably all be colocated in the node,
on the transmission-switching boundary, and the synergy argument would fail.
So, if synergy is not the principal motivation for photonic switching, then,
Why would anyone consider switching photonically, using devices that
still require more research, when one could switch electronically, with
devices that are already proven and practical?
Does this question sound familiar? Reread the second sentence of the third paragraph of
the Chauncey Depew letter in Figure 2.1. Here are a few of the good reasons for
switching optical signals directly.
• Protection switching. Sections 12.5.7 and 17.5 discussed the practice of installing
physical analog switching between transmission facilities and their corresponding
transmitters and receivers. This way if a facility is decommissioned or faulty, the
expensive transmitters and receivers can be moved to another facility. This is
called protection switching and it is even more common in optical facilities. So, if
we’re going to have some kind of photonic switching network anyway, as a
protection switch, maybe its role should be expanded to include facility switching,
capacity switching (Section 17.5), and even circuit switching.
• Noise reduction. Photons on fiber are inherently less prone to electromagnetic
interference than electrons on copper. While most noise accumulates in the
transmission facility, switching systems are also noisy environments. If noise and
crosstalk become very serious problems, particularly as the bit error rate (BER)
requirement drops to 10−12 , photonic switching may provide the answer.
• Modularity. Some firsthand personal experience has taught me an interesting and
surprising lesson. The construction of the DiSCO prototype, described in the next
chapter, entailed a great deal of complicated electronics and photonics equipment.
In DiSCO’s assembly and debugging, we spent far more time dealing with the
electronics than with the photonics. The electronics needed to be wire-wrapped
and we never did completely debug some problems in the electronics that I’m
convinced were caused either by some bizarre heat effect or by some strange
coupling through a power supply that might not have been perfectly regulated. The
photonics components were all pigtailed and every time we connected a photonic
component, it just worked the first time. As a newcomer to photonics and a veteran
of electronics, I was expecting to have more trouble with the photonics than with
the electronics. But, that is not what happened. While I’m not going so far as to
Photonic Switching in Space 687

suggest the use of photonics over electronics because the implementation and
system integration will be easier in photonics, I have no firsthand experience to
refute this controversial point.
• Immense connectivity. The inherent three-dimensional nature of light suggests that
a broad beam of light, or many parallel information-bearing rays, could be
projected onto a two-dimensional plane of interconnected nano-processors, each
with its own optoelectonics. One application is highly parallel picture processing
for video generation, pattern recognition, and robotic vision. But, three-
dimensional ‘‘free-space’’ systems have also been proposed to interconnect
thousands of parallel optical signals. This potential benefit of optics is so relevant
that we might even consider a photonic switching node, even if the transmission
plant is all digital electronics and copper.
• Immense data rate. Because optical transmission avoids the detrimental reactive
effects caused by parasitic inductors and capacitors in electronic transmission,
optical channels have extremely high theoretical bandwidth. This very high
channel capacity gives not only optical fiber, but also other photonic components,
the potential to carry vast amounts of information at extremely high data rates.
There is a related characteristic that may be even more important. Not only do
optical fiber and other photonic components have extremely high channel capacity,
but their construction is independent of the rate of the data that will go through
them. Data rate is not part of the specification of such components, as it is with
almost all electronic and copper components. Networks built with this technology
have rate-scalable infrastructure.
These last two potential advantages that optics might have over electronics have the
potential to be architectural breakthroughs. The ultimate goal, of course, is a system with
both these characteristics — a physically small system that can provide rapid
interconnection among thousands of optical fibers, each carrying 20 Gbps this year (and
50 Gbps 5 years later). Without a device breakthrough, this ultimate system will elude us
because, so far, photonic devices come in a natural dichotomy that forces us to select one
characteristic or the other. It appears that we can have immense connectivity or immense
data rate, but we can’t have both. The devices that have the potential for high
connectivity can’t support extremely high data rate, and the devices that have the
potential for high data rate can’t provide high connectivity.
20.1.3 Photonics and Optoelectronics
One of the drawbacks of electronic devices is that, by their nature, electrons move around
inside them. Moving electrons are subject to resistance and reactance. We’ll use the term
‘‘optoelectronics’’ to describe devices that operate externally on an optical signal but,
when they do, a corresponding electronic signal moves around internally.
The term ‘‘photonics’’ suggests a family of technology based on the photon,
analogous to how electronics technology is based on the electron. We’ll use the term
‘‘photonics’’ to describe devices, which may be electronically controlled, but through
which externally applied optical signals remain as optical signals internal to the device.
The goal of research in photonic switching is to use these new technologies to
construct complete networks that would switch data in optical form, without conversion
688 Telephone Switching Systems

to electronics. While these same technologies might be used for optical computing,
which is discussed briefly in this chapter, this chapter focuses on photonic switching.
Because optical communications channels are defined in space, time, and wavelength,
photonic switching in each of these three divisions is important. Photonic switching in
the space division is described in this chapter. Photonic switching in the time division,
the wavelength division, and in combinations of the three divisions, is described in the
next chapter.
This chapter describes several exotic photonic switching devices that could
significantly impact the future of telecommunications and computing. While interesting
devices are described, the coverage is superficial and the chapter, like the book, is
presented primarily at the architectural level. Because many of these photonic devices
are immature, a mediocre device characteristic is seen, not as a ‘‘device problem,’’ but as
an ‘‘architectural challenge.’’ Simply stated, if photonic devices don’t have satisfactory
characteristics, we will architect around their problems.

20.2 DEVICE DICHOTOMY


The devices that might be useful for photonic switching have a natural dichotomy. The
optoelectronic and photonic devices that may be relevant to computing and commu-
nications come in two general categories. This section presents an interesting argument,
based on the thermodynamics of data rate, that increases the relevance of one of these two
types of devices, analog switching, and the block-multiplexed data format.
20.2.1 Two Categories of Devices
Two categories of devices are ‘‘free-space’’ and ‘‘guided-wave.’’
Free-Space Devices. Because these devices operate by some form of optoelectronic
physical phenomenon, they regenerate any digital signals that pass through them but also
impose internal reactance effects on them. Because, in this class of devices, each
individual component is physically oriented normal to the surface of its substrate, they
have a very high potential for integrability, but it is difficult to interconnect many
neighboring devices into an integrated network. Because their input/output access is via
the substrate’s surface, many parallel devices can operate on many parallel signals, either
as pixels in some common picture or as independent channels. Their two-dimensional
parallel nature, surface access, and high potential integrability make them ideal for a free-
space architecture that could provide immense connectivity. Regeneration makes them
ideal for switching digital signals.
Guided-Wave Devices. Because these devices operate by some form of photonic
physical phenomenon, they impose no reactance effects on signals that pass through
them, but they cannot regenerate those signals. Because, in this class of devices, each
individual component is physically oriented tangential to the surface of its substrate, they
have a lower potential for integrability, but they are easily interconnected into complex
integrated networks. Because their input/output access is via the substrate’s edge, fewer
parallel devices can operate on fewer parallel signals. The optical devices that will be
described in the next section, optical fiber itself and both kinds of photonic amplifiers,
can all support extremely high data rates. But, they are all analog devices and none of
them integrates very well. Their sequential nature and reactance-free operation make
Photonic Switching in Space 689

them ideal for a network architecture that could switch a few signals each carrying an
immense data rate and for networks where rate-independence is important. But, they
perform analog switching, even on digital signals.
The most fundamental distinction between the optoelectronic devices and the guided-
wave devices is illustrated next.
20.2.2 Thermodynamics of Data Rate
We present an interesting way to evaluate the long-term potential of a switching
technology by looking at thermodynamic limitations imposed by data rate and by
integration [1,2]. First, a chart of power versus rate is introduced and explained, and
then, three limiting cases are explained. Then, integration is shown to exacerbate the
thermodynamics (because many switching devices are all generating heat on the same
chip) and analog switching is shown to have a significant advantage — when data is
formatted using block multiplexing.

Power
1 kW -
Heat transfer
limit
1W-
1 uJ
1 mW - 1 nJ
1 pJ
1 uW - 1 fJ
Noise 1 aJ
1 nW - limit
Quantum
limit
1 pW -| | | | | |
1 Pbps 1 Tbps 1 Gbps 1 Mbps 1 Kbps 1 bps
Data rate
Figure 20.1. Power versus rate. © 1982 AT&T [1].
Figure 20.1 illustrates the thermodynamics of bit rate. The figure is also an exercise
in the obscure regions of the metric system. Data rate is shown on the horizontal axis and
total dissipated power is shown on the vertical axis. The dashed diagonal lines represent
technology families, by the energy dissipated whenever a switch operates. For example,
if data at 1 Gigabit per second is processed by a switch which dissipates 1 picojoule for
each bit, then the total dissipated power is:
10−12 joules/bit × 10+9 bits/second = 10−3 joules/second = 1 milliwatt
This algebraic result is obtained from Figure 20.1 by reading up from ‘‘1 Gbps’’ to the
dashed line for ‘‘1 pJ’’ and then across to ‘‘1 mW.’’ One milliwatt of dissipation is also
obtained when a 1-nJ device operates at 1 Mbps, a 1-fJ device operates at 1 Tbps, or a
1-aJ device operates at 1 Pbps.
While we want to be able to process bits as fast as technology allows, a higher data
690 Telephone Switching Systems

rate implies more switch operations per second and, hence, more heat dissipated in the
technology. While physical data transmission doesn’t require switch operations, many
switch operations are required in the transmitters, repeaters, and receivers along a
transmission line and, especially, in the system equipment on the ends of the lines. Many
switch operations are required whenever data is modulated and demodulated, encoded
and decoded, compressed and decompressed, multiplexed and demultiplexed. These
switch operations are unavoidable. Germane to this discussion, however, switch
operations are also required anywhere data is switched. The number of switch operations
per bit of data and hence the thermodynamics of the switching fabric, depends on the
mode of switching, the data format, and the architecture of the fabric.
Figure 20.1 indicates that increasing the number of switch operations per second is
accompanied by greater heat dissipation. Three solid lines on Figure 20.1 represent three
different physical limits on bit processing.
• One of the lines that proceeds diagonally downward is labeled ‘‘Quantum Limit’’.
This line represents the optical power that is carried by an optical signal at f bits per
second that has only 10 photons in each bit and only one cycle of carrier in each
bit.
P q = 10 e p per second = 10 hf × f = 10 × 6.626 ×10−34 × f 2
At 1 Pbps, the optical power in such a signal is 6.6 mW. Because Figure 20.1 is a
log-log scale, a quadratic in f proceeds linearly to the right with slope = -2.
Because we’ve assumed that the data rate equals the carrier frequency, signals in
the high tera bits per second must be optical. While we can exceed one bit per
cycle in today’s electronic modems, it’s hard to imagine doing it at high giga bits
per second data rates. And, because no amount of energy can be lower than the
energy of one photon, it’s unlikely that any system can operate in the region to the
lower left of this line.
• The other line that proceeds diagonally downward is labeled ‘‘Noise Limit’’. The
power produced by random thermal noise is given by:
eTN
P TN = = 4kT × f
time
where Boltzmann’s Constant, k = 1.38 × 10−23 J/°K. Because the components’
ambient temperature is at least T = 300°K and since it would be difficult to detect a
signal if its SNR is lower than 3 dB, this line represents 2 P TN = 32 × 10−21 × f .
No system can operate to the lower left of this line either.
• The line that proceeds diagonally upward is labeled ‘‘Heat Transfer Limit’’. This
line represents the approximate limit for conducting heat away from existing
semiconductor technology, about 1 watt per nanosecond. Any switching device
that operates in the region to the upper left of this line, such as a 1-pJ device at 100
Gbps, accumulates heat faster than it can be conducted away — and overheats.
Given these three constraints, it appears from Figure 20.1 that the maximum possible data
rate is about 30 Tbps and that this rate is achievable only if devices are developed that can
switch at less than 1 aJ/bit. Existing devices are approaching 10 fJ/bit, which makes even
100 Gbps a problematic data rate.
Photonic Switching in Space 691

