Optical Switching and Networking xxx (xxxx) 100621

Contents lists available at ScienceDirect

Optical Switching and Networking

journal homepage: www.elsevier.com/locate/osn

Review

A Survey of Reconfigurable Optical Networks

Matthew Nance Hall a, ∗ , Klaus-Tycho Foerster b , Stefan Schmid b , Ramakrishnan Durairajan a
a Computer and Information Science, University of Oregon, 1447 E. 13th Ave, Eugene, OR, 97403-1202, USA
b Fa culty of Co mput er Sc ience, University of Vienna , Waeh ringer St r. 29, 1090, Vienna , Austria

ARTICLE INFO ABSTRACT

Keywords: Reconfigurable optical networks have em erged as a promising tech nology to efficiently serve the fast-
Reconfigurab le optical network s growing traffic produ ced by the digital society. This paper provides a survey of the field. We first review en-
Data center network s abling optical hardware tech nologies in general and then consider tech nologies that are specific to data cen-
Wide ar ea network s
ter networks and wide-area networks in more detail. We further provide an overview of the cost models
Network ed Sy stems
used in the literature as well as the algorithmic problems introdu ced by these tech nologies, their first solu-
Algo ri thms
tions, and discu ss systems and impleme ntation aspects. We conclud e with a discu ssion of open ch allenges.

1. Introduction However, with reconfigurable optical networks being a relatively a

new technology, the community is still discussing their conceptual fun-
The popularity of data-centric applications related to business, sci- damentals, benefits, and limitations.
ence, social networking or entertainment, as well as the rise of machine
learning and artificial intelligence, leads to an explosive growth of 1.1. Novelty and contribution
communication traf fic, especially inside data centers but also in the
wide-area networks (WANs) connecting the data centers to users and To the best of our knowledge, this is the first and up-to-date survey
each other. Optical communication is one of the most promising tech- on emerging reconfigurable optical networks, considering both data
nologies to cope with the resulting increase in communication require- center and WAN settings. While there are surveys with broad overviews
ments, due to the high bandwidth and energy efficiency it provides. Our of reconfigurable optical technologies in these settings (e.g., softwa re-
society's communication infrastructure is hence likely to become more defined optical networks [1], routing and spectrum allocation [2],
and more optical soon. Accordingly, we currently witness many efforts wa velength switching hardwa re architecture [3]), the functioning of
to further advance optical interconnects and lightwa ve telecommuni- such systems (e.g., reconfigurable metropolitan networks [4] and data
cation in data centers and WANs. center networks [5]) requires a full-stack perspective on optical net-
Reconfigurable optical technologies are a particularly interesting works. We address this requirement, in this paper, by presenting an end-
innovation in this context. Reconfigurable optical networks enable to-end perspective on reconfigurable optical networks by (a) emphasiz-
adaptation: either of the topology itself (in case of data centers) or of ing the interdependence of optical technologies with algorithms and
the network capacity (in case of WANs). Such adaptations may be ex- systems and (b) identifying the open challenges and future work at the
ploited by next-generation systems to improve performance and effi- intersection of optics, theory, algorithms, and systems communities.
ciency, e.g., by making the network demand aw are. For example, re- This survey is timely, as interest in dynamic optical layer network-
cent technologies based on free-space optics or optical circuit switches ing technologies is gaining attention from the networked systems com-
(OCSs) support very fa st topology adaptations in data centers. Mean- munity. Upon reviewing the last five years of publications from five op-
while, technologies based on reconfigurable optical add-drop multi- tical and systems network journals, we ran a clustering analysis to see
plexers (ROADM s) can add or drop wa velengths carrying data chan- how much overlap there has been between the two fields. Ta ble 1
nels from a transport fiber without the need to convert the signals to shows the journals, and their raw publication counts since 2015. The
electronic signals and back. In both cases, the entire bandwidth assign- clustering analysis wa s conducted using CitNetExplorer softwa re [6],
ment's planning need not be carried out during the initial deployment of and used a clustering resolution of 0.75. Connections between papers
a network.

∗ Correspond ing author.

E-mail address: mh [email protected] (M . Nanc e Hall).

https://doi.org/10.1016/j.osn.2021.100621
Received 29 September 2020; Received in revised form 12 February 2021; Accepted 18 March 2021
1573-4277/© 2021
M. Nance Hall et al. Optical Switching and Networking xxx (xxxx) 100621

Table 1 1.2. Scope

Papers pu blished sinc e 20 15 in various networking journals.
Journa l Papers There already exist some excellent surveys on optical networks
which at least partially cover reconfigurable aspects, both in the con-
Journa l of Li ghtwav e Technolo gy (JLT) 3988
text of data centers [7–9] and (to a lesser extent) in the context of
IEEE Tran sa ctions on Co mm unications (TCO M) 2703
WANs [10]. We extend these surveys while providing an up-to-date
IEEE-AC M Tran sa ctions on Networ ki ng (TON) 1226
Journa l of Op tical Co mm unications (JOC N) 817 overview of the literature. Our paper aims to provide an understanding
IEEE Tran sa ctions on Networ k an d Serv ice Ma na gement (TNSM) 533 of the underlying fundamental issues and concepts and identify com-
monalities and differences in different reconfigurable optical networks
are first-order citations, and the publication records are from Web of (spanning both data center and WANs). To this end, we proceed from
Science [182]. practical technological constraints to theoretical models and solutions
Fig. 1 shows the results of the clustering analysis in the five largest back to implementations. Within this domain, we restrict our coverage
clusters. Ba sed on the largest cluster, 1, it appears there is a strong rela- to enterprise and core networks. For those interested in last-mile passive
tionship between the Journal of Lightwa ve Technology (JLT) and the optical networks, mobile front-haul, and multiple-access networks we
Journal of Optical Communications and Networking (JOCN) publica- defer to the following related surveys [11–13].
tions. This is no surprise, as the two journals have a strong emphasis on
optical networking. Cluster 2 is mainly composed of IEEE Transactions 1.3. Organization
on Communications (TCOM) papers, with a few IEEE/ACM Transac-
tions on Networking (TON) papers. TCOM papers are generally con- The remainder of this paper is organized as follows. Section 2 de-
cerned with physical layer networking, including radio signals and fines the network architecture model for optical networks and its con-
copper mediums in addition to optical transmission. Cluster 3 shows a nection to the packet switched network model. Section 3 introduces the
strong relationship between TON and IEEE Transactions on Network key hardwa re technologies that underpin reconfigurable optical net-
and Service Ma nagement (TNSM) papers. These journals have a re- works. Section 4 showcases research on reconfigurable optical data
lated interest in networked systems, applications, and management. center networks (DCNs) by highlighting DCN specific hardwa re capa-
The three predominantly physical-layer journals (JLT, TCOM, and bilities, cost modeling, algorithms, and systems implementations. Sec-
JOCN) appear together in cluster 4. Cluster 5 shows a relation between tion 5 looks directly at reconfigurable WANs with an emphasis on
physical-layer and systems journals, TCOM, and TON. Clusters 6 and WAN-specific challenges in addition to cost modeling, algorithms, and
beyond are mostly singleton clusters, comprising predominantly one systems implementations. Section 7 concludes with the overarching
journal and are therefore omitted from the figure. Our analysis under- open challenges for the field of reconfigurable optical networks span-
scores a division between optical and higher layer publishing venues. ning hardwa re, data centers, and WANs.
The missing piece in this picture appears to be any significant over-
lap between the networked systems journals and optical journals. We 2. Network architectures
hope this survey will illuminate common areas of interest between these
two communities in order to bring more attention to cross-layer net- In this section, we briefly discuss two network architecture models
working research, as cross-domain knowledge and expertise will be in- that can leverage reconfigurable optics, IP-over-OTN networks and
creasingly important for bringing greater flexibility and control to the hybrid electric-optical data center networks. Our focus in this survey is
management of optical networks. to highlight and categorize reconfigurable optical networks in enter-
Our survey is tutorial in nature and focuses on concepts rather than prise networks, and therefore leave last-mile optical networks, such as
exhaustive related work, concentrating on selected articles. Hence, our passive-optical networks and fiber-to-the-home networks beyond the
paper targets students, researchers, experts, and decision-makers in the scope of our discussion.
networking industry who would like to obtain an overview of the criti- We also briefly outline principles leveraged in different contexts by
cal concepts and state-of-the art results in reconfigurable optical net- reconfigurable optical networks, softwa re defined networking and
works. We start with an overview of the enabling optical hardwa re elastic optical networking. This discussion introduces key aspects for
technologies. We explore where data center and WAN systems have in- network designers to consider when building a reconfigurable optical
tegrated this hardwa re. We review cost models, discuss the novel algo- network. This discussion reinforces our illustration of how full-stack
rithmic challenges and solutions in the literature, and elaborate on sys- perspective aids in the network design process. In sections 4 and 5, we
tems and implementation aspects. We also identify the major open is- look at specific implementations of reconfigurable optical networks in
sues which require further exploration and research to design the next more detail.
generation reconfigurable optical networks.
2.1. IP-over-optical transport network

IP-over-Optical Transport Networks (IP-over-OTN), defined in

ITU-T G.709, is the standardized protocol that links metro, regional,
and long-haul networks, as illustrated in Fig. 2. Thus, we discuss IP-
over-OTN when referring to the network's IP and the optical layers. In
IP-over-OTN. hosts (e.g., data centers, points-of-presence or POPs,
servers, etc.) connect to routers, and these routers are connected
through the optical transport network (OTN). A node in the optical
layer is an Optical Cross-Connect (OXC). An OXC transmits data on
modulated light through the optical fiber. The modulated light is called
a lambda, wa velength, or circuit. The OXC can also act as a relay for
other OXC nodes to transparently route wa velengths. When acting as a
relay for remote hosts, an OXC provides optical switching capabilities,
thus giving the network flexibility in choosing where to send transmit-
ting lambdas over the OXC node.
Fig. 1. Paper clusters among five networking and optical systems journals.

2
M. Nance Hall et al. Optical Switching and Networking xxx (xxxx) 100621

networks for node placement, fiber placement, and wa velength alloca-

tion.

2.2. Data center architecture

Historically, data centers relied on packet-switched networks to

connect their servers; however, as scale and demand increased, the cost
to build and manage these packet-switched networks became too large.
As a result of this change, new reconfigurable network topologies
gained more attention from researchers and large cloud providers.
Many novel data center architectures with reconfigurable optical
topologies have been proposed over the last decade. These architec-
Fig. 2. Metro, regional, and long-haul networks are connected by the IP-over- tures have in common that they reduce the static network provisioning
OTN standard. requirements, thereby reducing the network's cost by presenting a
means for bandwidth between hosts to change periodically. Fig. 4
Fig. 3 illustrates the connectivity at different layers of the IP- shows one such example of a hybrid electrical-optical data center archi-
over-OTN model. The physical network connects points-of-presence tecture. These architectures reduce cost and complexity via scheduling
(POPs) with optical fiber spans. OXC nodes connects these POPs with methods, which change bandwidth on optical paths in the data center.
optical paths or circuits. The physical routes of the paths are ab- Various approaches have been demonstrated. Notable architectures
stracted aw ay, and shown in color for reference. In the IP topology, employ fixed, and deterministic scheduling approaches [5,15] or de-
the colors of light are also abstraced aw ay, and we see a mesh IP net- mand-aware changes that prioritize establishing optical paths between
work connected by routers and switches. Hosts connect to nodes at servers with mutual connectivity requests [16,17]. Switching fabrics
this layer, and their traf fic travels down the optical paths in the are also diverse for data center optical systems. These include fabrics
physical network to reach its destination. based on nanosecond tunable lasers [18], digital micromirror devices
IP-over-OTN networks are not new. However, they are built at a (DMD) [19], and liquid crystal on silicon (LCOS) wavelength selective
great cost. Historically network planners have engineered them to ac- switches (WSS) [20].
commodate the worst-case expected demand by (1) over-provisioning
of dense wa velength division multiplexing (DWDM) optical channels 2.3. Software defined networking
and (2) laying redundant fiber spans as a fa il-safe for unexpected traf -
fic surges. These surges could come from user behavior changes or fa il- Modern data center, metro, and wide-area networks have been sub-
ures elsewhere in the network that forces traf fic onto a given path. stantially influenced by developments in Softwa re Defined Networking
Only recently have reconfigurable optical systems begun to gain atten- (SDN) [22], and this trend has also been making its wa y to optical net-
tion in the data center and wide-area network settings. For more infor- works [1]. The SDN paradigm decouples the control and data plane in
mation about early IP-over-OTN, we defer to Ba nnister et al. [14] and network hardwa re, giving operators greater control and flexibility for
references therein, where the authors present work on optimizing WDM controlling traf fic within their network. Without this decoupling, it is
more difficult to make lock-step changes to network functions, such as
routing. SDN offers a logically centralized point of control for imple-
menting policies across the network, thus enabling better network uti-
lization for bandwidth, latency, security policies, etc. These concepts
can also map further down the network stack to manage optical infra-
structure, thereby 1) improving optical layer performance with tech-
nology, which we describe in Section 3, and 2) allowing management
algorithms to adapt the optical paths in a demand-aw are fa shion,
which we describe in Section 4 for data center networks and in Section
5 for metro and wide-area networks.
Notwithstanding, providing a standardized stable and reliable pro-
grammable optical physical layer control plane for SDNs is still an on-
going effort, as recently outlined by the TURBO project [23]. One im-
portant step in this direction is the development of virtual testbeds to
evaluate the cross-layer operation of SDN control planes [24].