In this context, consider two additional factors: integrated circuits and analog
switching. The thermodynamics of these two factors significantly impacts the two
principal characteristics that motivate photonic switching: connectivity and throughput.
20.2.3 Level of Integration
While the ability to integrate many devices onto a single chip is important for commercial
practicality of any device, it is essential if we are to construct that projection plane
required for a three-dimensional high-connectivity system. Many of the optical devices
that show promise for a high level of integration are optoelectronic (not photonic) and
digital.
The Self Electro-optic Effect Device (SEED) is such a device [3,4]. The SEED is an
optical threshold device, typically implemented as an optical two-input NOR gate. As a
true digital device, the SEED has most of the same advantages that electronic digital
devices have, including that binary pulses are regenerated as they pass through. A SEED
can perform a computer logic operation and it can serve as a simple switch in
communications applications. Furthermore, integrated circuits have been built that have
thousands of SEEDs on a chip.
Researchers have proposed (on paper) many optical computer architectures, and
actual working prototypes have been built: one using SEEDs and another using switched
directional couplers. While both prototypes are bit-serial computers, built with discrete
components in a two-dimensional architecture, they are still important demonstrations of
proof of concept. While optical computing is important research, there is still a lot to be
done. It’s clear that we will be executing programs on electronic computers for many
years to come.
Three-dimensional free-space architectures for switching networks have also been
prototyped, with their projection planes made from a variety of 1 × 1 optoelectronic
devices: mechanical and liquid-crystal shutters, devices that operate like optical silicon
controlled rectifiers (SCR), individual SEEDs, and integrated circuits of SEEDs. All
these prototypes of switching fabrics are also important proofs of concept. They have
illustrated the need for economical imaging optics if the three-dimensional architecture is
ever to become practical, especially in multistage fabrics. But, they have illustrated
another, more fundamental, problem.
The subtle disadvantage to being digital is that the switching device operates on
every bit, and because it’s also optoelectronic, electrons move around deep inside the
device with every logic operation. So, each bit that passes through produces heat. If a
digital optoelectronic device expends 10 pJ with each bit of throughput, then a discrete
device operating at 10 Gbps dissipates 0.1 W of thermal energy. If 10,000 of them on a
single chip could all operate at this data rate, then the chip would need to dissipate 1 kW.
Because it can’t, it would overheat. The reader is cautioned to be wary of the research
reports on some of the switching technologies, including photonic technologies. Devices
have been reported that can operate at 1 Gbps. It is also reported that these same devices
have been integrated to more than 10,000 devices per chip. However, these two reports
cannot be taken together to imply that 10,000 devices on one chip can all operate
simultaneously at 1 Gbps. The digital optoelectronic devices that can be highly
integrated, while they might be operated quickly as discrete components, have
thermodynamic limitations when integrated together because, in the close proximity of an
integrated circuit, all the devices on a single chip contribute to the chip’s temperature.
692 Telephone Switching Systems

20.2.4 Analog Switching for Throughput


In an analog device, like an electromechanical relay (Section 2.3.1), the signal that
operates the device is different from the data that passes through the device.
If data is organized and formatted so that a ‘‘block’’ of consecutive bits
is switched to the same switch port, then the device operates once per
block, just before the beginning of each block, instead of having to
operate on each individual bit.
This block multiplexing format is discussed in more detail in Section 21.1.1. If such a
device expends 10 pJ with each switching operation, then a discrete device operating on
one timeslot every 100 ns, dissipates 0.1 mW of thermal energy, for any data rate. Even
if 100 such devices on a single chip could all operate at this timeslot rate, then the entire
chip would dissipate only 10 mW. We see that analog devices will excel at throughput
and are consistent with serial time-multiplexed optical data streams that are confined to
separate channels; that is, the existing real world.
We saw that, in the very act of regenerating, a digital device operates with each bit,
and because it is also optoelectronic, electrons move around inside the device each time it
operates. So, an analog device that switches a block with 1000 bits dissipates the same
amount of heat as a digital device that switches 1 bit. A 1-pJ device will overheat if it
operates at 1 Tbps but if it’s analog and if it’s switching 1000-bit blocks of data, it only
has to operate for each block (not each bit). Because the device switches at a rate of 1
Gigablock per second, it operates quite comfortably. In fact, a chip with 10 or 100 such
devices integrated together would still operate within the heat-transfer limit.
In the block multiplexed data format, data is packed into blocks at the line rate, but a
guard-band interval is provided between consecutive blocks of data. No data is allowed
in the guard band between the blocks. Two issues contribute to the required duration of
the guard band: switching speed and variation in packet arrival time. We assume that data
is organized into a block that is bunched into the temporal center of a timeslot. A time-
multiplexed N × N switching network would have to be switched into its assigned
configuration for a given timeslot at the beginning of that timeslot, before any of the data
actually arrived at the switches. It would have to hold this configuration until all the data
in every parallel block was throughput.
• The guard band must be long enough to allow all the switches time to operate and
thus configure the state of the network.
• The data in each timeslot will not arrive at the switching network at exactly the
same time. The different transmitters will not be perfectly synchronized, the fiber
links will have different lengths, and the index of refraction in each link may be
different. So, the guard band must also be long enough to allow for any variability
in the arrival time of the data at the switching network.
20.2.5 Connectivity or Throughput?
We’ve seen that highly integrable digital devices, like SEEDs, are optimal for
connectivity and that the slightly integrable analog devices are optimal for throughput.
Figure 20.1 illustrates an interesting design space in which we can try to get a little of
both. The bottom line is that, with existing technology, these four things — (1) three-
dimensional connectivity, (2) digital regeneration, (3) integrated circuits, and (4) high
Photonic Switching in Space 693

data rate — do not scale together. Systems with only three of these characteristics, any
three of the four, are not nearly as interesting as systems with all four. In terms of
optimizing its application, we must choose between connectivity and throughput — we
can’t have both.
There just isn’t a perfect photonic switching device. If there were, we could think
about having high connectivity and high throughput in the same system. With today’s
devices, one must choose between connectivity and throughput. So, which is more like
the real world?
• A photonic switching node that is built with highly integrated digital photonic
devices can terminate thousands of fibers, but all the fibers must operate at very low
data rates.
• A photonic switching node that is built with slightly integrated analog photonic
devices can terminate fibers that operate at very high data rates, but it can only
terminate tens of fibers.
Because the latter is more consistent with the evolving national network, the remainder of
this chapter deals only with these systems. Despite how good it sounds to be digital and
highly integrable, when it comes to photonic switching, scalability seems to favor the
devices that allow high throughput, even if this means that they are analog devices and
are only slightly integrable. This chapter and the next one emphasize this analog
photonic switching.

20.3 GUIDED-WAVE SWITCHING DEVICES


Three guided-wave devices capable of photonic switching are described here: the
semiconductor laser amplifier, the Mach-Zender interferometer, and the Lithium Niobate
switched directional coupler.
20.3.1 Semiconductor Laser Amplifiers
A semiconductor laser amplifier (SLA) is just a semiconductor diode laser (SDL) with
the following differences [5 – 7].
• The SDL is forward biased to produce a 1 and reverse biased to produce a 0. The
SLA is forward biased to turn the amplifier on and reverse biased to turn it off.
• In a forward-biased SDL, photons are spawned by spontaneous emission in a
confined physical channel that spans the junction, and photons (within the SDL’s
narrow gain bandwidth) are multiplied by stimulated emission each time they
traverse the channel. While an SLA’s forward-biased current is large enough to
sustain multiplication by stimulated emission of those photons that enter the
channel from the input fiber, this current is kept small enough to keep the diode’s
spontaneous emission (seen here as noise) at a reduced level. While an SLA is
constructed like an SDL, it acts more like an LED than an SDL when there is no
input signal.
• The SDL’s channel is partially reflective (optically resonant) on both ends, so many
of the photons in a signal bounce back and forth, being amplified for each pass
down the channel and establishing the standing wave that gives the SDL its modes.
694 Telephone Switching Systems

In the SLA, the chip’s end-facets are given an antireflective coating so most
photons make only one pass down the channel, from input fiber to output fiber, and
the physical resonance that causes longitudinal modes is reduced.
• In an SDL, the photons that don’t reflect are captured in the output fiber as an
optical 1 pulse. But, SLAs are manufactured with two fiber ‘‘pigtails,’’ one for the
attenuated input signal and one for the amplified output signal. The input fiber
couples an attenuated photonic signal into one end of the channel, and the
amplified photonic signal couples from the other end of the channel into the output
fiber. Some SDLs are also pigtailed with two fibers: one for the output signal and
one that feeds the output signal back to the electronics to provide for signal
stability and gain control.
• An SDL’s output signal has longitudinal modes corresponding to the standing
wavelengths that lie within the gain spectrum. An SLA’s output signal is an
amplified version of the input signal, corrupted by the SLA’s spontaneous emission.

Optical
power
Wavelength
w0
Figure 20.2. Longitudinal modes in a laser’s spectrum.
The antireflective (AR) coating is similar to the radar-absorbing coating applied to
spy planes. The boundary between two media is painted with a coating that is a quarter-
wavelength thick and has a reflectivity such that the ideal net reflected signal is the sum
of two signals of equal magnitude, 180° out of phase. An SLA with a poor AR coating,
with high reflection at its edges, has a gain spectrum that resembles an SDL’s power
spectrum (Figure 20.2). As the AR coating improves, providing less reflection, the gain
spectrum changes: (1) the center wavelength shifts, (2) the optical bandwidth increases,
and (3) the shape of Figure 20.2 smooths such that the peaks decrease and the valleys
increase. An SLA with a perfect AR coating, with no reflection at its edges, has a gain
spectrum that resembles an LED’s power spectrum.
While SLAs have been reported with gains around 20 dB, these are typically highly
modal SLAs where the wavelength of the input signal is tuned to a high-gain mode —
probably not a very practical arrangement because of the tuning. High-AR SLAs with
smooth gain spectra have lower gain but have lower variation in gain over the range of
wavelength. The optical bandwidth of an SLA’s gain is typically about 50 nm, limited by
the static nature of the antireflective coating.
Some of an SLA’s gain is lost because of the coupling between the device’s channel
and the input and output fibers. A net gain of 6 – 10 dB is reasonable. The useful
property of the SLAs is that the gain can be turned on and off rapidly by switching the
current signal to the device. If an SLA is a switched amplifier, it is also an amplifying
switch. Networks of SLA switches are discussed in the next section.
Photonic Switching in Space 695

20.3.2 Mach-Zender Interferometer


While the ultimate speed of light is a true constant, the speed of light in a material — and
many other optical properties of materials — are not true constants, but are complicated
functions of the material’s temperature, stress, optical intensity, and electric field. Any
material in which these optical properties are especially dependent on electric field is said
to be electro-optic. Crystalline Lithium Niobate, LiNbO3 , is a popular electro-optic
material.

+V x / 2

Fiber in Fiber out

Core Core
−V x / 2

Figure 20.3. Mach-Zender interferometer.


A classical Mach-Zender interferometer (MZI) is fabricated using discrete compo-
nents: two half-silvered mirrors for splitting and combining the optical signal and two
separately controlled electro-optic wave-guides. The more useful fabrication is integrated
on a chip of crystalline Lithium Niobate, as shown in Figure 20.3. By diffusing Titanium
into the material, a pattern of parallel optical wave-guides, with passive splitters and
combiners, is fabricated on the upper surface of a small crystalline chip of LiNbO3 . The
MZI’s input and output optical signals are carried over optical fibers whose ends are
physically butt-coupled against opposite edges of the chip so that the fiber cores align
with the ends of the diffused wave-guides.
The input signal splits so that equal halves of this signal traverse the two parallel
optical wave-guides. Then, these two components of the signal are combined again to
form the MZI’s output optical signal. Recall that the light in each of these parallel wave-
guides is a sinusoidal electromagnetic traveling wave. If the two wave-guides have equal
length and equal index of refraction, then the two signal components add, in phase, at the
output combiner, and the output signal exactly equals the input signal, except for any
internal attenuation.
In the MZI’s structure, however, each of these parallel optical wave-guides lies under
the plate of a different capacitor. So, if a voltage is applied to the structure, each wave-
guide resides in a different electric field. Because the material is electro-optic, this
external voltage gives each wave-guide a slightly different index of refraction. This
difference means that light travels a little faster in one wave-guide and a little slower in
the other. Some exact voltage, V x , changes the speed of light in the two wave-guides by
just the right amount so that the two optical signals arrive at the output combiner exactly
180° out of phase. Because the two signals would exactly cancel, or interfere with, one
another, the output signal would be completely extinguished. So, when the control
696 Telephone Switching Systems

voltage is 0, the input signal transmits through to the MZI’s output, and when the control
voltage = V x , the input signal is blocked.
Since the electro-optic effect occurs extremely rapidly, the MZI switches very
quickly, in even less than 1 ns, limited by the rate at which the capacitors can be charged.
Obviously, V x must have a different value for every different optical frequency
(wavelength). So, an MZI works only on coherent light, with all the optical energy at the
same frequency (wavelength). The MZI is a single-pole, single-throw switch in which an
electronic control voltage switches an optical data signal. MZIs are commonly used as
external modulators for lasers.
20.3.3 2 × 2 Lithium Niobate Switched Directional Coupler
This subsection discusses another analog photonic device, the switched directional
coupler (SDC) made from Lithium Niobate [8 – 11]. The SDC’s operation and
fabrication are presented here and its considerable shortcomings are explained in detail.
The SDC’s story continues in the next section, which describes some novel switching
networks that architect around the SDC’s device problems.

V X volts V B volts

Figure 20.4. LiNbO3 switched directional coupler.