Fig. 3. IP-over-OTN network architecture model, showing the connection be-

tween IP and optical layers. Fig. 4. Data center architecture proposed in c-Through [21].

3
M. Nance Hall et al. Optical Switching and Networking xxx (xxxx) 100621

2.4. Elastic optical networks 3.1. Wavelength selective switching

A span of optical fiber enables transmission of data over a spectrum In contrast to packet-switched networks, optically circuit-switched
or set of wa velengths. These wa velengths can be allocated in a fixed or systems operate at a coarser granularity. The transmission of informa-
flexible (flex) grid. Networks that allow flex grid allocations are also tion over a circuit requires an end-to-end path for the communicating
called Elastic Optical Networks (EONs). For example, according to the parties. Although packet switching has generally prevailed in today's
ITU-T G.694.1 fixed grid standard, frequencies must be 12.5, 25, 50, or Internet, recent research has revitalized the prospect of circuit switch-
100 GHz apart [1]. However, in elastic optical networks (EONs), also ing for data centers and wide-area networks by illuminating areas in
known as flex-grid networks, the frequency of a channel can be any which flexible bandwidth benefits outweigh the start-up cost of circuit
multiple of 6.25 GHz aw ay from the central frequency (193.1 THz) building.
and have a width that is a multiple of 12.5 GHz. Fig. 5 illustrates the Technological advancements for optical hardwa re, primarily driven
difference between a flex-grid and fixed-grid allocation. by physics and electrical engineering research, have been instrumental
Flex grid networks can greatly improve the spectral efficiency of IP- in making circuit-switched networks a viable model for data center
over-OTN, allowing the network to pack data channels more densely networks. Among these technologies are low-cost/low-loss hardwa re
within a span of optical fiber. However, they can also lead to unique architectures. Here we give a brief overview of technological advance-
challenges, particularly fragmentation. Fragmentation occurs when ments in this domain that have had the most significant impact on net-
spectrum allocated on a fiber has gaps in it that are too narrow to be worked systems.
filled. Novel approaches to managing EONs with fragmentation-aw are Kachris et al. [26] have an in-depth look at optical switching archi-
algorithms are covered in depth by Chatterjee et al. [25]. tectures in data centers from 2012. In their survey, they primarily look
at competing data center architectures and switch models. In this sec-
2.5. Summ ary tion, we choose to focus instead on those architectures’ physical mani-
festations (i.e., the base components that make them up). Furthermore,
Our survey relates the latest developments in reconfigurable net- exciting new developments have occurred since then, which we high-
works for data centers and WANs. The IP-over-OTN model is a useful light in this section.
framework for reasoning about and managing optical metropolitan, Polymer waveguides are a low-cost architecture for optical cir-
regional, and wide-area networks. Similarly, we are seeing data center cuit switches. These have been fa bricated and studied in depth over the
architectures become more reconfigurable and demand-aw are with op- last 20 years, including work by Ta boada et al. [27] in 1999, Yeniay
tical circuit switching. SDN is poised to bring substantial changes to the et al. [28] in 2004, and Felipe et al. [29] in 2018. Early implementa-
operation of optical networks in both domains by offering a centralized tions such as Ta boada et al. [27] showed fa brication techniques for
point for management and control for more network infrastructure, simple polymer wa veguide taps. Multiple wa veguide taps can be com-
from routing of packets to routing of optical paths. Moreover, EONs bined to form an Array Waveguide Grating (AWG), and the signals tra-
are also enabling better spectral efficiency. versing the AWGs can then be blocked or unblocked to create an optical
circuit switch. A major inhibitor of the polymer wa veguide architecture
3. Enabling hardware technologies wa s signal-loss, which wa s as high as 0.2 dB/cm until Yeniay et al.
[28] discovered an improvement on the state-of-the-art with ultralow-
In this section, we discuss hardwa re technologies that enable recon- loss wa veguides in 2004. Their wa veguides, made with fluorocarbons,
figurable optical networks. In our end-to-end discussion on reconfig- have 4× less loss (0.05 dB/cm) than the next best wa veguides at the
urable optical networks, the hardwa re is the foundational layer from time, made from hydrocarbons. Felipe et al. [29] demonstrate the ef-
which systems are built. Understanding these devices and their capabil- fectiveness of a polymer wa veguide-based switching architecture for
ities is crucial for designing and building real-world reconfigurable op- reconfiguring groups of optical flows of up to 1 Tbps, proving that
tical networking systems. We show examples of different optical tech- AWG is a viable and competitive switching architecture for data cen-
nologies, including optical switches and transponders, and examples of ters. More recently, in 2020, AWGs were demonstrated to work in con-
systems that use them. We also highlight recent advances in silicon pho- junction with sub-nanosecond tunable transmitters to create flat
tonics, and the implications this may have for reconfigurable optical topologies, significantly reducing power consumption for data center
networks in the near future. Finally, we discuss open challenges in re- networks due to the passive—no power required—nature AWGs [30].
configurable optical networks that might be solved with next- Switching speeds below 820 ps have been demonstrated using a 1 × 60
generation hardwa re. AWG and tunable laser [18]. AWGs with as many as 512 ports have
been demonstrated [31].
Microelectromechanical Systems (MEMS), introduced by
Toshiyoshi et al. [32] in 1996, offered a lower-loss and more flexible

Fig. 5. Example of fixed grid and flex grid spectrum allocation.

4
M. Nance Hall et al. Optical Switching and Networking xxx (xxxx) 100621

alternative to polymer wa veguide systems of the day. MEMS devices Table 2

are made up of small mirrors, which can be triggered between states Summary of systems implementations of reconfigurable wide area networks.
(i.e., on and of f). Therefore, in a MEMS system light is reflected rather Port Switching
than guided (as in the polymer wa veguide systems). This distinction Scalability Speed
between reflection and guiding implies generally slower switching
AWG 512 × 512 < 820 ps Highly scalable with unsurpassed demonstrations
speeds for MEMS based systems, as the mirror must be physically for short-reach applications with tunable lasers.
turned to steer light out of the desired switch-port. Despite this limita- MEMS 32 × 32 < 0.5 ms Higher scalability and lower insertion loss, less
tion, MEMS systems evolved to be competitive with polymer wa ve- crosstalk.
guides in modern systems. Advances in MEMS technology have yielded LCOS 16 × 16 10 − 100 Lower scalability and optical performance, but
μs more modular design than MEMS.
wa velength selective switches (WSS) scalable to 32 ports with switch-
ing speeds under 0.5 ms [33]. Da ta center solutions leveraging MEMS
based switches include Helios [34]. 3.2. ROADMs
Liquid Crystal on Silicon (LCOS) wa s demonstrated as another
viable optical switching architecture by Ba xter et al. [35] in 2006. An Reconfigurable add-drop multiplexers, or ROADMs, are an integral
LCOS switch is depicted in Fig. 6. Multiplexed optical signals enter the component of IP-over-OTN networks. These devices have evolved over
system from a fiber array. These signals are directed to a conventional the years to provide greater functionality and flexibility to optical
diffraction grating where the different colors of light are spatially sep- transport network operators. We briefly describe the evolution of
arated from each signal. These colors are then projected onto a unique ROADM architectures. Fig. 7 shows a broadcast and select ROADM ar-
position in the LCOS switching element. This element is divided into chitecture. Please refer to Ref. [3] for more information about ROADM
pixels or cells, and charged with an electrical current. The voltage ap- architectures.
plied to any cell in the switching element determines which output Colorless (C). Early ROADMs were effectively programmable
fiber a given channel will leave through. From there, the signal travels wa velength splitter-and-blockers, or broadcast-and-select devices. A
back through the system and into a different fiber in the array. wa velength splitter-and blocker can be placed before an IP-layer
Switches based on this technology have a response time of 10–100 switch. If the switch is intended to add/drop a wa velength (i.e., trans-
μs [36]. Recent work by Yang et al. [37] demonstrates the construction ceive data on it), then the blocker prohibits light on the upstream path
of a 12 × 12 and 1 × 144 port WSS based on a 1 × 12 LCOS architec- and enables light on the path to the switch. These splitter-and-blocker
ture. Chen et al. [38] developed an improved LCOS architecture with systems are better known as Colorless, or C-ROADMs, as the splitter-
which they demonstrated a 16 × 16 optical switch. LCOS switches are and-blocker architecture is independent of any specific frequency of
commercially available and are recognized as a key enabler for recon- light. To receive the maximum benefit from C-ROADMs, operators
figurable optical networks [20]. should deploy their networks with tunable transceivers as they allow
Summary. Table 2 summarizes optical switch performance met- more flexibility for the end hosts when connecting to remote hosts.
rics. Each architecture comes with advantages under distinct circum- Colorless, Directionless (CD). The CD-ROADMs extend the ar-
stances. Highly scalable data center architectures have been developed chitecture of C-ROADMs by pairing multiple C-ROADMs together in
with sub-nanosecond tunable lasers and AWGs [5,18,30]. MEMS have the same unit to allow for a wa ve to travel in one of many directions.
generally better scalability, lower insertion loss, and less crosstalk One shortfa ll of this architecture is that the drop ports from each direc-
over LCOS systems [39] but also demand higher precision manufactur- tion are fixed, and therefore if all of the drop ports are used from one
ing to ensure that all N × M mirrors configurations are accurately direction, the remaining points from other directions cannot be used.
aligned. LCOS elements can also be packed more compactly into a Due to the limitation of drop ports in different directions, the CD archi-
modular unit due to the absence of moving parts that are present in tecture is not contentionless.
MEMS. Colorless, Directionless, Contentionless (CDC). The CDC-
Summary. Table 2 summarizes optical switch performance met- ROADM solves the contention problem by providing a shared add/
rics. Each architecture comes with advantages under distinct circum- drop port for each direction of the ROADM. This allows contention-
stances. Highly scalable data center architectures have been developed less reconfiguration of the ROADM as any drop-signal is routed to a
with sub-nanosecond tunable lasers and AWGs [5,18,30]. MEMS have common port regardless of the direction from which the wa ve begins/
generally better scalability, lower insertion loss, and less crosstalk terminates.
over LCOS systems [39] but also demand higher precision manufactur- Colorless, Directionless, Contentionless w. Flexible Grid
ing to ensure that all N × M mirrors configurations are accurately (CDC-F). Flexi-grid, or elastic optical networks, are networks carrying
aligned. LCOS elements can also be packed more compactly into a optical channels with non-uniform grid alignment. This contrasts with
modular unit due to the absence of moving parts that are present in a fixed-grid network, where different wa velengths are spaced with a
MEMS. fixed distance (e.g., 50 GHz spacing). Wideband spacing allows signals
to travel fa rther before becoming incoherent due to chromatic disper-
sion. Thus, CDC-Flex or CDC-F ROADMs enable the reconfiguration of

Fig. 7. Broadcast and Select colorful ROADM. The add/drop node, R1, has
ports for two optical ch annels. These ch annels are directed at the ROADM.
The ROADM uses a splitter to broadcast the ch annels onto two outboun d
ports, wh ere a wavelength blocker selects the appropriate ch annel for the
Fig. 6. Liquid crystal on silicon wavelength selective switch. next router.