The switched directional coupler, shown in Figure 20.4 is a little like the Mach-
Zender interferometer described above and a little like a passive optical coupler (POC).
While the MZI is a 1 × 1 device and the POC is a 1 × 2 device, the SDC is a 2 × 2 device,
with two input fibers and two output fibers. Instead of being passive like the POC, the
SDC is active like the MZI. Like the POC, any optical signal coming into the device on
either input fiber can transfer, partially or fully, to the other wave-guide because of
evanescent wave coupling.
SDCs are implemented in crystalline LiNbO3 because Lithium Niobate happens to
have a high electro-optic effect; its index of refraction is quite sensitive to any electric
field. The parallel wave-guides lie between the parallel plates of a capacitor. The
coupling of optical energy between wave-guides can be controlled electronically, as in an
MZI.
SDCs have been made that switch using a variety of different physical processes over
different electro-optic properties. While the more precise physics of SDCs is omitted
here, all SDCs have a common effective operation. The connectivity between an SDC’s
two input fibers and its two output fibers can be controlled by an external voltage, as
shown in Figure 20.4. With V B volts applied, signals emerge from output ports on the
same physical channel as their respective input ports. This is called the BAR state. With
V X volts applied, the input signals cross over to their opposite physical channels. This is
called the CROSS state. The SDC is seen to be a single device that is an optical
implementation of the beta element, discussed in Section 7.3.4.
Photonic Switching in Space 697

The SDC has two very important properties.


• Because electrons don’t move around whenever photons flow through an SDC, the
flow of optical data produces no heat. But, electrons do have to move around when
the device is switched between BAR and CROSS. So, if data is block multiplexed,
then the only device switching occurs on the timeslot boundaries. This makes
SDCs independent of data rate, so they don’t have to be replaced whenever the
network’s data rate is increased.
• Like the MZI, the SDC switches very quickly between its two states. In block
multiplexing, guard bands can be quite short. But, even if data is switched by the
bit, the SDC can keep up with relatively high data rates.
An SDC’s length, switching voltage, and switching speed are interrelated. Because SDCs
can be fabricated to optimize any one parameter, if you read about new world’s records in
these parameters, you must be careful not to assume that one device can be fabricated that
meets all these characteristics. In Lithium Niobate, a typical SDC is less than 1 cm long
and less than 100 microns wide. A typical switching voltage is in the range of 5 – 100
volts. Obviously, a lower voltage is preferable, but that would require that the device be
longer. While they can be operated in less than 100 ps, a more typical switching speed is
a couple of nanoseconds.
Because the device works by changing an electric field and the wave-guides are
embedded between the parallel plates of a capacitor, the applied switching voltage
actually does charge a capacitor. This magnitude of this capacitor can also slow down the
switching speed, but it is typically only a couple of picoFarads. SDCs work on single-
mode light only, and the switching voltage for light that is polarized normal to the plane
of the substrate is quite different from the switching voltage for light that is polarized
tangential to the plane of the substrate. So, incoming light must have its polarization
linearized and this polarization must be properly aligned before the light can enter the
SDC’s wave-guides.
As a true analog gate, an SDC not only doesn’t regenerate its binary pulses, it
introduces loss and crosstalk to them. Attenuation, crosstalk, and the sensitivity of the
SDC’s different parameters will be discussed later.
While several SDCs can be integrated onto one Lithium Niobate substrate, it would
be difficult to get a hundred of them on a chip. Since wafers are cut from crystal
cylinders, called boules, the chips are typically 3 – 4 inches long and about 1 cm wide.
While the SDCs are narrow enough that many could fit top-to-bottom, the wave-guides
must be separated at the chip’s edge to accommodate the thickness of adjoining parallel
optical fibers. SDCs cannot be densely packed along the length dimension because the
interstage wiring (should we call it channeling?) uses up a significant amount of this
dimension. Wave-guides diffused in the crystal that are allowed to bend at too great a
radius of curvature have too much attenuation. So, architectures with many SDCs in the
width dimension require more length to accommodate the increased curvature of the
interstage wiring.
698 Telephone Switching Systems

20.3.4 Lithium Niobate’s Problems


Besides their analog nature, photonic SDCs were criticized for other inherent problems
with Lithium Niobate technology, which made these devices appear to be problematic in
practical implementations of photonic switching fabrics. Lithium Niobate’s problems
include high attenuation, high crosstalk, narrow optical bandwidth, dependence on
polarization, low tolerance to small variations in the switching voltages, and variation of
all parameters to temperature and even age. While this long list of problems makes
Lithium Niobate seem impractical, it’s not nearly as bad as it appears because there are
really only two problems: attenuation and crosstalk. All the other problems are related to
crosstalk and the crosstalk problem can be architecturally finessed.
This subsection carefully explains Lithium Niobate’s problems. First, attenuation is
discussed, and second, crosstalk is discussed. Then, Lithium Niobate’s sensitivity
problems are carefully explained. This subsection concludes with a philosophical
discussion of the difference between a device solution and an architectural solution.
Section 20.4 discusses how these problems have been solved — not by refining the
devices, but by architecting around the problems — an architectural solution, called
dilation.
Attenuation. Various sources of attenuation are given, followed by fabrication
techniques and architectural techniques for dealing with the problem.
• Coupling loss. The largest source of attenuation in Lithium Niobate occurs at the
chip boundary, where the fiber butts against the chip and the optical signal couples
from the fiber’s core into the wave-guide that is Titanium diffused into the Lithium
Niobate. The pattern of the optical cross-section in a fiber’s core differs from the
corresponding pattern in one of Lithium Niobate’s Titanium-diffused wave-guides.
This modal mismatch causes a high attenuation where an optical signal couples
between the fiber’s core and the chip’s wave-guide at the boundary where a fiber
butts against the edge of a Lithium Niobate chip. When fabricating a chip, giving
the wave-guide a more complex structure at the chip boundary can reduce the
coupling loss. But, an architectural solution is even more effective. Coupling loss
is worse when all the SDCs are discrete, each residing on its own chip of Lithium
Niobate. Coupling loss can be greatly reduced from the discrete case if at least
several adjacent stages can be integrated onto a single chip. Coupling loss is
minimized if the entire fabric can be integrated onto a single chip. Even if it can’t
fit on one chip, coupling loss can be reduced if a network can be made to fit onto
two butt-coupled chips — coupling from one wave-guide directly into another is
less lossy than coupling from a fiber into a wave-guide.
• Wave-guide loss. Because the Titanium-diffused wave-guides are quite lossy, there
is considerable loss along the wave-guides and internal to the SDCs.
Manufacturing progress has lowered the loss in wave-guides and switches.
Attenuation along the wave-guides and especially internal to the SDCs can be
reduced by selecting a fabric architecture that has a reduced number of stages —
like the Benes network (Section 7.3.4). Furthermore, because amplifiers can be
simplified if we have approximately uniform loss in all paths, this suggests
selecting a fabric architecture that has an equal number of stages in every path —
again the Benes network.
Photonic Switching in Space 699

• Bend and crossover loss. Loss increases wherever wave-guides bend. Loss in
wave-guide bends can be reduced by making the bends with a relatively large
radius of curvature. There is also a significant loss wherever wave-guides cross
over each other. Wave-guides really pass through each other and photon’s
momentum keeps most of the desired optical power proceeding straight through the
intersection. Crossover loss can be reduced by making the angle of intersection
relatively large. Crossovers can be completely eliminated and wave-guide bends
can be reduced by a careful choice of architecture, but the architectures that have
no bends or crossovers are typically quite poor in other characteristics like
connectivity, number of stages, and crosstalk. Increasing the bend radii and
crossover angles requires increasing the spacing between adjacent columns of SDC
switches. This effect is exacerbated in switch matrices with many rows of SDCs.
But, fitting several stages on a single chip, to reduce coupling loss, requires
decreasing the space between the SDCs. Unfortunately, we can’t have it both
ways. Fitting several stages on a single chip also requires decreasing the length of
each SDC, which increases the switching voltage and the crosstalk. We see that
fabrication is intricately tied to operating parameters.
Crosstalk. Crosstalk in a 2 × 2 switch is quantified by its extinction ratio (ER), a
measurement that is similar to signal-to-noise ratio. Assume that two optical signals
move left-to-right through an SDC, that all four paths through the switch have identical
attenuation, and that the SDC is in the BAR state. Then, the signal detected at the upper
right port has two components.
• The desired signal, detected at this upper right output port, is an attenuated version
of the desired signal at the selected upper left input port. If P s is the input power
of the desired signal, then the signal detected at the desired output has power
r × P s . The attenuation, r, is the fraction (hopefully close to 1) of the desired input
signal that reaches the selected output and it’s assumed to be identical in all four
paths through the switch.
• Crosstalk is an undesired leakage of an attenuated version of the signal that is input
to the lower left port and that is intended to be detected at the lower right port. The
crosstalk has power r × x × P u , where r is the expected attenuation in any path
through the SDC, P u is the input power of the undesired signal, and x is the
fraction (hopefully close to 0) of the undesired input signal that reaches the
unselected output, by traversing an unselected path through the SDC.
The SDC’s extinction ratio is the ratio of these two components of the total signal
detected at the upper right output port. If all selected paths have equal attenuations and
P d = P u , then ER = 1/x, usually expressed in decibels.
Consider a signal that passes through K consecutive SDCs, each with identical loss,
crosstalk, and extinction ratio. Accurately calculating the end-to-end extinction ratio is
quite difficult because crosstalk in the stages near the output is more significant since it is
not attenuated by as many subsequent stages of switches. If we ignore attenuation and
only consider crosstalk, then crosstalk adds to the signal in each successive SDC and the
end-to-end extinction ratio = K/x (be careful not to use decibels). It would actually be
significantly better than this.
700 Telephone Switching Systems

Sensitivity. After attenuation and crosstalk, Lithium Niobate’s next most notorious
problem is its high sensitivity to small variations in most of its operational parameters.
While, initially, Lithium Niobate might seem discouraging and impractical, the
discussion below shows that the parameter sensitivities arise because of the relatively
high crosstalk requirement.

Voltage
0.707 V max

Band-
width

FL F0 FH Frequency
Figure 20.5. Power transmitivity’s dependence on frequency.
Figure 20.5 shows a classical plot of transmitivity versus frequency, illustrating the
definition of bandwidth. If the figure shows signal power transmitivity versus frequency
for some channel, then this channel’s bandwidth is the band of frequencies between F L
and F H , the frequencies for which the power transmitivity is at least 50% of the peak
mid-band transmitivity. If F H − F L is very small, the channel is said to have high Q or to
be narrowband — its transmitivity is extremely sensitive to frequency. Similar to how a
signal’s transmitted power depends on frequency and how this dependency specifies the
channel’s frequency sensitivity (quantified by the definition of bandwidth), Lithium
Niobate’s crosstalk depends on many operational parameters and this dependency
specifies its sensitivity to these parameters.

Extinction ratio
20 dB

Tolerance

XL X0 XH Parameter X
Figure 20.6. Crosstalk’s dependence on parameters.
Figure 20.6 looks like the previous figure, only the definitions of the axes have been
changed. Similar to how bandwidth is defined and specified in the previous figure, Figure
20.6 defines and specifies Lithium Niobate’s tolerance (or intolerance) to any of several
parameters. Where bandwidth is defined as the frequency range over which power is
within some specification (greater than half the peak power), parameter tolerance is
defined as the range in parameter value over which crosstalk is within some specification.
The y axis represents the crosstalk in a Lithium Niobate SDC. Figure 20.6’s x axis can
represent any one of a number of some SDC’s operational parameters.
Photonic Switching in Space 701

• If x represents the switching voltage, then the figure shows that voltage X 0 is the
ideal voltage to switch the SDC into its desired state, to maximize the ER. Some
other voltage, far removed from X 0 , is optimal for switching the SDC into its
opposite state. For other voltages, the SDC is in some in-between state, neither
BAR nor CROSS, where ER is unacceptably low at both output ports.
• Let x represent the signal’s wavelength. Suppose the perfect optimum voltage is
applied that places some SDC in its BAR (or CROSS) state. But, this optimum
voltage depends on the wavelength, X 0 , of the signal switched through the SDC. If
the signal’s wavelength changes, the SDC’s extinction ratio worsens, unless the
switching voltage changes correspondingly.
• Let x represent the signal’s polarization. Suppose the perfect optimum voltage is
applied that places some SDC in its BAR (or CROSS) state. But, this optimum
voltage depends on the polarization, X 0 , of the signal switched through the SDC.
If the signal’s polarization changes, the SDC’s extinction ratio worsens, unless the
switching voltage (or wavelength) changes correspondingly.
• Let x represent the switch’s temperature. Suppose the perfect optimum voltage is
applied that places some SDC in its BAR (or CROSS) state. But, this optimum
voltage depends on the switch’s temperature, X 0 . If the switch’s temperature
changes, the SDC’s extinction ratio worsens, unless the other parameters change
correspondingly.
• Let x represent the length of time that a switch has been in its state. Suppose the
perfect optimum voltage is applied that places some SDC in its BAR (or CROSS)
state. But, this optimum voltage only pertains to the time, X 0 , when the state
change occurs. As the switch is in this state for a longer time, the SDC’s extinction
ratio worsens, unless the other parameters change correspondingly.
Lithium Niobate has a reputation for being extremely sensitive to relatively small changes
in relatively many parameters. The problem lies in the y axis, not the x axis or the many
x axes. While the scale in Figure 20.6 isn’t shown, Lithium Niobate’s real problem is that
a reasonable ER of, say, 100:1 (20 dB) is high on the y axis. The two intercepts,
corresponding to X L and X H , represent a narrow tolerance in the parameter represented
on the x axis. Requiring that an SDC have an extinction ratio > 20 dB is what underlies
Lithium Niobate’s high sensitivity to all its operational parameters. If the SDC’s required
extinction ratio can be lower, then all the parameter dependencies can be desensitized.
But wouldn’t a lower extinction ratio requirement result in a poor bit error rate? Not
necessarily. The previous subsection discussed that Lithium Niobate’s attenuation
problem has an architectural solution. So also does Lithium Niobate’s crosstalk problem.
Device or Architectural Solution. Lithium Niobate’s problems were exacerbated by an
argument over who owned the problem — is it a device problem or is it a systems
problem? Imagine a physicist and a systems engineer discussing this technology. The
physicist tells the systems engineer, ‘‘Tell me the requirements, so I can build you a
device that works within these parameters.’’ The systems engineer tells the physicist,
‘‘Tell me the device’s inherent characteristics, so I can build an architecture that works
within these parameters.’’
702 Telephone Switching Systems

Lithium Niobate’s problems are not that this technology is overly sensitive to every
one of these parameters. Its problem is that we’ve been requiring this technology to meet
a crosstalk requirement that is unable to be met in a practical way. If we relax the
crosstalk requirement, we remove the sensitivity to all the other parameters, and LiNbO3
is OK. But, if we relax the crosstalk requirement, doesn’t the fabric become impractical?
That depends on the fabric’s architecture, not on the devices. Repeat this: that depends
on the fabric’s architecture, not on the devices. We need an architecture for SDC
switching networks in which the end-to-end crosstalk is better than the extinction ratio of
each individual SDC in the network.
The breakthrough was not discovered in the physics lab. It was a casual observation
by a systems researcher. We have crosstalk in these SDC fabrics, not because the devices
are inadequate, but because we have two talking paths in each SDC. If the extinction
ratio is too small in some switch, you can try to improve the switch or you can try to
eliminate the source of the crosstalk. If there is only one switched connection set up
through an SDC fabric, there will be no crosstalk and the end-to-end path through the
switch will be extremely insensitive to operating parameters. If there are two switched
connections set up through an SDC fabric, but the two connections have no SDCs in
common (neither connection passes through the same SDC) there will be very little
crosstalk and the end-to-end path through the switch will be extremely insensitive to
operating parameters. It is only when some SDC inside the fabric supports two different
switched connections that we have crosstalk. While this observation is obvious in hind-
sight, it had remarkable consequences.

20.4 PHOTONIC SPACE-DIVISION FABRICS


While many network architectures for photonic space switching have been proposed, only
some of these have been prototyped. We have demonstrated the proof of concept that the
technologies will work in a switching application. While much of the early work was
based on classical space-switching architectures, significant new architectures have been
proposed for time, space/time, and even space/time/wavelength switching. While optical
packet-switching networks have been proposed, the technology is more easily applied to
circuit switching.
This section describes several photonic switching fabrics that are constructed from
guided-wave analog devices and that operate in the space division only. The matrix
architecture is generalized four different ways and then, the dilated Benes network is
shown to be a good architectural solution for Lithium Niobate’s problems. The third
subsection illustrates some switching architectures suitable for construction from SLAs.
20.4.1 Generalizing the Matrix Architecture for SDCs
Many different architectures have been proposed, and even prototyped, for switching
fabrics built from SDCs. Some of them, even some that were prototyped, are really quite
silly and came about only because the implementers were very naive about the theory of
interconnection switching networks. Only two architectures are described here: the
generalized matrix, in this subsection, and the dilated Benes, in the next one.
Figure 20.7 illustrates an M × N switching matrix using four different combinations
of passive/active splitting/combining [12]. The matrix switches shown in the figure
provide full connectivity for M = 3 input ports and N = 3 output ports. The discussion
Photonic Switching in Space 703

A X A X

B Y B Y

C Z C Z

(a) Passive splitter/passive combiner (b) Passive splitter/active combiner

A X A X

B Y B Y

C Z C Z

(c) Active splitter/passive combiner (d) Active splitter/active combiner

Figure 20.7. Four variations of a photonic switching matrix. © 1987 IEEE [12].
following assumes that M = N, but this isn’t necessary. The four architectures are called:
(a) the passive splitter/passive combiner (PS/PC), (b) the passive splitter/active combiner
(PS/AC), (c) the active splitter/passive combiner (AS/PC), and (d) the active
splitter/active combiner (AS/AC).
In Figure 20.7, a smaller box, with channels shown passing through it, represents a
passive 1 × 2 splitter or combiner, while a larger box, with the channels shown
interrupted, represents an active 1 × 2 switch, like an SDC. Each input or output port is
connected to the root-side of a tree-like structure of passive or active 1 × 2 devices.
Because these figures lack the rectilinear structure and ‘‘banjo wiring’’ of a traditional
matrix switching architecture, they don’t look like matrix switches. While the physical
splitters and combiners are necessary because optical fiber doesn’t ‘‘bus’’ well, they
704 Telephone Switching Systems

perform logically like a rectilinear bus. So, each structure in Figure 20.7 operates
logically like a matrix switch.
Figure 20.7(a) is a 3 × 3 extension of Figure 7.1(c). It illustrates a simple 3 × 3
photonic switching matrix that uses an optical shutter or MZI as the switching device.
An SDC with only one input and only one output wired can also be used as a 1 × 1 switch
in this architecture. This 3 × 3 matrix requires nine switches, a generalized M × N matrix
requires M × N SDCs, and a generalized N × N matrix requires N 2 SDCs. Because
optical signals can’t simply fan out, each of the M = 3 input signals must be passively
split to the left side of N = 3 SDCs and each of the N = 3 output signals must be passively
combined from the right side of M = 3 SDCs. This splitting and combining requires two
PDCs for each port, for a total of 12 PDCs, and the structure is called the passive
splitter/passive combiner.
Each of the SDCs in the center of Figure 20.7(a) can double as a 2 × 1 device by
using two of its inputs instead of only one. So, one or both stages of passive splitting or
combining can be absorbed into the switches in the center column, reducing not only the
number of splitters and/or combiners, but also the total number of switches. These
structures have been named by whether the input splitting and/or the output combining is
passive or active.
Figure 20.7(b) is a passive splitter/active combiner. Each input signal is passively
split to three junctors in the center of the network. From there, only one would be
actively switched to the desired output. Figure 20.7(c) is an active splitter/passive
combiner. Each input signal is actively switched to one of three junctors in the center of
the network. From there it is passively combined with two other signals, presumably
empty, to make the desired output signal. Figure 20.7(d) is an active splitter/active
combiner. Each input signal is actively switched to one of three junctors in the center of
the network. From there, it would be actively switched to the desired output.
Compare the four architectures in Figure 20.7 to each other and to the Benes (described
next) [13,14].
• Attenuation. In any network architecture that uses PDCs, signals suffer 3 dB of
attenuation in every PDC they pass through. So, consider generalizing the
networks above to N × N, let each PDC contribute 3 dB and each SDC contribute 1
dB, and ignore all losses except losses in PDCs and SDCs. Then, the AS/PC’s and
the PS/AC’s losses are 4 × log2 N dB, and the PS/PC’s loss is 6 × log2 N dB. The
AS/AC’s loss is 2 × log2 N dB and the Benes’ loss is even 1 dB less than the
AS/AC’s loss. So, for an 8 × 8 network, the PS/PC’s loss is 18 dB, the AS/PC and
PS/AC have loss of 12 dB, the AS/AC’s loss is 6 dB, and the Benes’ loss is 5 dB.
• Switch count. Generalizing to an N × N architecture, the PS/PC requires N 2 SDCs,
or N SDCs per input port, and 2(N – 1) PDCs per port (N – 1 1 × 2 switches in each
binary tree). The AS/PC and the PS/AC require N – 1 SDCs per port and an equal
number of PDCs. The AS/AC requires 2(N – 1) SDCs per port and the Benes
requires 2 log2 N − 1 SDCs per pair of ports. So, for an 8 × 8 network, the PS/PC
has 64 SDCs and 112 PDCs, and the AS/PC and PS/AC have 56 of each. Needing
no PDCs, the AS/AC has 112 SDCs, but the 8 × 8 Benes network has only 20
SDCs.
Photonic Switching in Space 705

• Crosstalk. Using PDCs on the input side of a generalized matrix distributes input
signals to many undesired places in the middle of the network. Using PDCs on the
output side of a generalized matrix collects undesired signals from anywhere they
might appear in the middle of the network. Consider an 8 × 8 network, ignore
attenuation, and let each SDC’s extinction ratio be 100 (20 dB). In the PS/PC and
AS/PC, each output port collects 0.01P from each of two off-SDCs and their end-
to-end SNR is 1/0.02 = 50 (not in dB). By its symmetry with the AS/PC, the
PS/AC has the same value. But, in the AS/AC, since the highest crosstalk term
passes through two off-SDCs, its end-to-end SNR is a little less than 1/0.0001 =
10K = 40 dB. Because a signal through the Benes accumulates 0.01P of crosstalk
in each stage, the end-to-end crosstalk through a five-stage network is 0.05 and the
end-to-end SNR is 1/0.05 = 20 = 13 dB. We will describe a modification of the
Benes that brings its end-to-end SNR in line with that of the AS/AC.
• Broadcast. Because of all the PDCs, the PS/PC scores poorly in attenuation and
switch count. While using PDCs on both sides of the network contributes to the
PS/PC’s low score on crosstalk, its redeeming quality is that it allows broadcast.
Also having passive splitting on the input side, the PS/AC also allows broadcast.
20.4.2 Dilated Benes Architecture
The Benes architecture is illustrated in Figure 7.11 and was described in detail in Section
7.3.4. The Benes architecture brings several obvious benefits to photonic switching: it is
constructed from a 2 × 2 switching device (the beta element), it has the lowest number of
stages for any nonblocking network built from 2 × 2 switches, and all paths use the same
number of switches. But, the Benes architecture is rich in wave-guide bends and wave-
guide crossovers and it still has crosstalk. We solve Lithium Niobate’s crosstalk by
modifying the Benes architecture.
When we try to see in the dark, the pupils of our eyes become dilated. We
automatically compensate for a lack of light by making the transparent opening larger, so
more light can get into our eye. Similarly, we will architecturally dilate a switching
network by providing a bigger optical channel for the light that goes through any
switching network. Dilation is a transformation on any switching network architecture
under which the network has more optical paths for the same amount of light as in the
untransformed network. In this section, we will dilate the Benes architecture, but any
known architecture can be dilated [15].
After describing dilation in general, the dilated Benes architecture is developed by an
inductive construction: the elementary 2 × 2 network is developed, a recursive
construction procedure is presented, and the induction is completed. Then, an 8 × 8
dilated Benes network (DBN) is illustrated and its prototype is described.
Dilation. Consider the 8 × 8 Benes network in Figure 7.11. Suppose all the ports on
both sides are idle. Now, suppose the upper left port wishes to be connected to the upper
right port. There are four different paths, one through each different 2 × 2 switch in the
center column. Let the upper path be selected arbitrarily. Notice two things.
• While the path goes through one switch in each stage of the Benes, it uses only half
of each 2 × 2 switch. Each switch in this path could also be used as part of one
other path.
706 Telephone Switching Systems

• There is absolutely zero crosstalk in this path, not because we picked a good path,
but because there are no other calls through the network.
If a second connection is added to this first connection, there may (or may not be) any
practical level of crosstalk, depending on whether the second path can be found through
the fabric such that the paths use different switches in every stage; that is, no one switch
is used for both paths. This may (or may not) be possible.
• Certainly if the second connection starts from the second port on the left side (or
terminates on the second port on the right side), there will be crosstalk, because
both paths must use half of the upper left switch (or upper right switch).
• If the second connection starts and terminates on the bottom half of both sides of
the network, there will be crosstalk only if we select the path that uses the upper
switch in the center column. If we select any of the other three paths for this
connection, there will be practically zero crosstalk. There will still be some
crosstalk, and the first connection will notice an increase in crosstalk, because some
unwanted signal can leak through more than one switch.
Suppose an N × N Benes fabric is fully connected (has N connections through it). Then,
because each column holds N/2 2 × 2 switches, each switch in the entire fabric must
support two paths. If the Benes fabric is implemented with photonic SDCs, each of the N
complete paths through the fabric is exposed to crosstalk from some other path in each
stage of the fabric.
Consider this observation. Do we have crosstalk because of the SDC’s inadequacy?
Or, do we have crosstalk because we’re too greedy — we’re trying to use both paths
through every SDC?