5
M. Nance Hall et al. Optical Switching and Networking xxx (xxxx) 100621

wa velengths with heterogeneous grid alignments. These are most use- gram for 16-QAM modulation, which offers 4 bits per symbol, or twice
ful for wide area networks, with combinations of sub-sea and terres- the baud of QPSK.
trial circuits. Fiber optic communications are subject to noise. The noise level is
Signal to Noise Ratio (SNR), and this metric determines the highest pos-
3.3. Bandwidth-variable transponders sible modulation format. In turn, the modulation format yields a poten-
tial capacity (Gbps) for an optical channel. For example, in Ref. [40],
Before we discuss bandwidth-variable transponders, we must first the authors claim that SNR of just 6 dB is sufficient to carry a 100 Gbps
take a moment to illuminate a common concept to all physical commu- signal, while a circuit with an SNR of 13 dB can transmit 200 Gbps.
nications systems, not only optical fiber. This concept is modulation Ba ndwidth Variable Transponders (BVTs) [41] have recently
formats. Modulation formats determine the number of binary bits that a proven to have significant applications for wide-area networks. These
signal carries in one symbol. Two parties, a sender and receiver, agree devices are programmable, allowing for the operator to choose from
on a symbol rate (baud), which determines a clock-speed to which the two or more different modulation formats, baud rates, and the number
receiver is tuned when it interprets a symbol from the sender. The sim- of subcarriers when operating an optical circuit. For example, the same
plest modulation format is on-off keying (OOK), which transmits one transponder may be used for high-capacity/short-reach transmission
bit per symbol. In OOK, the symbol is sent via a high or low power level, (16-QAM or greater) or lower-capacity/longer-reach transmission
as shown in Fig. 8A. A higher-order modulation technique is Quadra- (e.g., QPSK). Higher modulation formats offer higher data rates. They
ture Phase Shift Keying (QPSK), in which the symbol is a sinusoidal are also more sensitive to the optical SNR, which decreases in a step-
wave whose phase-offset relates the symbol. In QPSK, there are four wise manner with distance, as illustrated in Fig. 9. We note that BVTs
phase shifts agreed upon by the communicating parties, and therefore enable network operators to meet the ever-growing demand in back-
the system achieves two bits per symbol, or two baud, seen in Fig. 8B. A bone traf fic by increasing optical circuits’ spectral efficiency.
constellation diagram for QPSK is shown in Fig. 8C. As modulations be- Low spectrum utilization, or wa ste, can be an issue for BVT circuits.
come more complex, it is more useful to visualize them in the phase For example, a BVT configured for a low-modulation circuit such as
plane shown by their constellation diagram. Higher order modulation QPSK instead of 16-QAM has a potential for untapped bandwidth.
formats are of the type, N-Quadrature Amplitude Modulation (QAM) Sambo et al. [42] introduced an improvement to the BVT architecture,
techniques (Fig. 8D), and these permit log2(N) bits per symbol, where N known as Sliceable-BVT (S-BVT), which addresses this issue. They de-
is generally a power of 2. In QAM, the symbol is denoted by phase and scribe an architecture that allows a transponder to propagate numer-
amplitude changes. Fig. 8D shows an example of a constellation dia- ous BVT channels simultaneously. Channels in the S-BVT architecture
gram for 16-QAM modulation, which offers 4 bits per symbol, or twice are sliceable in that they can adapt to offer higher or lower modulation
the baud of QPSK. in any number of the given subchannels.
Before we discuss bandwidth-variable transponders, we must first
take a moment to illuminate a common concept to all physical commu- 3.4. Silicon photonics
nications systems, not only optical fiber. This concept is modulation
formats. Modulation formats determine the number of binary bits that a Various materials (e.g., Ga As, Si, SiGe) can be used to make photon-
signal carries in one symbol. Two parties, a sender and receiver, agree ics hardwa re required for data transmission. These devices include pho-
on a symbol rate (baud), which determines a clock-speed to which the todetectors, modulators, amplifiers, wa veguides, and others. Silicon
receiver is tuned when it interprets a symbol from the sender. The sim- (Si) is the preferred material for these devices due to its low cost. How-
plest modulation format is on-off keying (OOK), which transmits one ever, there are challenges to manufa cturing these silicon devices, such
bit per symbol. In OOK, the symbol is sent via a high or low power level, as optical power loss and free carrier absorption. Other materials, no-
as shown in Fig. 8A. A higher-order modulation technique is Quadra- tably Ga As, have better properties for propagating light; however,
ture Phase Shift Keying (QPSK), in which the symbol is a sinusoidal Ga As is more costly to manufa cture. Despite these challenges, research
wave whose phase-offset relates the symbol. In QPSK, there are four into efficient and quality transmission using silicon-based photonic de-
phase shifts agreed upon by the communicating parties, and therefore vices has boomed in the last decade. Early advances were made to-
the system achieves two bits per symbol, or two baud, seen in Fig. 8B. A wa rds silicon photonics (SiP) in the 80s, particularly for wa veguides,
constellation diagram for QPSK is shown in Fig. 8C. As modulations be- which are the basis for circuit switches and multiplexers. Today, SiP is
come more complex, it is more useful to visualize them in the phase an integral part of almost all optical hardwa re, including lasers, modu-
plane shown by their constellation diagram. Higher order modulation lators, and amplifiers.
formats are of the type, N-Quadrature Amplitude Modulation (QAM) A significant challenge for power-efficient SiP transceivers is cou-
techniques (Fig. 8D), and these permit log2(N) bits per symbol, where N pling loss between the laser source and passive wa veguide on Si inte-
is generally a power of 2. In QAM, the symbol is denoted by phase and grated circuit wa veguides, which can be as high as 2.3 dB, or 25%
amplitude changes. Fig. 8D shows an example of a constellation dia- power loss [43]. Recent work by Billah et al. [44] explores the integra-
tion of indium pihosphide (InP) lasers on chips, demonstrating a cou-

Fig. 8. Modulation examples of on-off keying, quadrature phase shift keying (QPSK), quadrature amplitude modulation (QAM), and constellation diagrams for QPSK
and 16QAM.

6
M. Nance Hall et al. Optical Switching and Networking xxx (xxxx) 100621

into technology, cost modeling, and algorithms. In technology, we sup-

plement the discussion from Section 3 with hardware capabilities that
currently exist only for DCNs. Such features include free-space optics
and sub-second switching. Next, we highlight cost modeling research,
whose goal is to derive formal estimates or guarantees on the benefit of
reconfigurable optical networks over static topologies for DCNs. Fi-
nally, we survey the relevant algorithms for managing and optimizing
reconfigurable optical networks in the data center. Many of these algo-
rithms focus on the interdependencies between optical path set-up and
routing and optimize them across layers. Notwithstanding, there is also
work that optimizes the physical layer simultaneously as well, respec-
tively focuses on the interplay between software defined networking
(SDN) and the physical layer, as illustrated in Fig. 10. We discuss these
Fig. 9. Conceptualization of the trade-off between modulation/data rate and examples in more detail and also survey further related work across the
distance/noise with BVT. Noise, which can be measured with bit error rate, Q next subsections.
factor, or SNR, increases with the distance covered by an optical circuit. As In this section, we illuminate efforts to improve DCNs with reconfig-
more noise is accumulated over greater distance, the highest-order modulation urable optics. Related surveys on this subject include Foerster et al. [7]
that the circuit can support, and thereby the data rate on that circuit, falls in a
and Lu et al. [9]. We divide the state of reconfigurable optical DCNs
piece-wise manner.
into technology, cost modeling, and algorithms. In technology, we sup-
plement the discussion from Section 3 with hardware capabilities that
pling with only 0.4 dB of loss, or roughly 10%. InP appears to be a
currently exist only for DCNs. Such features include free-space optics
promising compound for other SiP technology too, as evident by
and sub-second switching. Next, we highlight cost modeling research,
demonstrations of InP in-line amplification for WSS [45]. Costs are
whose goal is to derive formal estimates or guarantees on the benefit of
fa lling for optical hardwa re as more efficient and scalable manufa ctur-
reconfigurable optical networks over static topologies for DCNs. Fi-
ing techniques are enabled by SiP [46], thus allowing network opera-
nally, we survey the relevant algorithms for managing and optimizing
tors to deploy newer technology into their systems at a more advanced
reconfigurable optical networks in the data center. Many of these algo-
pace as the devices’ quality and guarantees have continued to improve.
rithms focus on the interdependencies between optical path set-up and
For more information on silicon photonics, see the survey by Thomson
routing and optimize them across layers. Notwithstanding, there is also
et al. [47].
work that optimizes the physical layer simultaneously as well, respec-
tively focuses on the interplay between software defined networking
3.5. Summ ary
(SDN) and the physical layer, as illustrated in Fig. 10. We discuss these
examples in more detail and also survey further related work across the
Hardwa re for reconfigurable optical networks is improving at rapid
next subsections.
scales, where researchers are developing more scalable optical switches
A key challenge for data centers is to optimize the utilization of the
with fa ster response times year af ter year. These WSS architectures are
data center network (DCN). In a DCN, many different services are run-
quickly being integrated with ROADMs to offer CDC-F flexibility for
ning and competing for shared bandwidth. Communication patterns
networks. Meanwhile, improvements to transponder technology are
between top-of-rack (ToR) switches vary with the underlying applica-
also paving the wa y for reconfigurable optics at network endpoints. In
tions that are running (e.g., map-reduce, video stream processing,
particular, S-BVTs offer dramatic CAPEX savings as one transponder
physics simulations, etc.). Thus, as future applications and user's needs
can deliver multiple modulated signals in parallel. These improvements
change, it is challenging to predict where bandwidth will be needed.
are accelerated by silicon photonics, bringing CMOS manufa cturing to
Static and reconfigurable network solutions have been posed by re-
optical hardwa re and greatly reducing the cost to deploy optical
search and industry to address this challenge. There is an assumption
switches and upgraded transponders in networks.
that the connectivity graph of the network cannot change in static net-
work solutions. These solutions also assume fixed capacity (or band-
4. Optically reconfigurable data centers
width) on links. In reconfigurable network solutions, by contrast, these
assumptions regarding connectivity and bandwidth are relaxed.
In this section, we illuminate efforts to improve DCNs with reconfig-
Servers and switches (collectively referred to as nodes) may connect
urable optics. Related surveys on this subject include Foerster et al. [7]
some subset of the other nodes in the network, and the nodes to which
and Lu et al. [9]. We divide the state of reconfigurable optical DCNs

Fig. 10. Solving the challenges involved in reconfigurable optics for data center networks requires bridging the gap between different technologies and goals for dif-
ferent layers of the network protocol stack.

7
M. Nance Hall et al. Optical Switching and Networking xxx (xxxx) 100621

they are adjacent may change over time. Further, the bandwidth of a pheric attenuation [53]. Another aspect is misalignment due to, e.g.,
connection may also change over time. vibrations, requiring active alignment systems [54] respectively a
Under the assumption of a static physical topology, different net- tradeoff between beamwidth and received power density, depending
work architectures and best practices have been established. Some of on the distance covered [49]. In summary, even though free-space op-
these architectures include Clos, fa t-tree, and torus topologies. Best tics is an attractive alternative for many scenarios [48], and can be
practices include (over)provisioning all links such that the expected uti- seen as “fiber without the fiber” [54], these technologies “are not used in
lization is a small fraction of the total bandwidth for all connections. commercial data centers yet” [55], and hence the main challenge is
These solutions can incur high cabling costs and are inefficient. working towards their practical deployment. We refer to two recent
Reconfigurable network solutions circumvent the limitations of the specialized surveys for more details [55,56].
static network solutions by reducing cabling costs or reducing the need Sub-second Switching. In data centers, distances are short be-
to over-provision links. The flexibility of light primarily empowers tween hosts, and therefore they do not lose their strength to such a de-
these reconfigurable solutions. Some of these flexibilities include the gree that mid-line devices such as amplifiers are necessary. Therefore,
steering of light (e.g., with MEMs or polymer wa veguides) and the high applications can benefit from all of the agility of optical layer devices
capacity of fiber-optics as a medium (e.g., dense wa velength division without accounting for physical-layer impairments, which can slow
multiplexing, or DWDM , enables transmitting O(Tb/s) on a single down reconfiguration times in wide-area networks. Research has
fiber). shown that micro-second switching of application traffic is possible in
data center environments [57–59]. The ability to conduct circuit
4.1. DCN-specific technologies switching at microsecond timescales has illuminated further intrigue,
particularly for transport protocols running on top of these networks.
Innovations in reconfigurable optical networks are enabled by hard- In c-Through [21], the authors observed that throughput for TCP ap-
ware's evolution, as discussed in Section 2. There is a subset of innova- plications dropped when their traffic migrated to the optical network.
tions that are well-suited for data centers only. These are free-space op- They showed how to mitigate this by increasing the queue size for op-
tics and sub-second switching. Although we have separated these below, tical circuit switches and adjusting the host behaviors. Mukerjee et al.
there may be overlaps between free-space optics and sub-second [60] augmented their solution by expanding TCP for reconfigurable
switching systems as well. data center networks. Another method to deal with rapid reconfigura-
Free-space Optics. In free-space optics systems, light propagates tion times at a micro-second level is using traffic matrix scheduling, as
through the air from one transceiver to another. Free-space optics en- we will further elaborate in Section 4.3.
ables operators to reduce their network's complexity (a function of ca- However already e.g., Alistarh et al. [61] showcased the possibility
bling cost. These closed environments and their highly variable nature of switching in the order of nano-seconds in a thousand port 25 Gbps
of intra-data center traffic make such solutions appealing, we refer to + optical switch design. Notwithstanding, a challenging question is
the overview by Hamza et al. [48] for further application scenarios. how to make use of such fast reconfiguration times, when accounting
Recent works such as Firefly [49] have demonstrated that free-space for computation and routing update delays. Mellette et al. follow an in-
optics are capable of reducing latency for time-sensitive applications triguing design choice with their rotor switches [62], by creating de-
by routing high-volume/low-priority traffic over the wireless optical mand-oblivious connections that change in the order of micro-seconds,
network while persistently serving low-volume/high-priority traffic on in turn pre-configuring the routing in RotorNet [15] and Opera [63].
a packet-switched network. High fan-out (1-to-thousands) for free- Project Sirius expands such ideas to the sub-nano-second level [18,30],
space optics is enabled with DMDs, or Dense Micro-mirror Devices, as resulting in a demand-oblivious design that can perform end-to-end re-
shown by ProjecToR [50]. The DMDs are placed near Top-of-Rack configurations in less than 4 nano-seconds at 50 Gbps [5]. We further
(ToR) switches and pair with disco-balls, fixed to the ceiling above the discuss these strategies in Section 4.3.
racks. The DMD is programmed to target a specific mirror on the Summary. Unlike in the WAN, data center technologies allow ex-
disco-ball, guiding the light to another ToR in the data center. Fig. 11 tremely fast switching times and high fan-out across the whole net-
illustrates the main properties of the free-space optics deployment work, the latter in particular in the case of free-space optics. Hence
proposed in Ref. [50]. The deployment and operation of a free-space especially the algorithmic design ideas allow substantially more flexi-
optics data center are fraught with unique challenges, e.g., geometri- bility and often differ fundamentally, as we will see in Section 4.3.
cal placement as investigated in 2D in OWCell [51] and in 3D in Dia- Innovations in reconfigurable optical networks are enabled by hard-
mond [52], but also particularly for keeping the air clear between ware's evolution, as discussed in Section 2. There is a subset of innova-
transceivers and DMDs. Any particulate matter that the light comes tions that are well-suited for data centers only. These are free-space op-
into contact with can severely degrade performance and cause link tics and sub-second switching. Although we have separated these below,
failures should they persist. This phenomenon is known as atmos- there may be overlaps between free-space optics and sub-second
switching systems as well.
Free-space Optics. In free-space optics systems, light propagates
through the air from one transceiver to another. Free-space optics en-
ables operators to reduce their network's complexity (a function of ca-
bling cost. These closed environments and their highly variable nature
of intra-data center traffic make such solutions appealing, we refer to
the overview by Hamza et al. [48] for further application scenarios.
Recent works such as Firefly [49] have demonstrated that free-space
optics are capable of reducing latency for time-sensitive applications
by routing high-volume/low-priority traffic over the wireless optical
network while persistently serving low-volume/high-priority traffic on
a packet-switched network. High fan-out (1-to-thousands) for free-
space optics is enabled with DMDs, or Dense Micro-mirror Devices, as
shown by ProjecToR [50]. The DMDs are placed near Top-of-Rack
(ToR) switches and pair with disco-balls, fixed to the ceiling above the
racks. The DMD is programmed to target a specific mirror on the
Fig. 11. Free-space optics switching architecture for data centers [50].