A Y

B Z

Figure 20.8. 2 × 2 dilated switching module.


2 × 2 Dilated Network Using SDCs. Dilation is illustrated on the most elementary
network, a 2 × 2. Even though the network in Figure 20.8 has four switches, in two rows
and two stages, it’s functionally equivalent to a single SDC. The network has only two
useful configurations, corresponding to an SDC’s BAR and CROSS states. If all four
switches are placed in their BAR state, A connects to Y and B connects to Z. If all four
switches are placed in their CROSS state, A connects to Z and B connects to Y. Notice
that, in both configurations, the two coexisting paths never share a common switch.
Consider the network of Figure 20.8 in its BAR configuration, with all four SDCs in
their BAR states. Consider the four junctors in the center of the network. The upper
junctor carries most of the signal that A transmits to Y. Some of this signal is attenuated
in the upper left SDC, where some leaks to the junctor that proceeds to the lower right.
The lower junctor carries most of the signal that B transmits to Z. Some of this signal is
attenuated in the lower left SDC, where some leaks to the junctor that proceeds to the
Photonic Switching in Space 707

upper right. Because one input port on each first-stage SDC is unterminated, the unused
path through each of these SDCs carries absolutely no photons. So, the signals in the two
straight-across junctors are not corrupted by any crosstalk. But, the two diagonal junctors
carry leakage signals to the second stage. Here, in this second stage, the actual signals
are nominally switched to their respective output ports and the leakage signals are
nominally switched to the SDCs’ unterminated output ports. But, the leakage signals can
leak onto the actual signals inside these second-stage SDCs. While there is some
crosstalk in the output signals, the crosstalk is second order; it must leak through two
SDCs.
Some observations are made.
• If each SDC’s ER = X, then the network’s end-to-end extinction ratio is X 2 . So, for
example, if system BER requires that end-to-end ER = 22 dB, then each SDC must
operate with an individual switch ER = 11 dB. This figure is so easily attainable
that SDC parameters can be greatly relaxed.
• The network of Figure 20.8 is wide-sense dilated, but only because the control
algorithm is trivial (there is only one path for each pair of endpoints). While not
clear from this example, dilated networks are generally strict-sense dilated because
the network path must be carefully selected. In general, the number of paths
through a dilated network is the same as the number of paths through its non-
dilated parent.
• Where the regular 2 × 2 Benes network has one single SDC, the 2 × 2 dilated
Benes network has four SDCs. In general, a dilated network has more than twice
as many switches as its nondilated parent.
• Where the regular 2 × 2 Benes network has one stage of switching, the 2 × 2
dilated Benes network has two stages. In general, a dilated network has one more
stage (not double) than its nondilated parent.
• Compare Figure 20.8 to a 4 × 4 nondilated Benes network. Figure 20.8 results
from doing two things to a 4 × 4 nondilated Benes network: (1) excise its center-
stage, and (2) use only one port on each edge switch. This generalizes.
N × N Dilated Network by Recursion. In Benes’ recursion, described in Section 7.3.4,
an N × N Benes network is built by wiring together a column of 2 × 2 switches on the
left, a column of 2 × 2 switches on the right, and a center stage consisting of two parallel
identical (N/2) × (N/2) Benes networks. If the resulting N × N Benes network must
provide a single connection, from one of the N left ports to one of the N right ports, it can
choose from among N/2 different paths through the fabric. Any one of N/2 different
paths can be selected.
In a similar construction, illustrated in Figure 20.9, an N × N dilated network of type
X is built by wiring together a column of 1 × K switches (commutators) on the left, a
column of K × 1 switches (decommutators) on the right, and a center stage consisting of
K parallel identical N × N regular networks of type X. If each of the interior
subnetworks is a regular N × N Benes fabric, then Figure 20.9 is the template to create
an N × N dilated Benes network. If the N × N network in Figure 20.9 must provide a
single connection, from one of the N left ports to one of the N right ports, its controller
708 Telephone Switching Systems

1×K K×1
switch #1 switch
N × N network
of type X

N ports N ports

#K
N × N network
1×K of type X K×1
switch switch

Figure 20.9. Template for dilation. © 1987 IEEE [15].


can choose from among N/2 different paths through each of the K interior subnetworks in
the center stage. Any of the K × N /2 selected paths uses half of one 2 × 2 switch in each
stage. Obviously, if a network supports only one active path, this path cannot be affected
by crosstalk from any other path through the network.
If a second connection is added to this first connection, there may (or may not) be
crosstalk, depending on whether two paths can be found through the fabric such that the
paths use different switches in every stage; that is, no one switch is used for both paths.
If both paths use the same interior subnetwork, and both paths through this subnetwork
have no switch in common, then the two paths will not mutually crosstalk. If switch
sharing is unavoidable when two paths use the same interior subnetwork, crosstalk can
still be avoided by routing the second connection through a different interior subnetwork;
assuming that K ≥ 2.
If the network in Figure 20.9 is fully connected (has N connections through it), we
want to select K to be large enough so that none of the N paths through the fabric shares
any switch with any of the other paths, under any switching configuration. Such fabrics,
where every path through the fabric uses switches that are not shared by any other paths,
are called dilated fabrics. If no switch in the network of Figure 20.9 ever carries more
than one path, then there is no first-order crosstalk. While dilation is an optical term that
implies that the switch is somehow open to more light than before, the concept could be
applied to an electronic fabric if one were built from devices that have too much inherent
electronic or electromagnetic crosstalk.
So, can this work for some finite value of K and if so, how large does K have to be?
Obviously, K has to be larger than 1 — hopefully, not too much larger. With a maximum
of N paths through an N × N network, replication by K = N would have to completely
eliminate all first-order crosstalk. But, replicating a nonblocking network K = N times
Photonic Switching in Space 709

would produce a very inefficient and expensive fabric. While it’s probably not obvious,
we’ll show that K = 2 is enough replication to implement dilation.
The Inductive Step. It’s probably not intuitive that K = 2 would be sufficient for
dilation. The convincing argument is a constructive proof by induction. Figure 20.8
illustrates the basis of the constructive induction.
• Figure 20.9 shows the architecture of a generalized dilated network. Figure 20.8 is
seen to be a special case, where N = 2 and K = 2. In the leftmost column, each 2 ×
2 SDC has an unused left port. Each represents a 1 × K commutator, for K = 2.
The rightmost column is similar and the stage containing K N × N switching
modules in Figure 20.9 is empty in Figure 20.8.
• Figure 7.11 shows the architecture of a generalized Benes network. Figure 20.8 is
seen to be a 4 × 4 Benes network, but with two modifications. (1) The center
column of 2 × 2 switches is removed from the 4 × 4 Benes, and the columns
adjacent to the center column are reconnected. (2) On the outer edge of the outer
stages of the 4 × 4 Benes, only one of the two ports on each switch is used.
So, the switching module of Figure 20.8 operates as a 2 × 2 switch. The two inputs can
be connected to the two outputs via a BAR or CROSS configuration of the module. In
either configuration of the module, no two connections use the same switch. The module
is dilated. When the module of Figure 20.8 is used in the inductive construction, all of its
ports are used.
Now, use the template of Figure 20.9, with K = 2, to construct an N × N dilated
Benes network. It has K = 2 interior subnetworks, each a regular N × N Benes network.
Its leftmost and rightmost stages use 1 × 2 devices, which could be implemented as 2 × 2
SDCs with an unused port on their outer edge. So far the construction algorithm for an
N × N dilated Benes network is the same as that for a 2N × 2N regular Benes network
except that in the dilated version, each of the 2 × 2 switches on the two edges is
connected to only one external port. Obviously, a network can’t be dilated if its edge
switches are connected to two ports. It turns out that in the dilated Benes network the
center column of switches is redundant and can be excised from the architecture without
affecting its properties.
The architecture for any fabric constructed from 2 × 2 devices can be transformed
into its equivalent dilated architecture. Furthermore, dilation can be accomplished by
only requiring K = 2.
8 × 8 Dilated Benes Network. Figure 20.10 shows an 8 × 8 dilated Benes network
[16,17]. The 2 × 2 SDCs in the leftmost and rightmost columns are connected to only
one external port and have an unused port. The four interior columns of SDCs comprise
two separate interior subnetworks, each a 4 × 4 dilated Benes network, but with
connections to both ports on each edge switch. The network in Figure 20.10 would be a
regular 16 × 16 Benes network if all 16 ports were connected on both edges and the
center column of SDCs were replaced.
An N × N dilated Benes network has two more stages of switching, compared to a
regular N × N Benes network, but one is compensated by the excised center column.
While the number of stages is important in a photonic network because of the loss
incurred in each stage, adding one stage of switching over the minimum number seems
710 Telephone Switching Systems

A S
B T
C U
D V
E W
F X
G Y
H Z

Figure 20.10. 8 × 8 dilated Benes network. © 1987 IEEE [15].


like a small penalty when it solves Lithium Niobate’s crosstalk problem. An N × N
dilated Benes network also has twice as many 2 × 2 switches per stage, compared to a
regular N × N Benes network. However, recalling that SDCs are very thin, Lithium
Niobate technology is easily squandered in this dimension. In practice, this has little
effect on the level of integration in the fabrication of the photonic chip.
An 8 × 8 dilated Benes network was built, using Lithium Niobate SDCs, for the
DiSCO project, described in the next chapter. The entire network is enclosed in a
physical package, about 1 inch by 8 inches. The package has two ribbons of eight fibers
each and 96 pins for the electrical control of the network’s 48 SDCs. While the
network’s entire structure could not fit on one chip of Lithium Niobate, half of the
structure could. Since Figure 20.10 is symmetric about a vertical line up the center,
DiSCO’s 8 × 8 DBN was built on two identical chips that were butt-coupled inside the
package. Each input fiber required an individual polarization controller. The typical
switching voltage was around 9V.
The four generalized matrix architectures and the regular Benes were all compared at
the end of the previous subsection. The DBN’s characteristics should be also compared.
The DBN’s attenuation would be comparable to that of the AS/AC, only slightly worse
than the loss in the regular Benes network, which has the minimum number of stages for
any fully connected network architecture built from beta elements. While the DBN’s
switch count is more than double that of the regular Benes, it’s still better than any of the
generalized matrix architectures. An 8 × 8 DBN has 48 SDCs. But, because of the shape
of an SDC and its integrability, doubling the switch count over the regular Benes is not
that expensive. Because, in a DBN, a signal accumulates second-order crosstalk in each
SDC that it passes through, using the assumptions of this comparison, the DBN’s end-to-
end SNR is about 1/0.0006 = 1.667K = 33 dB, a little poorer than the AS/AC’s end-to-
end SNR, but much better than any of the others. The DBN can only provide point-to-
point circuit-switched connections and cannot support broadcast.
Just as dilation is a useful architectural design to solve Lithium Niobate’s crosstalk
problem, it is similarly useful for solving Lithium Niobate’s polarization problem.
Optical signals can be separated into their orthogonal polarization components, each
component can be switched through a separate network, and then the separate
Photonic Switching in Space 711

polarizations can be added together again. This doubling of the fabric completely
eliminates the need for polarization adjusters and all the manual attention they require.
20.4.3 SLA Switching Networks
This subsection describes the semiconductor laser amplifier (SLA), not as an amplifier as
we did in the previous section, but as a photonic switching device [18]. First, the SLA is
shown to be a 1 × 1 switch with unusual properties. Then, a 2 × 2 switch is designed as a
small fabric with four SLAs and its states and control are described. Then, many
different matrix and modular photonic switching architectures are compared.
SLA as a 1 × 1 Optical Switch. While EDFAs have become the optical amplifier
technology of choice, at 1.5 microns, the SLA still has an important role. Not only is the
SLA the only feasible optical amplifier at 1.3 microns, but it has an important role in the
future at both wavelengths because it can be used as an electronically controlled photonic
switching device. While the Lithium Niobate SDC has many device drawbacks, the SLA
is an excellent photonic switch.
• While the EDFA has an extremely slow switching time, the SLA can be turned on
and off as quickly as an SDL, in the order of 1 ns.
• While the SDC has a very poor extinction ratio, requiring tight parameter
tolerances or architectural cleverness, the SLA’s ER is better than good. When an
SLA is not forward biased, virtually none of the input photons (as low as -40 dB)
reach the output fiber. When an SLA is forward biased, it not only allows a signal
to pass, it amplifies any signals (by as much as +20 dB) that pass through it.
Extinction ratios have been observed as high as +60 dB.
The SLA’s most significant drawback is that, because it is only a 1 × 1 switch (single-pole
single-throw), even simple network architectures need many devices. While conventional
SLAs don’t promise to integrate even as well as the SDCs, a highly integrable variation of
the SLA, manufactured in Indium Phosphide, appears to be extremely promising.
Elementary 2 × 2 Network. A 2 × 2 switching module can be built easily and obviously
with SLAs as the switching elements, shown in Figure 20.11. This figure is identical to
Figure 7.1(c). Each optical input feeds a 1 × 2 splitter, and each optical output is fed
from a 2 × 1 combiner. The module has four SLAs in a logical 2 × 2 array. The ith input
is connected to the jth output by operating the <i,j>-th SLA.
If each SLA has enough gain to compensate for the net optical loss in the module,
the output signal will have the same intensity as the input signal. For example, if fiber-to-
device coupling losses and splitter/combiner losses are all -3 dB, then the SLAs in Figure
20.11 would each need +12 dB of optical gain. This is not unreasonable.
Switching States and Their Control. Besides having the potential for 0 net loss, this 2
× 2 optical switching module has another advantage over a 2 × 2 LiNbO3 SDC. While
the 2 × 2 SDC has only two states, BAR and CROSS, the 2 × 2 SLA module is more
versatile.
• The 2 × 2 SLA module can be placed in a broadcast state. The ith input is
connected to the 0 and 1 output by operating both the <i,0>th and the <i,1>th SLA.
This broadcast state doesn’t affect the net gain. The signal at the ith input appears
712 Telephone Switching Systems