8
M. Nance Hall et al. Optical Switching and Networking xxx (xxxx) 100621

disco-ball, guiding the light to another ToR in the data center. Fig. 11 the cost differences between optically and electrically switched data
illustrates the main properties of the free-space optics deployment center networks can be found in the work of Kassing et al. [64], with
proposed in Ref. [50]. The deployment and operation of a free-space an analysis for non-wired topologies in the works of Shin et al. [65]
optics data center are fraught with unique challenges, e.g., geometri- and Terzi and Korpeoglu [55].
cal placement as investigated in 2D in OWCell [51] and in 3D in Dia- To address the concerns surrounding complexity and value while
mond [52], but also particularly for keeping the air clear between raising aw areness for the necessity of optically switched interconnects,
transceivers and DMDs. Any particulate matter that the light comes researchers have constructed cost models to demonstrate the benefit of
into contact with can severely degrade performance and cause link optical switching and hybrid architectures. Wang et al. [66] developed
failures should they persist. This phenomenon is known as atmos- one such model. They conducted intra-DC traf fic measurements, which
pheric attenuation [53]. Another aspect is misalignment due to, e.g., consisted of mixed workloads (e.g., Ma pReduce, MPI, and scientific ap-
vibrations, requiring active alignment systems [54] respectively a plications). They then played the traces back in simulation, assuming
tradeoff between beamwidth and received power density, depending that three optical circuits could be created and reconfigured between
on the distance covered [49]. In summary, even though free-space op- racks every 30 s. Their data center with seven racks showed that rack-
tics is an attractive alternative for many scenarios [48], and can be to-rack traf fic could be reduced by 50% with circuit switching.
seen as “fiber without the fiber” [54], these technologies “are not used in The following sections present more cost modeling work in the con-
commercial data centers yet” [55], and hence the main challenge is text of algorithmic simulations and systems implementations.
working towards their practical deployment. We refer to two recent
specialized surveys for more details [55,56]. 4.3. Algorithms
Sub-second Switching. In data centers, distances are short be-
tween hosts, and therefore they do not lose their strength to such a de- The capability of optical circuit switching for data center networks
gree that mid-line devices such as amplifiers are necessary. Therefore, comes with the need to define new algorithms for optimizing utiliza-
applications can benefit from all of the agility of optical layer devices tion, bandwidth, fa irness, latency, or any other metric of interest. Re-
without accounting for physical-layer impairments, which can slow search has presented many different approaches for optimizing the
down reconfiguration times in wide-area networks. Research has metric relevant to the network operator in static networks. Traf fic En-
shown that micro-second switching of application traffic is possible in gineering (TE) generally refers to the determination of paths for flows
data center environments [57–59]. The ability to conduct circuit through the network, and the proportion of bandwidth levied for any
switching at microsecond timescales has illuminated further intrigue, particular flow. If the data center has a static network topology (e.g.,
particularly for transport protocols running on top of these networks. fa t-tree), then TE is simple enough that switches can conclude how to
In c-Through [21], the authors observed that throughput for TCP ap- route flows. However, introducing reconfigurable paths complicates
plications dropped when their traffic migrated to the optical network. the process of TE significantly: network elements (e.g., switches) must
They showed how to mitigate this by increasing the queue size for op- now also determine with whom and when to establish optical paths,
tical circuit switches and adjusting the host behaviors. Mukerjee et al. and when to change them.
[60] augmented their solution by expanding TCP for reconfigurable Overview. The current algorithmic ideas to establish such optical
data center networks. Another method to deal with rapid reconfigura- paths can be classified into roughly five different areas, which we will
tion times at a micro-second level is using traffic matrix scheduling, as discuss next. Due to the inherent hardware constraints (forming cir-
we will further elaborate in Section 4.3. cuits), all of them rely on 1) matchings, where on its own the main idea
However already e.g., Alistarh et al. [61] showcased the possibility is to maximize matching's weight, e.g., representing throughput, la-
of switching in the order of nano-seconds in a thousand port 25 Gbps tency, etc. However maximum matchings can be slow to compute, and
+ optical switch design. Notwithstanding, a challenging question is hence there has been interest in 2) demand-oblivious approaches, cy-
how to make use of such fast reconfiguration times, when accounting cling through different network designs, 3) traffic matrix scheduling, to
for computation and routing update delays. Mellette et al. follow an in- batch-compute a whole set of matchings ahead of time, and also lever-
triguing design choice with their rotor switches [62], by creating de- aging the speed-up of 4) machine learning algorithms. Lastly, another
mand-oblivious connections that change in the order of micro-seconds, way of quickly reacting to demand changes is by borrowing ideas from
in turn pre-configuring the routing in RotorNet [15] and Opera [63]. 5) self-adjusting data structures, in particular adapting the aspect of
Project Sirius expands such ideas to the sub-nano-second level [18,30], purely local circuit changes.
resulting in a demand-oblivious design that can perform end-to-end re- Overview. The current algorithmic ideas to establish such optical
configurations in less than 4 nano-seconds at 50 Gbps [5]. We further paths can be classified into roughly five different areas, which we will
discuss these strategies in Section 4.3. discuss next. Due to the inherent hardware constraints (forming cir-
Summary. Unlike in the WAN, data center technologies allow ex- cuits), all of them rely on 1) matchings, where on its own the main idea
tremely fast switching times and high fan-out across the whole net- is to maximize matching's weight, e.g., representing throughput, la-
work, the latter in particular in the case of free-space optics. Hence tency, etc. However maximum matchings can be slow to compute, and
especially the algorithmic design ideas allow substantially more flexi- hence there has been interest in 2) demand-oblivious approaches, cy-
bility and often differ fundamentally, as we will see in Section 4.3. cling through different network designs, 3) traffic matrix scheduling, to
batch-compute a whole set of matchings ahead of time, and also lever-
4.2. Cost mo deling aging the speed-up of 4) machine learning algorithms. Lastly, another
way of quickly reacting to demand changes is by borrowing ideas from
Momentum has been building for data centers to move to optically 5) self-adjusting data structures, in particular adapting the aspect of
switched and electrical/optical hybrid networks. However, there is a purely local circuit changes.
general reluctance to wa lk aw ay from the old paradigm of a packet- Matchings can be computed quickly [67] and often provide a
switched-only network (PSO) due to the additional complexity of opti- good approximation, especially in settings where the goal is to maxi-
cal circuit switching (e.g., the control plane management of optical cir- mize single-hop throughput along with reconfigurable links. Ma tching
cuits with shifting demand, and the variety of optical switching archi- algorithms hence frequently form the basis of reconfigurable optical
tectures available). Further, without a quantitative measure of value- networks, e.g., Helios [68], c-Through [21,69] rely on maximum
added by optical switching over PSO, DCN operators are understand- matching algorithms. If there exist multiple reconfigurable links (say b
ably reluctant to spend capital on an unvetted system. A discussion on many), it can be useful to directly work with a generalization of

9
M. Nance Hall et al. Optical Switching and Networking xxx (xxxx) 100621

matching called b-matching [70]: b-matchings are for example used in each round. Each node can only send a certain number of packets in one
Proteus [71] and its extension OSA [72], as well as in BM A [73] round. The approach by Kulkarni et al. [94] considers a model where
which relies on an online b-matching algorithm; BM A also establishes communication requests arrive online over time and uses an analysis
a connection to online (link) caching problems. In some scenarios, for based on LP relaxation and dual fitting.
example, when minimizing the average weighted path length under Self-Ad justing Data structures. A potential draw back of traf fic
segregated routing, maximum b-matching algorithms even provide matrix scheduling algorithms is that without countermeasures, the op-
optimal results [74,75]. This however is not alwa ys true, e.g., when timal topology may change significantly from one traf fic matrix
considering non-segregated routing policies [74,75], which require snapshot to the next, even though the matrix is similar. There is a se-
heuristics [49, §5.1], [76]. ries of algorithms for reconfigurable networks that account for recon-
Oblivious Ap proaches. Ma tchings also play a role in reconfig- figuration costs, by making a connection to self-adjusting data struc-
urable networks which do not account for the traf fic they serve, i.e., in tures (such as splay trees) and coding (such as dynamic Huffman cod-
dema nd-oblivious networks. The prime example here is RotorNet [15] ing) [90,95–101]. These networks react quickly and locally two new
which relies on a small set of matchings through which the network cy- communication requests, aiming to strike an optimal tradeoff be-
cles endlessly: since these reconfigurations are “dumb”, they are fa st tween the benefits of reconfigurations (e.g., shorter routes) and their
(compared to demand-aw are networks) and provide frequent and peri- costs (e.g., reconfiguration latency, energy, packet reorderings, etc.).
odic direct connections between nodes, which can significantly reduce To be more specific, the idea of the self-adjusting data structure-
infrastructure cost (also known as “bandwidth tax”) compared to based algorithms is to organize the communication partners (i.e., the
multi-hop routing, see also Teh et al. [77]. In case of uniform (delay destinations) of a given communication source in either a static binary
tolerant) traf fic, such single-hop forwa rding can saturate the network's search or Huffman tree (if the demand is known), or in a dynamic tree
bisection bandwidth [15]; for skewed traf fic matrices, it can be useful (if the demand is not known or if the distribution changes over time).
to employ Valiant load balancing [78] to avoid underutilized direct The tree optimized for a single source is sometimes called the ego-tree,
connections, an idea recently also leveraged in Sirius [5] via Chang and the approach relies on combining these ego-trees of the different
et al. [79]. Opera [63] extends RotorNet by maintaining expander sources into a network while keeping the resulting node degree con-
graphs in its periodic reconfigurations. Even though the reconfigura- stant and preserving distances (i.e., low distortion). The demand-aware
tion scheduling of Opera is deterministic and oblivious, the precompu- topology resulting from taking the union these ego-trees may also be
tation of the topology layouts in their current form is still randomized. complemented with a demand-oblivious topology, e.g., to serve low-
Expander graphs (and their variants, such as random graphs [80]) are latency flows or control traffic; see the ReNet architecture for an exam-
generally considered very powerful in data center contexts. An exam- ple [99].
ple of a demand-aw are expander topology wa s proposed in Ta le of To be more specific, the idea of the self-adjusting data structure-
Tw o Topologies [81], where the topology locally converts between based algorithms is to organize the communication partners (i.e., the
Clos and random graphs. destinations) of a given communication source in either a static binary
Traffic Matrix Scheduling. Another general algorithmic ap- search or Huffman tree (if the demand is known), or in a dynamic tree
proach is known as traffic ma trix scheduling: the algorithmic optimiza- (if the demand is not known or if the distribution changes over time).
tions are performed based on a snapshot of the demand, i.e., based on The tree optimized for a single source is sometimes called the ego-tree,
a traf fic matrix. For example, Mordia [82] is based on an algorithm and the approach relies on combining these ego-trees of the different
that reconfigures the network multiple times for a single (traffic de- sources into a network while keeping the resulting node degree con-
mand) snapshot. To this end, the traf fic demand matrix is scaled into a stant and preserving distances (i.e., low distortion). The demand-aware
bandwidth allocation matrix, which represents the fraction of band- topology resulting from taking the union these ego-trees may also be
width every possible matching edge should be allocated in an ideal complemented with a demand-oblivious topology, e.g., to serve low-
schedule. Next, the allocation matrix is decomposed into a schedule, latency flows or control traffic; see the ReNet architecture for an exam-
employing a computationally efficient [83] Birkhoff-von-Neumann ple [99].
decomposition, resulting in O(n 2) reconfigurations and durations. This Machine Learning. Another natural approach to devise algo-
technique also applies to scheduling in hybrid data center networks rithms for reconfigurable optical networks is to use machine learning.
which combine optical components with electrical ones, see e.g., the To just give two examples, xWeaver [17] and DeepConf [102] use
heuristic used by Solstice [84]. Eclipse [85] uses traf fic matrix schedul- neural networks to provide traf fic-driven topology adaptation. An-
ing to achieve a (1 − 1/e(1−ε))-approximation for throughput in the other approach is taken by Kalmbach et al. [103], who aim to strike a
hybrid switch architecture with reconfiguration delay, but only for di- balance between topology optimization and “keeping flexibilities”,
rect routing along with single-hop reconfigurable connections. Re- leveraging self-driving networks. Finally, Truong-Huu et al. [104] pro-
cently Gupta et al. [86] expanded similar approximation guarantees posed an algorithm that uses a probabilistic, Ma rkov-chain based
to multi-hop reconfigurable connections, for an objective function model to rank ToR nodes in data centers as candidates for light-path
closely related to throughput. creation.
While Eclipse is an offline algorithm, Schwartz et al. [87] presented Ac counting for Ad ditional Aspects. Last but not least, several
online greedy algorithms for this problem, achieving a provable com- algorithms account for additional and practical aspects. In the context
petitive ratio over time; both algorithms allow to account for reconfigu- of shared mediums (e.g., non-beamformed wireless broadcast, fiber1
ration costs. Another example of traffic matrix scheduling is DANs (rings)), contention and interference of signals can be avoided by using
[88–91] (short for demand-aware networks, which are optimized to- different channels and wa velengths. The algorithmic challenge is then
ward a given snapshot of the demand). DANs rely on concepts of de- to find (optimal) edge-colorings on multi-graphs, an NP-hard problem
mand-optimized data structures (such as biased binary search trees) for which fa st heuristics exist [106]. However, on specialized topolo-
and coding (such as Huffman coding) and typically aim to minimize the gies, optimal solutions can be found in polynomial time, e.g., in Wave-
expected path length [88–91], or congestion [89]. In general, the prob- Cube [107]. Shared mediums also have the benefit that it is easier to
lem features intriguing connections to the scheduling literature, e.g., distribute data in a one-to-many setting [108]. For example, on fiber
the work by Anand et al. [92], and more recently, Dinitz et al. [93] and rings, all nodes on the ring can intercept the signal [105, §3.1]. One-to-
Kulkarni et al. [94]; the latter two works however are not based on
matchings or bipartite graphs. In Dinitz et al. [93], the demands are the
edges of a general graph, and a vertex cover can be communicated in 1 In the context of data center proposals, shared fibe r is the more popular

me dium , e.g., in Refs. [72 ,82 ,10 5].

10
M. Nance Hall et al. Optical Switching and Networking xxx (xxxx) 100621

many paradigms2 such as multicast can also be implemented in other Table 3

technologies, using e.g., optical splitters for optical circuit switches or Summa ry of systems impleme ntations of reconfigurable data center net-
half-reflection mirrors for free-space optics [111–115]. works.
Fa br ic Dema nd- Novelty
4.4. Systems implementations Awar e

Heli os Hy br id ✓ Firs t hybr id sy stem using WDM for

There have been many demonstrations of systems for reconfig- [68] busty lo w-la tency traf fic
urable optics in data centers. Ma ny of the papers that we discuss in Sec- c-Thro ugh Hy br id ✓ Enla rg ed buffers for optical ports
tion 4.3 are fully operational systems. Another notable research devel- [21] increases util ization
Proj ecToR Hy br id/FSO ✓ Introduces DMDs for fr ee-sp ace
opment that does not fit into algorithms is the work by Mukerjee et al.
[50] sw itching thus enab li ng a fa n-out
[60]. They describe amendments to the TCP protocol to increase the ef- potentia l to thousa nds of nodes
ficiency of reconfigurable data center networks. These amendments in- Proteus All-optical ✓ Desi gn of an al l-optical an d
clude dynamic buffer re-sizing for switches and sharing explicit net- [71] reconfig urab le DCN.
work feedback with hosts. Moreover, Yang et al. [116] showcase an OS A [72] All-optical ✓ Demo nstrates gr eater reconfig uration
flexibil ity an d bisection ba ndwidth than
interesting cross-layer aspect where the physical layer itself is con-
hybr id ar chitectures
trolled by SDN, in the sense that they allow for transceiver tuning in Ro torNet Hy br id ✗ An al l-optical dema nd-obli vi ous DCN
real-time. Their main contributions relate to new SDN control modules [15] ar chitecture for si mp li fied networ k
and interfa ces, being orthogonal to (scheduling) algorithms. Much of ma na gement
the other work on reconfigurable DCNs are summarized in Ta ble 3. Op era All-optical ✗ Extends Ro torn et to include expa nder
[63] gr ap hs ro tations
We see two main conceptual differences in current reconfigurable
Fl at-tree Hy br id ✓ A hybr id of ra ndom gr ap hs an d Cl os
data center network designs, namely concerning 1) the demand-aw are [81] topolo gi es br ings reconfig urab le optics
or -oblivious circuit control plane and the 2) all-optical or hybrid fa b- clos er to exis ting DCNs.
ric. Sirius [5], Opera [63], and RotorNet [15] all propose a demand- So ls tice Hy br id ✓ Exploits sp ar se traf fic pa tterns in DCNs
oblivious optical layer, in essence rotating through a set of topologies, [84] to achieve fa st scheduli ng of
reconfig urab le networ ks .
letting the higher layers take advantage of the changing optical con-
Ecli ps e Hy br id ✓ Ou tperform s So ls tice by ap plyi ng
nections. To this end, there is no computational delay, but on the other [16] subm odular optimi zation theory to
hand, specifically skewed demands can suffer from performance degra- hybr id networ k scheduli ng.
dation. Demand-aw are control planes can adapt to any demands but xWeaver Hy br id ✓ Trai ns neural networ ks to construct
need careful tuning to avoid scaling and prediction issues, which then [17] perf or ma nt topolo gi es ba sed on trai ning
da ta fr om histor ic traf fic traces.
again can be inferior to demand-oblivious network designs, depending
DeepCo nf Hy br id ✓ Presents a generi c mo del for
on the scenario. Notwithstanding, the three listed demand-oblivious [102] constructing lear ning sy stems of
designs currently rely on specialized and experimental hardwa re. Re- dyna mi c optical networ ks
garding the choice of fa bric, hybrid designs are highly beneficial for WaveCu be Hy br id ✓ A mo dula r networ k ar chitecture for
small and short-lived flows, and hence a combination of packet and cir- [107] suppor ting divers e traf fic pa tterns.
Si ri us [5] All-optical ✗ Achieves na nosecond-gr an ular ity
cuit switching, such as in RotorNet [15] or Eclipse [16], can combine
reconfig uration for thousa nds of nodes
the best of both worlds. Notwithstanding, provisioning for both types
of networks leads to overheads in cost and terms of cross-fa bric effi-
ciency, and thus are not a silver bullet solution. An intriguing design in 5. Reconfigurable optical metro and wide-area networks
this context is Opera [63], as it alwa ys provisions a small diameter net-
work with optical links, emulating classic DCN properties inside their In this section, we survey recent research in reconfigurable optics in
circuit choices. However, as mentioned above, this design choice comes metropolitan (metro) and wide-area networks (WAN). Reconfigurable
with the price of demand-obliviousness, and it would be interesting to optics refers to dynamism in the physical-layer technology that enables
see how other all-optical demand-aw are systems, such as e.g., OSA high-speed and high throughput WAN communications, fiber optics.
[72], can implement such properties as well. We divide reconfigurable optical innovations into two sub-categories,
rate-adaptive transceivers, and dynamic optical paths. Rate adaptive
transceivers, or bandwidth-variable transceivers (introduced in Section
4.5. Summ ary
3.3) are optical transceivers that can change their modulation format to
adapt to physical layer impairments such as span-loss and noise. Dy-
There is a wide range of data center specific technology and algo-
namic optical paths refer to the ability to steer light, thus allowing the
rithmic ideas that enable efficient circuit switching in data center net-
edges of the network graph to change (e.g., to avoid a link that has
works, with newer developments focusing on leveraging the benefits of
failed).
fa ster circuit reconfigurations. In contrast, there has also been some re-
In this section, we survey recent research in reconfigurable optics in
cent work [117] that discusses the idea of robust topology engineering,
metropolitan (metro) and wide-area networks (WAN). Reconfigurable
e.g., adapting the circuits only every few minutes or even days [118].
optics refers to dynamism in the physical-layer technology that enables
Notwithstanding, scaling current system designs can be problematic, in
high-speed and high throughput WAN communications, fiber optics.
particular, due to the speed of the control plane and fa n-out restric-
We divide reconfigurable optical innovations into two sub-categories,
tions. Whereas one solution for the latter is free-space optics, those still
rate-adaptive transceivers, and dynamic optical paths. Rate adaptive
fa ce significant practical deployment issues in data center contexts. On
transceivers, or bandwidth-variable transceivers (introduced in Section
the other hand, demand-oblivious system designs inherently overcome
3.3) are optical transceivers that can change their modulation format to
such control plane delays, but cannot adapt well to skewed demands.
adapt to physical layer impairments such as span-loss and noise. Dy-
In their current form, they are not available as off-the-shelf hardwa re.
namic optical paths refer to the ability to steer light, thus allowing the
Designing scalable demand-aw are reconfigurable data centers is hence
edges of the network graph to change (e.g., to avoid a link that has
one of the main next challenges.
failed).
Many groups have studied the programmability and autonomy of
optical networks. Gringeri et al. [119] wrote a concise and illuminating
2 Conceptually similar ch allenges arise for coflows [10 9,11 0]. introduction to the topic. In it, the authors propose extending Software