0 0

1 1

Figure 20.11. 2 × 2 SLA network. © 1988 IEEE [18].


on both outputs at the same power level. Of course, if the module is in one of its
broadcast states, the other input signal is blocked.
• The SDC is typically in its BAR or its CROSS state, regardless of whether the
inputs carry a true optical signal. While this simplifies the SDC’s control, it means
that any optical noise on either input to a switch always has a path through the
switch to one of the outputs. The SLA module, on the other hand, can by placed in
a ‘‘neither’’ state. This added flexibility ensures a quiet network.
The logical operation of a 2 × 2 network built from SLAs is summarized in Table 20.1.
TABLE 20.1. Control states for a 2 × 2 SLA network

0,0 0,1 1,0 1,1 Interpretation


0 0 0 0 Totally disconnected
1 0 0 0 Input 0 to output 0 only
0 1 0 0 Input 0 to output 1 only
1 1 0 0 Input 0 broadcast to both outputs
0 0 1 0 Input 1 to output 0 only
0 0 0 1 Input 1 to output 1 only
0 0 1 1 Input 1 broadcast to both outputs
1 0 0 1 BAR state
0 1 1 0 CROSS state
Matrix SLA Networks. The matrix construction of the 2 × 2 network can be generalized
to construct an optical switching network of any size. Figure 20.7(a) shows the passive
splitter/passive combiner architecture of a 3 × 3 photonic matrix network that uses SDCs
as only 1 × 1 switches. A 3 × 3 SLA photonic matrix network has an identical
Photonic Switching in Space 713

architecture. Following the one-stage construction described in Section 7.2.1, an N × N


SLA matrix network requires N 2 SLAs.
Generalizing from Figures 7.1(c) and 20.7, besides requiring N 2 SLAs, an N × N
matrix network also requires 2N 1 × N splitters, one for each input and one for each
output. If a 1 × N splitter is built from 1 × 2 splitters, then each such splitter tree requires
N – 1 individual 1 × 2 splitters. So an N × N photonic matrix network needs 2 (N − 1) 1
× 2 splitters. No matter what the implementation, any 1 × N optical splitter (and
combiner) has 3 log2 N dB of net loss.
Modular SLA Networks. The 2 × 2 SLA network can be generalized in another way.
The 2 × 2 SLA network of Figure 20.11 can be encapsulated into a 2 × 2 photonic
switching module. Then, this module can be used as a beta element in constructing a
larger network, like an N × N Benes network.

A W

B X

C Y

D Z

Figure 20.12. 4 × 4 Benes network from 2 × 2 SLA modules. © 1988 IEEE [18].
Figure 20.12 shows a 4 × 4 Benes network where each of the component modules is
the 2 × 2 SLA matrix from Figure 20.11. The internal components inside each module
are shown in the figure. Observe that each module’s output comes from the 1-side of a 2
× 1 combiner and that each module’s input goes into the 1-side of a 1 × 2 splitter. The
module boundary disguises an architectural optimization: that both functions can be
performed by a single 2 × 2 device.
This optimization is illustrated in Figure 20.13. Note that the interstage splitters and
combiners are merged and the module boundaries are eliminated. While it doesn’t look
like one, Figure 20.13 is a 4 × 4 Benes network. Figure 20.12 is better than Figure 20.13
if it is desired to construct a stage as an integrated-circuit module. Then, the 4 × 4 Benes
would be made from six identical modules. Figure 20.13 is better than Figure 20.12 if
the network is built from discrete components or if the entire network is built as one
integrated circuit.
714 Telephone Switching Systems

A W

B X

C Y

D Z

Figure 20.13. Nonmodular 4 × 4 Benes network from SLAs. © 1988 IEEE [18].
Comparing SLA Architectures. Comparing the relative characteristics of the different
SLA network architectures is quite complicated. Table 20.2 summarizes only the
numbers of switches and splitters that are needed in each respective architecture.
Different fabric architectures impose different requirements on the reflectivity, gain, and
other physical requirements of the SLA devices that are used to construct each fabric.
The details are beyond the book’s intended scope.
When an SLA is operated as a switch, it is an analog switch, similar to an SDC. The
SLA’s big advantage over the SDC is that the SLA amplifies the signals that flow through
it. While the SLA is a 1 × 1 switch and the SDC is a 2 × 2, 2 × 2 structures can be
fabricated using SLAs and passive optical couplers.
In a relatively new technology, network architectures are diffused onto an integrated
circuit, typically fabricated in Indium Phosphide (InP). Available components include
wave-guides, splitters, combiners, and fiber couplers; the wave-guides themselves are the
1 × 1 switches. This InP technology appears (at least, to this author) to hold the future of
photonic switching.
TABLE 20.2. Architectural and device specifications for SLA networks

Size Network # SLAs # Splitters


2-by-2 Matrix/Benes 4 4
4-by-4 1-Stage Matrix 16 24
4-by-4 Modular Benes 24 24
4-by-4 Non-Mod Benes 24 16
8-by-8 1-Stage Matrix 64 112
8-by-8 Modular Benes 80 80
8-by-8 Non-Mod Benes 80 48
16-by-16 1-Stage Matrix 256 480
16-by-16 Modular Benes 224 224
16-by-16 Non-Mod Benes 224 128
Photonic Switching in Space 715

20.5 PHOTONIC PACKET SWITCHING


While the previous section and the next chapter are based on circuit switching in
photonics, this section describes packet switching in optical technology. Packet
switching further exacerbates the thermodynamic problem discussed in Section 20.2. Not
only must every bit, at least in the packet’s header, be individually switched, the entire
header typically must also be stored and operated on with software. While a format for
packet switching, optimized for low thermodynamics, is described in this section, packet
switching is seen to have an inherent disadvantage at extremely high data rates.
Besides requiring that every packet be individually processed, packet switching’s
other major drawback is the lack of a really good way to handle congestion and
collisions. Packet congestion can be handled by software or by hardware. If handled by
hardware, packets can either be (1) queued up in memories or (2) deflection-routed into
the network. Because there is no photonic software, photonic networks have to use one
of the hardware-based solutions. Since the serial delay-line is the only practical form of
optical memory, packet queuing in optical packet-switched networks must use optical
fiber as memory. While optical fiber can be dedicated for storing packets, a photonic
network has fiber between its nodes anyway that could be used for memory and trans-
mission. So, most proposed packet-switched photonic networks use deflection routing.
This section’s two subsections describe two different styles of architecture for
photonic packet switching, each reminiscent of previously described systems. The first
subsection describes a progressive packet-switching network, reminiscent of Step-by-
Step (Chapter 6), for use with conventional packet formats. The second subsection
describes a technique based on a nonconventional header multiplexing.
20.5.1 Self-Routed Packet-Switched Networks
Assume that packets have the conventional format, with the header at the beginning.
Further assume that a binary-coded destination address is placed at the very beginning of
the header. Even more restrictive, assume that this address is not an absolute end-marked
name for the node but, instead, is the complete relative routing direction that steers a
packet from its source through some network to its destination.
Assume the source node got this information during the call setup of the virtual-
circuit connection and installed it into the packet’s header before launching the packet
into the network. Let this address contain a field for each node the packet passes through,
in the order of the hops. Let the entry in each node’s field be the address of this node’s
exit port that routes the packet to the next node in the path. Examine one node in this
network and assume that each arriving packet’s header begins with the binary code for the
proper exit port from this node. Let the node strip this information away so that, after
exiting the node, the packet’s header begins with the code for the next node’s exit port.
Inside the network node, incoming packets find their way to the assigned output port
by routing themselves through the node’s internal switching fabric. The fabric’s basic
structure has the Banyan architecture discussed in Section 7.5. When a packet arrives at a
2 × 2 switch in the Banyan, the first bit determines which output the remainder of the
packet should use. The steering bit is stripped off and the packet proceeds on to the next
stage of the network. This logical operation can’t be implemented in photonics (at least,
not in the foreseeable future). So, the implementation must be in optoelectronics, perhaps
SEEDs, or in electronics with O-to-E conversion on the fabric edges.
716 Telephone Switching Systems

Because each switch in the Banyan has two inputs, two packets could arrive
simultaneously. Both these packets could be destined to the same switch output. While
packet headers could be examined ahead of time so that collisions could be avoided by
buffering at the node’s input, this seems extraordinarily complicated, especially at high
data rates. If the Banyan is electronic, packet-sized buffers could be implemented as part
of each 2 × 2 switch, but with considerable performance and cost penalty. So, in
photonics, deflection routing seems like the best, if not the only, solution.
Consider one arbitrary node in a large network. Let this node have an internal
Banyan, a fabric of 2 × 2 switches. If two packets arrive at the same 2 × 2 switch and
both packets are destined for this switch’s same output port, then the switch delivers one
of the packets to the wrong port. Here are three (are there more?) alternative strategies.
• External network-wide deflection. The packet that goes to the wrong port on this
switch will progress on to the wrong output port on the internal Banyan. From
here, it is transmitted to the wrong next node. If the packet’s header also contains
the destination address, and not just the virtual circuit’s path, then this wrong node
can correct the address and redirect the packet. This procedure doesn’t seem
amenable to photonics.
• Instead of letting the packet go to a wrong node, the node can make another
attempt to steer the packet to the correct node. If a packet is steered to some
switch’s wrong port, then the switch might have the internal logic to change the rest
of the bits in the address and, thereby, steer the packet to an output port on the
Banyan that does not lead on to another node. There are two internal strategies.
• Feedback around the Banyan. Some of the Banyan’s output ports can feed
back around to its input ports. This is especially effective when the Banyan
is combined in tandem with a Batcher sorting network, as in two prototype
electronic packet switching networks, called Starlite and Sunshine.
• Iterate the Banyans. Banyans can be cascaded serially so that the correct
ports all lead to the corresponding next node and the wrong ports lead to a
subsequent Banyan still in this same node.
All these schemes are more amenable to electronic technology than to photonic
technology. However, somebody should try to build one using SEEDs.
20.5.2 Header Multiplexing
Recall that block multiplexing is a modification of the classical digital multiplexing
format that makes circuit switching more amenable to photonic technology. Similarly,
header multiplexing is a modification of the conventional statistical multiplexing format
that makes packet switching more amenable to optoelectronic technology. The
conventional packet format is ordered so that the packet’s header precedes the user’s data,
carried behind in the packet’s body. In electronic packet switching, the header is
processed after the entire packet is stored in the receiving node. In photonic packet
switching, while the header must still be processed electronically, we want the body to
remain in optical form and we want to avoid storing it in the receiving node.
With header multiplexing, the header for packet n + 1 coincides with the previous, or
nth, packet. This format allows the receiving node’s processing unit ample time to
Photonic Switching in Space 717

determine a packet’s next destination before the data actually arrives. Header
multiplexing can be implemented in space, time, or wavelength. The conventional packet
format is easily converted to the header multiplexed format in the time domain. The
header is stripped off an arriving packet and is delivered to the receiving node’s
processing unit, but the packet’s body is delayed by one timeslot in a fiber delay line.
With high data rates, even this conventional format may need to be modified by the
addition of a temporal guard band between the header and the body.
In space or wavelength, packet headers and packet bodies use separate time-
synchronized channels. In timeslot t, the header channel carries the header for the packet
that will arrive on the body channel in timeslot t + 1. The header channel and the body
channel can be two separate optical fibers or two separate wavelengths on the same fiber.
It is important to see that, while a packet’s header is much smaller than its body, they
occupy the same amount of time on their respective channels. So, the header channel can
not only have a lower line rate than the body channel, but the header channel’s line rate
can remain constant even if the packet’s body size is doubled (and also doubling the
optical line rate on the body channel).

In 0 b b Out 0
Dmx SDC Mux
h h
SDC
In 1 b b Out 1
Dmx SDC Mux
h h
Photonic Slow Fast Fast Slow
Electronic O-E O-E O-E O-E

Header
processor

Local
gateway

LAN
Figure 20.14. 2 × 2 Add-Drop packet-switching element. © 1994 IEEE [19].
Figure 20.14 shows a 2 × 2 switching element that performs packet switching in a
photonic environment with header multiplexing [19]. Two mutually synchronized input
packets enter the element over two separate physical channels at the left of Figure 20.14.
Both packets’ headers are O-to-E converted and are delivered to the electronic header
processing unit (HPU). If the headers are wavelength multiplexed, as assumed in the
figure, then they are simply demultiplexed, as shown. If the header channels are
physically separated, then the node has no demultiplexors and the input header channels
would connect directly to the O-to-E converters. If the packet format is conventional, but
with a header-to-body guard band, then the headers are switched, instead of being
demultiplexed. Figure 20.14 assumes that each packet’s header arrives one timeslot
ahead of the body of the packet for which it provides the signaling. If each packet’s
718 Telephone Switching Systems

header and body arrive in the same timeslot, then the body of the packet would be
delayed, by inserting the appropriate length of fiber between the demultiplexors and the
two leftmost SDCs.
The HPU has an entire timeslot in which to process both packet headers. If either (or
both) packet’s final destination is the LAN for which this node is the gateway, then the
HPU puts the corresponding leftmost SDC(s) in its CROSS state and the packet’s body
would be O-to-E converted and dropped to the local gateway processor. At the same
time, if the local gateway processor has an appropriately addressed local packet in its
internal queue, then this packet would be E-to-O converted and added for output
transmission in the place of the retained packet. If neither packet is local, then the HPU
puts both leftmost SDCs in their BAR states and both incoming packets proceed to the
rightmost SDC.
The HPU routes the two incoming packets, or any local packets that are added in
their place, by putting the rightmost SDC in its BAR or CROSS state, appropriately. If
both packets are destined for the same output link, then one would be deflected. If
deflection is not tolerated, one of the packets could be stored locally in the gateway and
transmitted in a later timeslot.
Figure 20.14 shows a photonic packet-switching node as if it is a node in a large
WAN. A similar architecture, without the local gateway, could be used to implement a 2
× 2 switching element for use in a Banyan or other self-directed packet-switched fabric.

EXERCISES
20.1 Compare the bandwidths of a 4-kHz analog telephone channel, a coaxial cable
carrying 70 4-MHz television signals, and two different ways for measuring
optical fiber’s ultimate bandwidth. Make the comparisons by using analogies to
pipe diameters. (a) Let 3/8-inch copper tubing represent the 4-kHz analog
telephone channel. Using its inside diameter of 1/4 inches, compute the ratio of
bandwidth to square feet. Use this ratio for the other three analogies. (b)
Compute the corresponding pipe diameter for a coax carrying 280 MHz. (c)
Designating the narrow bandwidth of optical fiber as a 60-nm window around 1.3
microns and a 90-nm window around 1.5 microns, find the optical frequencies for
each of the four limiting wavelengths and the frequency bandwidth of each
window. Then, compute the corresponding pipe diameter for the sum of both
windows. (d) Designating the broad bandwidth of optical fiber as the complete
wavelength window between 0.6 microns and 1.6 nm, find the optical frequencies
for each of these limiting wavelengths and the frequency bandwidth of this
window. Then, compute the corresponding pipe diameter for this window.
20.2 Consider a semiconductor device that dissipates 1 nJ each time it operates. (a)
How much is the maximum data rate, limited by thermal transfer? (b) If some
integrated circuit contains 10,000 such devices, operating concurrently, how much
is the maximum data rate, limited by thermal transfer? (c) If a single discrete
device operates in analog mode on 1000-bit blocks of data, how much is the
maximum data rate, limited by thermal transfer? (d) If some integrated circuit
contains 100 devices, operating as in part (c), how much is the maximum data
Photonic Switching in Space 719

rate, limited by thermal transfer?


20.3 What are the extinction ratios, in decibels, for a 2 × 2 switch that separates its
input signals by 80/20, 95/5, 99/1, 99.9/0.1?
20.4 Appproximate the curve in Figure 20.6 by the Gaussian probability function. Let
X 0 on Figure 20.6 correspond to the Gaussian mean and let X H and X L
correspond to a Gaussian tolerance of plus/minus 0.1 standard deviation. Scale
Figure 20.6 by assuming that some parameter X is specified at a value of 100 and
that the 20-dB extinction ratio requires a numerical tolerance of 98.5 < X < 101.5.
(a) What is the corresponding standard deviation? (This is approximately the
Gaussian scale for IQ.) Suppose dilation opens the equivalent Gaussian tolerance
to plus/minus 1.0 standard deviations. (b) What is the corresponding numerical
tolerance?
20.5 Consider a 2 × 2 matrix switch built with SDCs and passive splitters and
combiners. Corresponding to Figure 20.7, sketch the four implementations:
PS/PC, PS/AC, AS/PC, and AS/AC. In the PS/PC, use the SDCs as 1 × 1
switches. In the other three, use the SDCs as 1 × 2 and/or 2 × 1 switches.
20.6 Assume that both input signals to a 2 × 2 optical switching fabric have power = P,
that passive splitters and combiners are perfect at 50/50, and that the SDCs
perform an 80/20 separation of input signals. Compare the signal-to-crosstalk
ratio at either output from five different architectures for a 2 × 2 fabric built using
SDCs: a single SDC used as a 2 × 2 switch and the four implementations of the
matrix switch from the previous exercise.
20.7 Sketch a 4 × 4 dilated Benes network. Input 0 can be connected to output 0 over
two different paths. Arbitrarily select the upper path and show it on your sketch.
While input 1 could be connected to output 1 over two different paths in an idle
network, in the presence of the previous 0-to-0 connection, only one of these
paths is usable (and it’s dilated). (This illustrates that dilation is strict-sense —
see Section 7.3.1. Arbitrarily selected paths are not necessarily dilated; an
intelligent control algorithm has to be used.) Add the dilated 1-to-1 connection to
your sketch. Input 3 can be connected to output 3 over two different connections
and, even in the presence of the previous 0-to-0 and 1-to-1 connections, both
paths are dilated. Arbitrarily select the lower path and show it on your sketch.
Now assume that the 1-to-1 connection is released. Redraw your sketch, showing
only the 0-to-0 and 3-to-3 connections. Now, suppose input 1 must be connected
to output 2. While there are two connections, how many are dilated? (This
illustrates that dilation in the dilated Benes is only rearrangeable — see Section
7.3.3.) Show a rearranged network pattern that supports 0-to-0, 1-to-2, and
3-to-3, all over dilated paths.
20.8 Control of a 2 × 2 network built from semiconductor laser amplifiers (SLA) is
illustrated in Table 20.1. Design a digital control unit for this network. This is a
difficult digital design problem and shouldn’t be attempted by a beginner.
Because there are nine states in the table, the network must be controlled by a
4-bit word, abcd, the inputs to the controller. Since there are four SLAs to
720 Telephone Switching Systems

2×2
SLA matrix

Controller

a b c d

Figure 20.15. Controller for an SLA 2 × 2 switch.


control, the controller has four output signals, one to each SLA. Assign the four
input variables as: a = broadcast or not, b = input(s) are connected to
corresponding or opposite output(s), c = enable input 0 or not, d = enable input 1
or not. Show the 4-bit code that specifies each of the nine states in Table 20.1.
Taking advantage of don’t-care code words, design the controller. [I found a
solution requiring only five two-input logic gates. Can you find it? Can you beat
it? There is a lot of opportunity with how don’t-cares are assigned.]
20.9 Consider optical packets at the node shown in Figure 20.14. Suppose three
consecutive packets arriving over In 0 are destined for Out 0, the local node, and
Out 0, respectively. Synchronous with these three packets, suppose three
consecutive packets arriving over In 1 are destined for the local node, Out 1, and
Out 0, respectively. Suppose several packets are queued up in the local gateway
that are destined for both output ports. Assume that, while incoming packets may
have to be deflected to the wrong output, local packets wouldn’t be transmitted
unless the correct output is available. Walk the incoming and outgoing packets
through the node. Describe the state of all SDCs in each packet timeslot and
identify any deflected packets.

REFERENCES
[1] Smith, P. W., ‘‘On the Physical Limits of Digital Optical Switching and Logic Elements,’’ The Bell System
Technical Journal, Vol. 61, No. 8, Oct. 1982, pp. 1975 – 1993. © 1993 AT&T.

[2] Smith, P. W., ‘‘On the Role of Photonic Switching in Future Communications Systems,’’ IEEE Circuits
and Devices Magazine, May 1987.

[3] Cloonan, T. J., et al., ‘‘A 3D Crossover Switching Network Based on S-SEED Arrays,’’ in Photonic
Switching II, pp. 196 – 199, K. Tada and H. S. Hinton (eds.), Berlin, Germany: Springer-Verlag, 1990.

[4] Hinton, H. S., et al., ‘‘Photonic Switching Fabrics Based on S-SEED Arrays,’’ in Photonic Switching II,
pp. 30 – 39, K. Tada and H. S. Hinton (eds.), Berlin, Germany: Springer-Verlag, 1990.

[5] Jopson, R. M., and G. Eisenstein, ‘‘Optical Amplifiers for Photonic Switches,’’ in Photonic Switching, pp.
84 – 87, T. K. Gustafson and P. W. Smith (eds.), Berlin, Germany: Springer-Verlag, 1988.
Photonic Switching in Space 721

[6] Kobayashi, S., and T. Kimura, ‘‘Semiconductor Optical Amplifiers,’’ IEEE Spectrum, May 1984.

[7] Thylen, L., P. Granestrand, and A. Djupsjobacka, ‘‘Optical Amplification in Switching Networks,’’ in
Photonic Switching, pp. 212 – 215, T. K. Gustafson and P. W. Smith (eds.), Berlin, Germany: Springer-
Verlag, 1988.

[8] Alferness, R. C., ‘‘Guided-Wave Devices for Optical Communication,’’ IEEE Journal of Quantum
Electronics, Vol. QE-17, No. 6, Jun. 1981, pp. 946 – 955.

[9] Hinton, H. S., ‘‘Photonic Switching Using Directional Couplers,’’ IEEE Communications Magazine, Vol.
25, No. 5, May 1987, pp. 16 – 26.

[10] Izutsu, M., et al., ‘‘Lithium Niobate Devices and Their Role in Photonic Switching,’’ in Photonic
Switching II, pp. 318 – 328, K. Tada and H. S. Hinton (eds.), Berlin, Germany: Springer-Verlag, 1990.

[11] Schmidt, R. V., and R. C. Alferness, ‘‘Directional Coupler Switches, Modulators, and Filters Using
Alternating ∆β Techniques,’’ IEEE Transactions on Circuits and Systems, Vol. CAS-26, Dec. 1979,
pp. 1099 – 1108.

[12] Spanke, R. A., ‘‘Architectures for Guided-Wave Optical Space Switching System,’’ IEEE
Communications Magazine, Vol. 25, No. 5, May 1987. © 1987 IEEE. Portions reprinted with
permission.

[13] Lu, C. C., and R. A. Thompson, ‘‘The Double-Layer Network Architecture for Photonic Switching,’’
IEEE Journal of Lightwave Technology, Vol. 12, No. 8, Aug. 1994, pp. 1482 – 1489.

[14] Payne, W. A., and H. S. Hinton, ‘‘System Considerations for the Lithium Niobate Photonic Switching
Technology,’’ in Photonic Switching, pp. 196 – 199, T. K. Gustafson and P. W. Smith (eds.), Berlin,
Germany: Springer-Verlag, 1988.

[15] Padmanabhan, K., and A. N. Netravali, ‘‘Dialed Networks for Photonic Switching,’’ IEEE Transactions
on Communications, Vol. COM-35, No. 12, Dec. 1987, pp. 1357 – 1365. Also in Photonic Switching, pp.
142 – 145, T. K. Gustafson and P. W. Smith (eds.), Berlin, Germany: Springer-Verlag, 1988. © 1987
IEEE. Portions reprinted with permission.

[16] Thompson, R. A., et al., ‘‘An Experimental Modular Switching System with a Time-Multiplexed Photonic
Center-Stage,’’ Proc. Photonic Switching Topical Meeting, OSA, Salt Lake City, UT, Mar. 1 – 3, 1989,
pp. 92 – 93.

[17] Thompson, R. A., J. J. Horenkamp, and G. D. Bergland, ‘‘Photonic Switching of Universal Timeslots,’’
Proc. XIII Int. Switching Symposium, Stockholm, Sweden, May 27 – Jun. 1, 1990, Vol. 1, pp. 165 – 170.

[18] Evankow, J. D., Jr., and R. A. Thompson, ‘‘Photonic Switching Modules Designed with Laser Diode
Amplifiers,’’ IEEE Journal on Selected Areas in Communications, Vol. 6, No. 7, Aug. 1988, pp. 1087 –
1095. © 1988 IEEE. Portions reprinted with permission.

[19] Blumenthal, D. J., R. J. Feuerstein, and J. R. Sauer, ‘‘Multihop 2x2 All-Optical Photonic Packet Switch,’’
IEEE Conference on Communications, 1994, pp. 1364 – 1368. © 1994 IEEE. Portions reprinted with
permission.