11
M. Nance Hall et al. Optical Switching and Networking xxx (xxxx) 100621

Defined Network (SDN) principles to optical transport networks. They Dynamic optical networks must rapidly address these challenges (in
highlight challenges, such as reconfiguration latency in long-haul net- sub-second time frames) to achieve the highest possible utilization, pos-
works, and provide a trade-off characterization of distributed vs. cen- ing a significant challenge. For example, it requires multiple orders of
tralized control for an optical SDN system. They claim that a tiered hier- magnitude increases in the provisioning time for optical circuits beyond
archy of control for a multi-regional network (e.g., segregated optical what is typically offered by hardwa re vendors. Therefore, several re-
and network control loops) will offer the best quality solution. Further, search efforts have explored wa ys to automate WAN network ele-
they argue that centralized control should work best to optimize com- ments’ configuration concerning physical layer impairments in a robust
peting demands across the network, but that the controller's latency and time-efficient manner.
will be too slow to react to network events, e.g., link outages quickly. Chromatic Dispersion. DWDM makes efficient use of optical
Therefore, the network devices should keep some functionality in their fiber by putting as many distinct optical channels, each identified by a
control plane to respond to link failures in a decentralized manner, e.g., frequency (or lambda λ) onto the shared fiber. Each of these lambdas
reallocating the lost wavelengths by negotiating an alternative path be- travels at a different speed relative to the speed of light. Therefore, two
tween the endpoints. bits of information transmitted simultaneously via two different lamb-
The question of centralized vs. distributed network control is just das will arrive at the destination at two different times. Further, chro-
one example of the many interesting questions that arise when consid- matic dispersion is also responsible for pulse-broadening, which reduces
ering reconfigurable optical networks for metro and wide areas. This channel spacing between WDM channels and can cause FEC errors.
space is unique because many of the solutions here require understand- Therefore, DWDM systems must handle this physical impairment.
ing and sharing of information across layers of the network stack. For Amplified Spontaneous Emission (ASE) Noise. A significant
example, Fig. 12 illustrates interdependence between the objectives for limitation of circuit switching is the latency of establishing the circuit
communication across different layers of the stack; these features in- due to ASE noise constraints [120]. Although SDN principles can apply
clude algorithms, enabling technologies. We highlight several canonical to ROADMs and WSSs (to automate the control plan of these devices),
examples of systems that exist in those domains and across different physical layer properties, such as Noise Figure (NF) and Ga in Flatness
layers. In this section, we will explore these examples more deeply (GF) complicate the picture. When adding or removing optical channels
along with other related efforts. to or from a long-haul span of fiber, traversing multiple amplifiers, the
amplifiers on that path must adjust their gain settings to accommodate
5.1. Metro/WAN-specific challenges and solutions the new set of channels. To this end, researchers have worked to ad-
dress the challenge of dynamically configuring amplifiers. Oliveira
There are many reasons for the prevalence of optical fiber as the de- et al. [121] demonstrated how to control gain on EDFAs using GM PLS.
fa cto leader for long-distance communications. First, it has incredible They evaluated their solution on heterogeneous optical connections
reach compared to copper—optical signals can propagate 80–100 km (10, 100, 200, and 400 Gbps) and modulations (OOK, QPSK, and 16-
before being amplified. Second, it has an incredibly high bandwidth QAM). They used attenuators to disturb connections and allow their
compared to the radio spectrum. Third, optical fiber itself has proved to GM PLS control loop to adjust the amplifier's gains. They show that
be a robust medium over decades, as improvements to the transponders their control loop helps amplifiers to adjust while transmitting bits with
at the ends of the fiber have enabled operators to gain better value out BER below the FEC threshold for up to 6 dB of added attenuation.
of the same fiber year af ter year. Moura et al. [122] present a machine learning approach for config-
To design a WAN, the network architect must solve several difficult uring amplifier gain on optical circuits. Their approach uses case-based
challenges, such as estimating the demand on the network now and reasoning (CBR) as a foundation. The intuition behind CBR is that the
into the future, optimal placement of routers and quantity of ports on gain setting for a set of circuits will be similar if similar circuits are pre-
those routers within the network, and optimal placement of amplifiers sent on a shared fiber. They present a genetic algorithm for configuring
in the network. amplifiers based on their case-based reasoning assumption. They show
Ma ny design challenges solve more easily in a static WAN, where that their methodology is suitable for configuring multiple amplifiers
optical channels are initialized once and maintained for the network's on a span with multiple optical channels. In a follow-up study, they
life. For example, amplifiers carrying the channel must have their gain present FAcCBR [123], an optimization of their genetic algorithm,
set in such a wa y that the signal is transmitted while maximizing the which yields gain recommendations more quickly by limiting the num-
signal-to-noise ratio (SNR). This calculation can take minutes or hours ber of data-points recorded by their algorithm.
depending on the network's characteristics (e.g., the number of indeter- Synchronization. Ma naging a WAN requires coordinating ser-
minate hosts and the number of distinct channels on shared amplifiers). vices (e.g., end-to-end connections) among diverse sets of hardwa re
appliances (transponders, amplifiers, routers), logically and consis-

Fig. 12. To deploy and operate reconfigurable optical networks in metro and wide-area networks require expertise spanning the bottom three layers of the network
stack, including algorithms and enabling technology. We highlight several canonical examples of systems that exist in this space and explore other related works
along with these systems more deeply in this section.

12
M. Nance Hall et al. Optical Switching and Networking xxx (xxxx) 100621

tently. The Internet Engineering Ta sk Force (IETF) has defined proto- save 40%–60% of the required transponders in the network with this
cols and standards for configuring WAN networks. As the needs and multi-layer approach. The networks they looked at were Deutsche
capabilities of networks have evolved, so have the protocols. Over the Telekom [134] and Telefonica Spain core networks. These authors’
years, new protocols have been defined to bring more control and au- work provides a strong motivation for jointly optimizing IP and Opti-
tomation to the network operator's domain. These protocols are Sim- cal network layers and sharing of information between the two.
ple Network Ma nagement Protocol (SNMP) [124] and Network Con- Papanikolaou et. at. [135] propose a cost model for joint multi-
figuration Protocol (NETCONF) [125]. Additionally, network opera- layer planning for optical networks. Their paper presents three network
tors and hardwa re vendors have been working to define a set of gener- planning solutions; dual-plane network design, fa ilure-driven network
alized data models and configuration practices for automating WAN design, and integrated multilayer survivable network design. They
networks under the name OpenConfig [126]. Although OpenConfig is show that dual-plane and fa ilure-driven designs over-provision the IP
not currently standardized with the IETF, it is deployed and has layer, leaving resources on the table that are only used if link fa ilures
demonstrated its value in several unique settings. occur. They show that integrated multi-layer survivable network de-
In addition to the standardized and proposed protocols for general- sign enables a significant reduction in CapEx and that the cost savings
purpose WAN (re)configuration, there has been a push by various inde- increases beyond dual plane and fa ilure driven designs.
pendent research groups to design and test protocols specifically for re- Cost models for evaluating C-ROADM vs. CDC-ROADM network
serving and allocating optical channels in WAN networks. architectures are described by Kozdrowski et al. [136]. They show that
One protocol was developed in conjunction with the CORONET for three regional optical networks (Germany, Poland, USA), CDC-
[127] program, whose body of research has led to several other devel- ROADM based networks can offer 2 to 3× more aggregate capacity
opments in reconfigurable optical WANs. The proposal, by Skoog et al. over C-ROADM based networks. They evaluate their model with uni-
[128], describes a three-way handshake (3WHS) for reserving and es- form traf fic matrices (TMs) and apply various scalar multipliers to the
tablishing optical paths in single and multi-domain networks. In the TM . Their model accounts for many optical hardwa re related con-
3WHS, messages are exchanged over an optical supervisory channel straints, including the number of available wa velengths and cost fa c-
(OSC)—an out-of-band connection between devices isolated from user tors associated with manual-(re)configuration of C-ROADM elements.
traffic. The transaction is initiated by one Optical Cross-Connect However, their model doesn't include an optical-reach constraint. They
(OXCA) and directed at a remote OXC, OXCZ. At each hop along the limit solver computation time to 20 h and present the best feasible solu-
way, the intermediate nodes append the available channels to the mes- tion determined in that amount of time.
sage. Then, OXCZ chooses a channel via the first-fit strategy [129] and Service velocity refers to the speed with which operators may grow
sends a message to OXCA describing the chosen channel. Finally, OXCA their network as demand for capacity grows. Woodwa rd et al. [137]
activates the chosen channel and beings sending data over it to OXCZ. tackles the problem of increasing service velocity for WANs. In this
This protocol is claimed to meet the CORONET project standard for a context, they assume a network of colorless non-directional ROADMS
setup time of 50 ms + RTT between nodes. Bit arrays are used to com- (CN-ROADMs), in which any incoming wa velength can be routed on
municate the various potential channels between nodes and are any outgoing fiber. Note that CN-ROADMs are also called CD-
processed in hardware. The blocking probability is 10−3 if there is one ROADMs in other papers. These both refer to the same ROADM archi-
channel reserved between any two OXC elements so long as there are at tecture. They claim that one of the largest impedances for network
least 28 total channels possible between OXCs [128]. growth in these networks is the availability of regenerators. To solve
this problem, they present three algorithms for determining regenera-
5.2. Cost mo deling tors’ placement in a network as service demand grows. The algorithms
are: locally aw are, neighbor aw are, and globally aw are. Each algo-
Fiber infrastructure for wide-area networks is incredibly costly. Pro- rithm essentially considers a broader scope of the network, which a
visioning of fiber in the ground requires legal permitting processes node uses to determine if an additional regenerator is needed at the site
through various governing bodies. As the length of the span grows be- at a particular time. They show, via Monte Carlo simulations, varying
yond metropolitan areas, to connect cities or continents, the number of optical reach and traf fic matrices. The broadest scope algorithm per-
governing bodies with whom to acquire the legal rights to lay the fiber forms the best and allocates enough regenerators at the relevant sites
grows [130]. Then, keeping the fiber lit also incurs high cost; power re- without over-provisioning. This work shows that service velocity is im-
quirements are a vital consideration for wide-area network provision- proved with demand forecasting, enabling infrastructure to be placed
ing [131]. Therefore, reliable cost models are necessary for deploying to meet those projected demands.
and managing wide-area networks. In this section, we look at cost Programmable and elastic optical networks can also work together
modeling efforts particularly suited for reconfigurable optical net- with Network Function Virtualization (NFV) to offer lower-cost ser-
works. vice-chaining to users. Optimal strategies have been demonstrated,
An early study on the cost comparison of IP/WDM vs IP/OTN net- with heuristic algorithms, to quickly find near-optimal solutions for
works (in particular: European backbone networks) wa s conducted by users and service brokers by Chen et al. [138]. In their work, they take
Tsirilakis et al. in Ref. [132]. The IP/WDM network consists of core a game-theoretic approach to modeling the competition among service
routers connected directly over point-to-point WDM links in their brokers—who complete offering the lowest cost optical routs and ser-
study. In contrast, the IP/OTN network connects the core routers vice chains, and between users—who compete to find the lowest cost
through a reconfigurable optical backbone consisting of electro-optical and highest utility service chains among the brokers. They demonstrate
cross-connects (OXCs) interconnected in a mesh WDM network. both parties’ strategies, which converge on low-latency service chain
Capacity planning is a core responsibility of a network operator in solutions with low blocking probability for optical paths.
which they assess the needs of a backbone network based on the pro- Modeling opportunity cost of optically switched paths is explored by
jected growth of network usage. Gerstel et al. [133] relates the capac- Zhang et al. [139]. In their work, they present an algorithm for quickly
ity planning process in detail, which includes finding links that require evaluating the opportunity cost of a wa velength-switched path. Given
more transponders and finding shared-risk-link-groups that need to be a request and a set of future requests, the opportunity cost for accom-
broken-up, among other things. They note that in this process, the IP modating the initial request is the number of future requests blocked as
and Optical network topologies are historically optimized separately. a result of the accommodation. Thus, the network operator's goal is to
They propose an improvement to the process via multilayer optimiza- minimize opportunity cost by permitting connections that interfere
tion, considering the connection between IP and optical layers. They with the fewest future requests.