SELECTED BIBLIOGRAPHY
Any good physics textbook should provide further background in optics. For much
greater detail, there are any number of good textbooks that specialize in optical
communications. A very readable textbook on optical communications is Fiber Optic
Communications, by Joseph C. Palais. Greater depth can be found in Optical Fiber
Telecommunications, edited by Stewart E. Miller and Alan G. Chynoweth, or Fiber Optic
Networks, by Paul E. Green, Jr.
722 Telephone Switching Systems

Many of the devices described in this chapter and the next are not recent enough, or
of enough general interest, to have found their way into textbooks. Yet, most of this work
on photonic switching devices represents group efforts and the life’s work of many very
talented individuals. Besides citing some individual papers below, it is more appropriate
to cite the researchers themselves. Anyone seeking further background on these photonic
devices should browse through the collected works of the following researchers, whom I
acknowledge for their contributions, both to this book and my own research: R. C.
Alferness, E. DeSurvire, R. M. Jopson, D. A. B. Miller, M. J. O’Mahony, P. Smith, and J.
Veselka.
The reader interested in photonic switching should read Scott Hinton’s book, An
Introduction to Photonic Switching Fabrics, which covers many of these topics in greater
detail. As in the previous paragraph, besides citing specific references, it seems more
useful to cite people: A. Hill, H. S. Hinton, A. Huang, D. K. Hunter, H. Jordan, I.
Kaminow, S. Knauer, C. T. Lea, C. C. Lu, G. Luderer, C. M. Qiao, S. Ramanan, G.
Richards, J. Sauer, R. A. Thompson, and L. Thylen.
Ailawadi, N. K., et al., ‘‘Broadband Photonic Switching Using Guided-Wave Fabrics,’’ IEEE LTS, May 1991,
pp. 38 – 43.
Alferness, R. C., ‘‘Waveguide Electrooptic Switch Arrays,’’ IEEE Journal on Selected Areas in
Communications, Vol. 6, No. 7, Aug. 1988, pp. 1117 – 1130.
Bell, T. E., ‘‘Fiber Optics,’’ IEEE Spectrum Magazine, (25th Anniversary Issue), Vol. 25, No. 11, 1988.
Bergmann, E. E., A. M. Odlzko, and S. H. Sangani, ‘‘Half Weight Block Codes for Optical Communications,’’
AT&T Technical Journal, Vol. 65, No. 3, May/Jun. 1986.
Bianchini, R. P., Jr., and H. S. Kim, ‘‘Design of a Nonblocking Shared-Memory Copy Network for ATM,’’
IEEE Infocom ’92, 1992, pp. 876 – 885.
Boncek, R. K., et al., ‘‘1.24416 Gbits/s Demonstration of a Transparent Optical ATM Packet Switch Node,’’
Electronics Letters, Vol. 30, No. 7, Mar. 31, 1994, pp. 579 – 580.
Chaffee, C. D., The Rewiring of America — The Fiber Optics Revolution, Boston, MA: Academic Press, 1988.
Cisneros, A., and C. A. Brackett, ‘‘A Large ATM Switch Based on Memory Switches and Optical Star
Couplers,’’ ICC ’91, 1991, pp. 721 – 728.
DeBosio, A., C. DeBernardi, and F. Milindo, ‘‘Deterministic and Statistic Circuit Assignment Architectures for
Optical Switching Systems,’’ Topical Meeting on Photonic Switching, Technical Digest Series, Vol. 13, OSA,
1987, pp. 35 – 37.
DeBosio, A., et al., ‘‘ATM Photonic Switching Node Architecture Based on Frequency Switching Techniques,’’
in Photonic Switching II, pp. 300 – 303, K. Tada and H. S. Hinton (eds.), Berlin, Germany: Springer-Verlag,
1990.
Elion, H. A., and V. N. Morozov, Optoelectronic Switching Systems in Telecommunications and Computers,
New York, NY: Marcel Dekker, 1984.
Eiselt, M., et al., ‘‘Experimental Optical ATM 2 x 2 Switching Node,’’ European Conference on Optical
Communications, 1992, pp. 373 – 376.
Eng, K. Y., ‘‘A Photonic Knockout Switch for High-Speed Packet Networks,’’ IEEE Journal on Selected Areas
in Communications, Vol. 6, No. 7, Aug. 1988, pp. 1107 – 1116.
Evankow, J. D., and R. M. Jopson, ‘‘Nondestructive Measurement of Length Dependence of Gain
Characteristics in Fiber Amplifiers,’’ IEEE Transactions Photonics Technology Letters, Vol. 3, No. 11, Nov.
1991.
Photonic Switching in Space 723

Fortenberry, R., et al., ‘‘Photonic Packet Switch Using Semiconductor Optical Amplifier Gates,’’ Electronics
Letters, Vol. 27, No. 14, Jul. 1991, pp. 1305 – 1307.
Fye, D. M., ‘‘Practical Limitations on Optical Amplifier Performance,’’ Journal of Lightwave Technology, Vol.
LT-2, No. 4, Aug. 1984.
Giacopelli, J. N., et al., ‘‘Sunshine: A High-Performance Self-Routing Broadband Packet Switch Architecture,’’
IEEE Journal on Selected Areas in Communications, Vol. 9, No. 8, Oct. 1991, pp. 1289 – 1298.
Green, P. E., Jr., Fiber Optic Networks, Englewood Cliffs, NJ: Prentice Hall, 1993.
Ha, W. L., et al., ‘‘Photonic Fast Packet Switching at 700 Mbps,’’ OFC ’91, 1991.
Haas, Z., ‘‘The Staggering Switch: An Electronically Controlled Optical Packet Switch,’’ IEEE Journal of
Lightwave Technology, Vol. 11, No. 5/6, May/Jun. 1993, pp. 925 – 937.
Hardy, S., ‘‘Ascend Enters Optical Networking Fray,’’ Lightwave, Aug. 1998.
Hinton, H. S., ‘‘Photonic Switching Technology Applications,’’ AT&T Technical Journal, Vol. 66, No. 3,
May/Jun. 1987, pp. 41 – 53.
Hinton, H. S., ‘‘Architectural Considerations for Photonic Switching Networks,’’ IEEE Journal on Selected
Areas in Communications, Vol. 6, No. 7, Aug. 1988, pp. 1209 – 1226.
Hinton, H. S., An Introduction to Photonic Switching Fabrics, New York, NY: Plenum Press, 1993.
Huang, A., and S. Knauer, ‘‘Starlite: A Wideband Digital Switch,’’ IEEE Globecom, 1984, pp. 121 – 125.
Ikeda, M., ‘‘Tandem Switching Characteristics for Laser Diode Optical Switches,’’ Electronics Letters, Vol. 21,
No. 6, Mar. 14, 1985.
Jacobs, I., ‘‘Basics of Lightwave,’’ Telephone Engineer & Management, Feb. 1985, pp. 74 – 80.
Jajszczyk, A., and H. T. Mouftah, ‘‘Photonic Fast Packet Switching,’’ IEEE Communications Magazine, Vol. 31,
No. 2, Feb. 1993, pp. 58 – 65.
Karol, M. J., M. G. Hluchj, and S. P. Morgan, ‘‘Input Versus Output Queueing on a Space-Division Packet
Switch,’’ IEEE Transactions on Communications, Vol. COM-35, No. 12, Dec. 1987, pp. 1347 – 1356.
Kasahara, K., ‘‘In Pursuit of Perfection,’’ Photonics Spectra, Vol. 32, No. 2, Feb. 1998.
Kobayashi, H., B. L. Mark, and Y. Osaki, ‘‘Call Blocking Probability of All-Optical Networks,’’ Proc. 1st IEEE
Int. Workshop on Broadband Switching Systems, Poznan, Poland, Apr. 19 – 21, 1995, pp. 186 – 200.
Kuroyanagi, S., T. Shimoe, and K. Murakami, ‘‘Photonic ATM Switching Network,’’ in Photonic Switching II,
pp. 296 – 299, K. Tada and H. S. Hinton (eds.), Berlin, Germany: Springer-Verlag, 1990.
Lambert, P., and F. Barbetta, ‘‘The Shape of Light,’’ tele.com, Vol. 3, No. 9, Aug. 1998.
Laude, J. P., K. Liddane, and S. Slutter, ‘‘Diffraction Gratings Stretch Fiber’s Capacity,’’ Photonics Spectra, Vol.
30, No. 2, Feb. 1996.
‘‘Lightwave Technology Market Forecast,’’ Lightwave, Dec. 1993, pp. 30 – 47.
Marrackchi, A., Photonic Switching and Interconnects, New York, NY: Marcel Dekker, 1994.
Marx, B. R., ‘‘Opto-Electronics Research Thrives in Europe,’’ Laser Focus World, Vol. 32, No. 12, Dec. 1996.
Miller, S. E., and A. G. Chynoweth (eds.), Optical Fiber Telecommunications, Orlando, FL: Academic Press,
1979.
Nakagawa, K., ‘‘Role of Optical Amplifiers in Realizing All-Optical Communication Networks,’’ in Photonic
Switching II, pp. 14 – 29, K. Tada and H. S. Hinton (eds.), Berlin, Germany: Springer-Verlag, 1990.
Palais, J. C., Fiber Optic Communications, Fourth Edition, Englewood Cliffs, NJ: Prentice Hall, 1998.
‘‘Photonics in the Information Age,’’ special issue of Photonics Spectra, Feb. 1995.
Renaud, M., et al., ‘‘Network and System Concepts for Optical Packet Switching,’’ IEEE Communications
Magazine, Vol. 35, No. 4, Apr. 1997, pp. 96 – 102.
724 Telephone Switching Systems

Sato, Y., K. Aida, and K. Nakagawa, ‘‘An Erbium-Doped Fiber Active Switch,’’ in Photonic Switching II,
pp. 122 – 125, K. Tada and H. S. Hinton (eds.), Berlin, Germany: Springer-Verlag, 1990.
Schwartz, M. I., ‘‘Optical Fiber Transmission — From Conception to Prominence in 20 Years,’’ IEEE
Communications Magazine, (issue on 100 Years of Communications Progress), Vol. 22, No. 5, May 1984.
Shibata, J., and T. Kajiwara, ‘‘Optoelectronic ICs,’’ IEEE Spectrum Magazine, (issue on Supercomputers: A
Special Report), Vol. 26, No. 2, Feb. 1989.
Shimazu, Y. and M. Tsukada, ‘‘Ultrafast Photonic ATM Switch with Optical Output Buffers,’’ IEEE Journal of
Lightwave Technology, Vol. 10, No. 2, Feb. 1992, pp. 265 – 272.
Simon, J. C., ‘‘Semiconductor Laser Amplifier for Single Mode Optical Fiber Communications,’’ Journal of
Optical Communications, Vol. 4, No. 2, 1983.
Suematsu, Y., and K. I. Iga, Introduction to Optical Fiber Communications, New York, NY: Wiley, 1982.
Thompson, R. A., ‘‘Optical Fiber Communications,’’ in Encyclopedia of Computer Science and Technology, Vol.
41, pp. 235 – 266, Kent, A., and J. G. Williams (eds.), New York, NY: Marcel Dekker, 1999.
Thompson, R. A., ‘‘Traffic Capabilities of Two Rearrangeably Nonblocking Photonic Switching Modules,’’
AT&T Technical Journal, Vol. 64, No. 10, Dec. 1985, pp. 2331 – 2373.
Tsuda, H., Y. Sakai, and T. Kurokawa, ‘‘Optical Flip-Flops Using Two Nonlinear Etalons,’’ in Photonic
Switching II, pp. 118 – 121, K. Tada and H. S. Hinton (eds.), Berlin, Germany: Springer-Verlag, 1990.
Wakao, K., et al., ‘‘InGaAsP/InP Optical Switches Embedded with Semi-Insulating InP Current Blocking
Layers,’’ IEEE Journal on Selected Areas in Communications, Vol. 6, No. 7, Aug. 1988, pp. 1199 – 1204.
Yasui, T., and K. Kikuchi, ‘‘Photonic Switching System/Network Architectural Possibilities,’’ in Photonic
Switching, pp. 24 – 35, T. K. Gustafson and P. W. Smith (eds.), Berlin, Germany: Springer-Verlag, 1988.
Yeh, Y., M. G. Hluchyj, and A. S. Acampora, ‘‘The Knockout Switch: A Simple, Modular Architecture for
High-Performance Packet Switching,’’ IEEE Journal on Selected Areas in Communications, Vol. SAC-5, No. 8,
Oct. 1987, pp. 1274 – 1283.
Zhong, W. D., Y. Onozato and J. Kaniyil, ‘‘Copy Network with Shared Buffers for Large-Scale Multicast ATM
Switching,’’ IEEE/ACM Transactions on Networking, Vol. 1, No. 2, Apr. 1993, pp. 157 – 165.

Das könnte Ihnen auch gefallen