13
M. Nance Hall et al. Optical Switching and Networking xxx (xxxx) 100621

5.3. Algorithms in two directions, via theoretic scheduling results and for improvements
on deadline-constrained transfers.
Jointly optimizing both the optical and the network layer in wide- In DaRTree [152], Luo et al. develop an appropriate relaxation of
area networks leads to new opportunities to improve performance and the cross-layer optimization problem for bulk transfers under deadlines.
efficiency, while introducing new algorithmic challenges. In contrast to Their approach relies on a non-greedy allocation in an online setting,
the previously discussed data center networks, it is impossible to create which allows future transfers to be scheduled efficiently without need-
new topological connections in a wide-area network (without deploy- ing to reallocate currently utilized wa velengths. To enhance multicast
ing more fiber. Free-space optics solutions don't apply here). Instead, transfers (e.g., for replication), they develop load-adaptive Steiner Tree
reconfigurability is possible by adjusting and shifting bandwidth capac- heuristics.
ities along the fiber edges, possibly over multiple hops. Hence, we need Jia et al. [153] design various online scheduling algorithms and
a different set of algorithmic ideas that optimize standard metrics such prove their competitiveness in the setting of OWAN [150]. The authors
as throughput, completion time, blocking probability, and resilience. In consider the minimum makespan and sum completion time, analyzing
this section, we discuss recent papers that tackle these issues, starting and extending greedy cross-layer scheduling algorithms, achieving
with some earlier ones. Moreover, there is the need for some central small competitive ratios. Dinitz and Moseley [154] extend the work of
control to apply the routing, policy, lightpath, etc. changes, for which Jia et al. by considering a different objective, the sum of flow times in
we refer to recent surveys [140,141]. an online setting. They show that resource augmentation is necessary
Jointly optimizing both the optical and the network layer in wide- for acceptable competitive bounds in this setting, leading to nearly (of-
area networks leads to new opportunities to improve performance and fline) optimal competitive ratios. While their algorithms are easy to im-
efficiency, while introducing new algorithmic challenges. In contrast to plement (e.g., relying on ordering by release time or by job density), the
the previously discussed data center networks, it is impossible to create analysis is complicated and relies on linear program relaxations. More-
new topological connections in a wide-area network (without deploy- over, their algorithm also allows for constant approximations in the
ing more fiber. Free-space optics solutions don't apply here). Instead, weighted completion time setting, without augmentations.
reconfigurability is possible by adjusting and shifting bandwidth capac- Another (algorithmic) challenge is the integration of cross-layer al-
ities along the fiber edges, possibly over multiple hops. Hence, we need gorithms into current traf fic engineering systems. Such TEs are tried
a different set of algorithmic ideas that optimize standard metrics such and tested, and hence service providers are reluctant to adapt their de-
as throughput, completion time, blocking probability, and resilience. In signs. To this end, Singh et al. [155] propose an abstraction on how dy-
this section, we discuss recent papers that tackle these issues, starting namic link capacities (e.g., via bandwidth variable transceivers) can be
with some earlier ones. Moreover, there is the need for some central inserted into classic TEs. Even though the TE is oblivious to the optical
control to apply the routing, policy, lightpath, etc. changes, for which layer, an augmentation of the IP layer with fa ke links enables cross-
we refer to recent surveys [140,141]. layer optimization via the TE. A proposal [156] for a new TE for such
Routing aspects are explored intensively in this context. Algorithmic dynamic link capacities is discussed in the next Section 5.4. Singh et al.
approaches to managing reconfigurable optical topologies have been [155] also discuss consistent update methods [157] for dynamic link
studied for a decade, but are recently gaining new attention. In early capacities, which Tseng [158] formalizes into a rate adaption planning
work by Kodialam et al. [142] explores IP and optical wa velength problem, providing intractability results and an LP-based heuristic.
routing for a series of connection requests. Their algorithm determines OptFlow [159] proposes a cross-layer abstraction for programmable
whether a request should be routed over the existing IP topology, or if a topologies as well, but focuses on shifting wa velengths between neigh-
new optical path should be provisioned for it. Interesting work by boring fibers. Here, the abstraction concept is extended by not only cre-
Brzezinski and Modiano [143] who leverage matching algorithms and ating fa ke links but also augmenting the traf fic matrix with additional
Birkhoff–von Neumann matrix decompositions and evaluate multi-vs. flows. As both links and flows are part of the input for TEs, OptFlow en-
single-hop routing 3 in WDM networks under stochastic traf fic. How- ables the compilation of optical components into the IP layer for vari-
ever, the authors mostly consider relatively small networks, e.g., with ous traf fic engineering objectives and constraints. Concerning consis-
three to six nodes. For larger networks, shortest lightpath routing is a tent updates, classic flow-based techniques [157] carry over, enabling
popular choice [145]. Another fundamental aspect frequently consid- consistent cross-layer network updates too.
ered in the literature regards resilience [146–148]. For example, Xu Optimizing reconfigurable optical networks for circuit provisioning
et al. [146] investigate resilience in the context of shared risk link and per flow rate allocation is a complex and challenging endeavor; the
groups (SLRGs) and propose a method on how to provision the circuits static routing and wa velength allocation problem is NP-complete
in a WAN. To this end, they construct Integer Linear Programs to ob- [160]. Recent work by Guo et al. [161] explores the potential for an
tain maximally SLRG-diverse routes, which they then augment with artificial intelligence (AI) implementation of a network controller using
post-processing for DWDM system selection and network design issues. deep-learning. They describe a network control agent based on deep-
We now introduce further selected algorithmic works, starting with the learning which determines where and when to activate and deactivate
topic of bulk transfers [149]. a limited set of circuits given a snapshot of demand between hosts in
In OWAN [150], Jin et al. optimize bulk transfers in a cross-layer the network. They also explore inherent draw backs and precautions to
approach, which leverages both the optical and the network layer. consider settings in which such an agent is deployed. Their study offers
Their main objective is to improve completion time; while an integer insights for the potential benefit of an AI-assisted optical network con-
linear program formulation would be too slow, the authors rely on a troller, and novel challenges to consider for their given model.
simulated annealing approach. A local search shifts the wa velength al- Algorithms that optimize optical network topology for higher-layer
locations, allowing heuristic improvements to be computed at a sub- applications, such as virtual network functions (VNF) have recently
second scale. The scheduling of the bulk transfer then follows the stan- gained attention. In particular, VNF network embedding (VNF-NE) has
dard shortest job first approaches. When updating the network state, if been studied by various groups [162,163]. VNFs are an abstraction of
desired, OWAN can extend prior consistent network update solutions resources in networks that have traditionally been deployed as hard-
[151] by introducing circuit nodes in the corresponding dependency wa re devices (e.g., intrusion detection systems, firewa lls, load-balances,
graphs. OWAN also considers deadline constrained traf fic, implement- etc.). Now, instead of monolithic hardwa re appliances many of these
ing the earliest deadline first policy. Follow-up work extended OWAN devices are deployed as softwa re on commodity servers, giving more
flexibility to add and remove them at will and yielding cost-savings for
network operators. Network embedding is a physical layer abstraction
3 See also the idea of lightpath splitting in Elastic Optical Networks [14 4].

14
M. Nance Hall et al. Optical Switching and Networking xxx (xxxx) 100621

for creating end-to-end paths for network applications or network ing the transceivers off. Then, they propose a TE objective function via
function virtualization (NFV) service chains. Paths have requirements linear-programming, to minimize churn, or impact due to SNR fluctua-
for both bandwidth and CPU resources along the service chain. Wang tions, in a WAN. Finally, they evaluate their TE controller on a testbed
et al. [162] proves this problem to be NP-complete for elastic optical WAN and show that they improve network throughput by 40% over a
networks. Soto et al. [163] provides an integer linear program (ILP) to competitive softwa re-defined networking controller, SWAN [168].
solve the VNF-NE problem. The ILP solution is intractable for large net- Dynamic Optical Paths. In the early aughts, researchers explored
works. Thus, they provide a heuristic that uses a ranking-system for op- the benefit of dynamic optical paths for networks in the context of
tical paths. Their heuristic ranks optical paths by considering a set of grid-computing. Early efforts by Figueira et al. [169] addressed how a
end-to-end connection requests. Paths with higher rank satisfy a more system might manage dynamic optical paths in networks. In this work,
significant proportion of the demand for bandwidth and CPU among the authors propose a web-based interfa ce for submitting optical re-
all of the requests. configuration requests and a controller for optimizing the requests’ ful-
Optical layer routing with traf fic and application constraints is a fillment. They evaluate their system on OMNInet [170], a metropoli-
difficult problem. The running theme has been that linear programming tan area network with 10 Gbps interconnects between 4 nodes and
solutions can find provably optimal solutions [164], but take too long Wavelength Selective Switches between them. They claim that they
to converge for most use cases. However, network traf fic is not entirely can construct optical circuits between the OMNInet nodes in 48 s. Fur-
random and therefore has an underlying structure that may be ex- ther, they show that amortized setup time and transfer is fa ster than
ploited by offline linear program solvers, as shown by Kokkinos et al. packet-switching for files 2.5 Gb or larger (assuming 1 Gbps or greater
[165]. They use a two-stage approach for routing optical paths in an optical interconnect and 300 Mbps packet switching throughput).
online manner. Their technique finds periodic patterns over an epoch They go on to evaluate file transfer speeds using the optical intercon-
(e.g., daily, weekly, or monthly) and solves the demand characterized nect and show that they can archive average transfer speeds of 680
within the epoch with an offline linear program. Then, their online Gbps. Iovanna et al. [171] address practical aspects of managing mul-
heuristic makes changes to the topology to accommodate random tilayer packet-optical systems. They present a set of useful abstractions
changes in demand within the epoch. for operating reconfigurable optical paths in traf fic engineering using
an existing management protocol, GM PLS.
5.4. Systems implementations Stability is an important feature of any network. An interesting
question about reconfigurable optical networked systems arises regard-
The integration of reconfigurable optics with WAN systems has been ing the stability of optically switched paths. That is if the topology can
impracticable due to its cost and a lack of convergence on cross-layer continuously change to accommodate random requests, what service
APIs for managing the WAN optical layer with popular SDN controllers. guarantees can the network make? Can the fluctuation of the optical
However, some exciting work has demonstrated the promise for recon- layer be detrimental to IP layer services? Chamania et al. [172] explore
figurable optics in closed settings. Notably, RADWAN [156] and CORO- this issue in detail, providing an optimal solution to keep quality of ser-
NET [127] for bandwidth-variable WAN systems and systems with dy- vice guarantees for IP traf fic while also improving performance beyond
namic optical paths, respectively. In this section, we explore reconfig- static optical layer systems.
urable optical WAN systems more deeply in these two contexts. Table 4 Blocking probability is a crucial metric for assessing the flexibility
summarizes these systems. of an optical network. It is the probability that a request for an end-to-
Bandwidth Variable Transceivers. A team of researchers at Mi- end lightpath in the network cannot be provisioned. Turkcu et al.
crosoft evaluates bandwidth variable transponders' applicability for in- [173] provides analytical probability models to predict the blocking
creased throughput in Azure's backbone in North America [40]. They probability in ROADM based networks with tunable transceivers and
find that throughput for the WAN can increase if they replace the fixed- validate their models with simulation considering two types of ROADM
rate transponders in their backbone network with three-wa y sliceable architecture in their analysis, namely share-per-node and share-per-
transponders. They also show that for higher-order slices, bandwidth link. In share-per-link, each end of a link has a fixed number of
gran increases at diminishing returns. transponders that can use it. In share-per-node, a node has a fixed set of
Traf fic Engineering with rate-adaptive transceivers wa s recently transponders that may use any incident links. The authors show that a
proposed by Singh et al. [156]. The authors are motivated by a data- low tunable range (4–8 channels, out of 32 possible) is sufficient for re-
set of Microsoft's WAN backbone Signal-to-Noise ratio from all trans- ducing blocking probability in two topologies, NSF Net (14 Nodes),
ceivers in the North-American backbone, over two and a half years. and a ring topology with 14 nodes. As the tunable range moves beyond
They note that over 60% of links in the network could operate at 8 and up to 32, there is little to no benefit for split-per-node and share-
0.75× higher capacity and that 25% of observed outages due to SNR per-link architectures. As the load on the network increases, blocking
drops could be mitigated by reducing the modulation of the af fected probability increases, as well as the gap between blocking probability
transceivers. They evaluate the reconfigurability of Ba ndwidth- of split-per-node and split-per-link decreases.
Variable Transponders, showing that reconfiguration time for the Ba ndwidth-on-demand (BoD) is an exciting application of recon-
transceivers could be reduced from minutes to milliseconds by not turn- figurable networks. Von Lehmen et al. [127] describe their experience
in deploying BoD services on CORONET, DARPA's WAN backbone.
They implement protocols for add/dropping wa velengths in their WAN
Table 4
with a novel 3-wa y-handshake protocol. They demonstrate how their
Summary of systems implementations of reconfigurable wide area networks.
system can utilize SWAN [168] Traf fic Engineering Controller as one
BVT Network Amps. Algorithms
such application that benefits from the BoD service.
Design
More recently, there has been a resurgence of academic work high-
CORONET ✗ ✗ ✗ ROLEX protocol lighting the potential benefit of dynamic optical paths in the WAN. One
[127] such system, called OWAN (Optical Wide-Area Network) [174], pro-
OWAN [150] ✗ ✗ ✗ Simulated Annealing
poses how to use dynamic optical paths to improve the delivery time
FACcBR [122] ✗ ✗ ✓ Case Based Reasoning
RADWAN ✓ ✗ ✗ Linear Program
for bulk transfers between data centers. They build a testbed network
[156] with home-built ROADMs and implement a TE controller to orchestrate
DDN [167] ✗ ✓ ✓ Time-slotted packet scheduling bulk transfers between hosts in a mesh optical network of nine nodes.
Iris [4] ✗ ✓ ✓ Shortest path for any failure They compare their results with other state-of-the-art TE systems, em-
scenario

15
M. Nance Hall et al. Optical Switching and Networking xxx (xxxx) 100621

phasizing that OWAN delivers more transfers on time than any other modules, and wa velength and fiber collision [177]. We expect silicon
competing methods. photonics to bring down the cost of transport hardwa re, thereby in-
Dynamic optical paths increase the complexity of networks and ca- creasing access to such devices and lowering entry barriers for re-
pacity planning tasks because any optical fiber may need to accommo- search and development.
date diverse and variable channels. However, this complexity is re- Data center networks. Our understanding of algorithms and
wa rded with robustness or tolerance to fiber link outages. Gossels et al. topologies in reconfigurable networks is still early, but first insights
[175] propose dynamic optical paths to make long-haul networks into efficient designs are being published. One front where much more
more robust and resilient to node and link fa ilures by presenting algo- research is required concerns the modeling (and dealing with) recon-
rithms for allocating bandwidth on optical paths dynamically in a figuration costs. Indeed, existing works differ significantly in their as-
mesh network. Their objective is to protect networks from any single sumptions, even for the same technology, making it challenging to
node or link fa ilure event. To this end, they present an optimization compare algorithms. Related to this is also the question of how recon-
framework for network planners, which determines where to deploy figurations af fect other layers in the networking stack, and how to
transponders to minimize costs while running a network over dynamic design (distributed) controllers. In terms of algorithms, even though a
optical paths. majority of problems are intractable to solve optimally, due to inte-
Another effort in reducing the complexity of dynamic optical path gral connection constraints, the question of approximation guarantees
WAN systems wa s presented by Dukic et al. [4]. Their system, Iris, ex- is mostly open. For example, consider designing a data center with
ploits a unique property of regional connectivity, i.e., the vast abun- minimum average weighted path length. A logarithmic approxima-
dance of optical fiber in dense metropolitan areas [176]. They find that tion is easy to achieve by simply minimizing the diameter of a (con-
the complexity of managing dynamic optical paths is greatly reduced stant-degree) static topology. However, computing an optimal solu-
when switching at the fiber-strand level versus the (sub-fiber) wa ve- tion is NP-hard. So, can we obtain polynomial approximation algo-
length level. To this end, they detail their design trade-off space for in- rithms with constant performance trade-offs? Similarly, do good
ter-data center connectivity across metropolitan areas. They deploy (fixed) parameter characterizations enable efficient run times, and
their system in a hardwa re testbed to emulate connectivity between what can we expect from e.g., linear time and distributed algorithms?
three data centers, verifying that optical switching can be done in Moreover, beyond general settings, how do specific (oblivious) net-
50–70 ms over three amplifiers. They obviate amplifier reconfiguration work designs enable better algorithms, and how does their design in-
delays by conducting fiber-level switching rather than wa velength- terplay with topologies of the same equipment cost?
level. Thus, the amplifiers on a fiber path are configured once for the Next, going beyond scheduling, how can the framework of online
channel that traverses it. When a circuit changes its path, aw ay from algorithms be leveraged in this context? Ideally, we wa nt a reconfig-
one data center and towa rds another, it uses a series of amplifiers that urable link to exist befo re the traf fic appears. How can we balance this
have been pre-configured to accommodate the loss of that given cir- from a worst-case perspective? In this context, traf fic prediction tech-
cuit. niques might reduce the possible solution space massively, but we will
Inter-data center network connectivity over a regional optical still need extremely rapid reaction times to new traf fic information.
backbone wa s also investigated by Benzaoui et al. [167]. Their system, Another open challenge is the efficient interplay between reconfig-
Deterministic Dynamic Netwo rk (DDN), imposes strict constraints for urable and non-reconfigurable network parts. Theory for specific re-
application layer latency and jitter. They show that they can reconfig- configurable topologies (e.g., traf fic matrix scheduling for a single opti-
ure optical links in under 2 ms, and guarantee consistent latency and cal switch) has seen much progress. However, more general settings,
jitter through their time-slotted scheduling approach. particularly non-segregated routing onto both network parts, are still
an open issue, beyond an abstract view of the combination with a sin-
5.5. Summ ary gle packet switch.
Metro and Wi de-area Networks. Metro and wide-area optical
Reconfigurable optics for metro and wide-area networks have networks are rich with open challenges. The works presented in this
gained substantial attention in the last decade. This push requires cross- section highlight significant developments that have been made to-
domain collaboration as demand aw are changes at the optical layer wa rds reconfigurable WAN systems and illuminate great benefits for
are influenced by physical layer impairments (signal-loss, chromatic such systems. However, programmability, cross-layer information
dispersion, noise, etc.), in addition to higher-layer performance metrics sharing, and physical properties of light still must be solved. On the
(latency, demand, congestion, etc.). There are various novel works that programmability front, efforts such as OpenConfig [126], Open-
have addressed several fundamental questions in reconfigurable optical ROADM [178], and ONOS [179] are working to provide white-box
networks. Cost-modeling efforts predict network performance with system stacks for optical layer equipment. If these are widely
various classes of reconfigurable hardwa re. Algorithmic work suggests adopted and standardized, this will open the door for agile and effi-
efficient methods for efficiently managing network layer and optical cient use of wide-area networks for a variety of applications (e.g.,
layer elements in the fa ce of shifting traf fic demands. Researchers have new tools to combat DD oS [180]). Other challenges include wran-
proposed and prototyped several systems for reconfigurable optical gling with the physical constraints of efficient and rapidly reconfig-
networks in recent years, but much of this work is still in the design and urable WANs, for example, coordination of power adjustments across
proof-of-concept phase. All in all, there are still many open challenges amplifiers for long-haul circuits.
ahead to widely deploy and efficiently utilize reconfigurable optics in
production networks, as we discuss next. 7. Conclusion and future work

6. Open challenges in reconfigurable optical networks Reconfigurable optical networks are a young technology, and much
of their potential and limitations are not well-understood today. In this
Hardware technologies. The development of hardwa re for re- paper, we have specifically considered data center and wide-area net-
configurable optical networking is a burgeoning field in engineering works. Still, many other networks may benefit from similar technolo-
and research. While CDC-F ROADMs exist today, they are costly to gies, and even in our context, the tradeoffs between costs and benefits
produce, and their capabilities are found lacking. In particular, the (e.g., in terms of resilience, performance, efficiency) are not well under-
benefit of integrating CDC-F ROADMs with optical transport net- stood. In particular, these tradeoffs also depend on the specific technol-
works is limited by cascading fiber impairments, signal loss at WSS ogy, e.g., on the reconfiguration time, as well as the traf fic pattern; for

16
M. Nance Hall et al. Optical Switching and Networking xxx (xxxx) 100621

example, demand-aw are reconfigurable networks may only be useful if ar chitectures an d dyna mi c ba ndwidth al lo cation al go ri thms , IEEE
Co mm un. Surv . Tutori al s 10 (2008) 46–60, https: //doi. or g/10.1109/
the traf fic pattern exhibits temporal and spatial structure [181]. We CO MS T.2008.4625804.
currently specifically lack models for reconfiguration costs, and these [14] Bannis ter J.A. , Fr atta L. , Gerl a M. , Topolo gi cal desi gn of the wa velength-
costs, in turn, depend on the control plane, which is another open re- divi si on optical networ k, IEEE INFO CO M’90, IEEE Co mp uter So ciety, 1990,
pp. 1005–1006.
search challenge. It is not clear whether decentralized control planes [15] Mell ette W.M. , McGuinness R. , Ro y A. , Forencich A. , Papen G., Snoeren A.
are alwa ys superior to centralized ones, or whether hybrid designs are C. , Porter G., Ro torNet: a scal ab le, lo w-comp lexi ty, optical da tacenter
required. It is also not clear how to optimally design such control networ k, SIGCOM M, 2017.
[16] Venkatak ri shna n S. B., Alizad eh M. , Visw an ath P., Co stly circuits,
planes. From an algorithmic point of view, reconfigurable optical net-
subm odular schedules an d ap pr ox im ate cara th’eodor y theorems ,
works present a mostly uncharted complexity landscape. Whereas clas- SIGMETRICS , AC M, 2016, pp. 75–88.
sic networking problems can largely rely on decades of optimization [17] Wang M. , Cu i Y. , Xi ao S. , Wang X. , Ya ng D., Chen K. , Zhu J., Neural
and graph theory, reconfiguration adds new and different twists to net- networ k meets DCN: traf fic-dr iv en topolo gy ad ap tation with deep lear ning,
PO MACS 2 (2018) 26:1-26:25.
working problems. [18] La nge S. , Ra ja A. , Shi K. , Ka rp ov M. , Behrendt R. , Cl ethero e D., Ha ll er I.,
We hence hope that our survey can help to put the new concepts, Ka ri nou F. , Fu X. , Li u J., Luka shchuk A. , Thom sen B., Jozwik K. , Co sta P.,
technologies and challenges of reconfigurable optical networks into Ki ppenberg T.J., Ball an i H. , Sub-na nosecond optical sw itching using chip-
ba sed so li ton mi crocom bs , Op tical Fiber Co mm unication Co nference
perspective and hence help researchers to bootstrap and contribute to (OFC ’20), The Op tical So ciety (OSA ), 2020.
this emerging field. [19] Ghobad i M. , Ma haja n R. , Phan is hayee A. , Blan che P.-A. , Ra stegar fa r H. ,
Glick M. , Ki lper D., Desi gn of Mi rr or As semb ly for an Agil e Reconfig urab le
Data Center Interconnect, 2016 Technical Report MS R-TR-2016-33.
Funding [20] Shak eri A. , Garr ich M. , Brav al heri A. , Ca regl io D., So l´e-Pareta J.,
Fuma ga ll i A. , Traf fic al lo cation stra tegi es in ws s-ba sed dyna mi c optical
This work is supported by National Science Foundation (CNS networ ks , J. Op t. Co mm un. Netw. 9 (2017) B112–B123.
[21] Wang G., Anders en D.G., Ka mi nsky M. , Papa gi an na ki K. , Ng T.E., Ko zuch
1850297), a UO Faculty Research Award, a Ripple Faculty Fellowship,
M. , Ry an M. , c-thro ugh: Part-time optics in da ta centers, Proceedings of the
and by the European Research Council (ERC) (grant agreement AC M SIGCOM M 2010 Co nference, 2010, pp. 327–338.
864228). The views and conclusions contained herein are those of the [22] Xi a W., Wen Y. , Foh C. H. , Niya to D., Xi e H. , A surv ey on so ftwa re-defined
authors and should not be interpreted as necessarily representing the of- networ ki ng, IEEE Co mm un. Surv . Tutori al s 17 (2014) 27–51.
[23] Ki lper D., Berg ma n K. , TURBO: Tera bits/s Using Reconfig urab le Bandwidth
ficial policies or endorsements, either expressed or implied, of NSF, UO, Op tics (Final Report), 2020 https: //www. os ti.go v/biblio/1618041.
Ripple, or ERC. [24] La ntz B., D´ıa z-Mo ntiel A. A. , Yu J., Ri os C. , Ruffini M. , Ki lper D.,
Demo nstration of so ftwa re-defined pa cket-optical networ k emulation with
mi ninet-optical an d onos , Op tical Fiber Co mm unication Co nference (OFC )
Decl aration of competing interest 2020, Op tical So ciety of Am erica, 2020, p. M3Z.9.
[25] Chatterj ee B.C. , Ba S. , Ok i E., Fr ag mentation pr oblems an d ma na gement
The authors declare that they have no known competing financial ap pr oa ches in elas tic optical networ ks : a surv ey, IEEE Co mm un. Surv .
Tutori al s 20 (2017) 183–210.
interests or personal relationships that could have appeared to influ- [26] Ka chri s C. , Tomk os I., A surv ey on optical interconnects for da ta centers,
ence the work reported in this paper. IEEE Co mm un. Surv . Tutori al s 14 (2012) 1021–1036.
[27] Taboad a J.M. , Ma ki J.J., Tang S. , Sun L. , An D., Lu X. , Chen R. T., Thermo -
References opticall y tuned cascad ed poly mer wa veguide taps , Appl. Phys . Lett. 75
(1999) 163–165.
[28] Yeniay A. , Gao R. , Taka ya ma K. , Gao R. , Gari to A. F. , Ultra-lo w-lo ss
[1] Thya ga turu A. S. , Mercia n A. , McGarr y M. P., Reis sl ein M. , Kell erer W.,
poly mer wa veguides, J. Li ghtwav e Technol. 22 (2004) 154–158.
So ftwa re Defined Op tical Networ ks (SDO NS): A Co mp rehensiv e Surv ey, [29] De Feli pe D., Kl einert M. , Zawa dzki C. , Pola tynski A. , Ir ms cher G., Brinker
IEEE Co mm un. Surv . Tutori al s (2016). W., Mo ehrl e M. , Bach H. -G., Keil N., Schell M. , Recent develo pm ents in
[2] Chatterj ee B.C. , Sa rm a N., Ok i E., Ro uting an d sp ectrum al lo cation in
poly mer-ba sed photonic comp onents for disr uptive capa city upgr ad e in
elas tic optical networ ks : a tutori al , IEEE Co mm un. Surv . Tutori al s 17 da ta centers, J. Li ghtwav e Technol. 35 (2016) 683–689.
(2015) 1776–1800. [30] Cl ar k K. A. , Cl ethero e D., Gera rd T., Ha ll er I., Jozwik K. , Shi K. , Thom sen
[3] Ma ro m D.M. , Co lbourn e P.D., D’erri co A. , Fontai ne N.K. , Ikum a Y. , Proietti
B., Will ia ms H. , Zerv as G., Ball an i H. , et al ., Sy nchronous subnan os econd
R. , Zong L. , Ri va s-Mo scos o J.M. , Tomk os I., Surv ey of photonic sw itching clock an d da ta recovery for opticall y sw itched da ta centres using clock
ar chitectures an d technolo gi es in suppor t of sp atia ll y an d sp ectral ly flexible phas e caching, Nature Electronics 3 (2020) 426–433.
optical networ ki ng, IEEE/OS A J. Op tic. Co mm un. Netw. 9 (2017) 1–26. [31] Cheung S. , Su T., Ok am oto K. , Yo o S. J.B., Ultra-comp act si li con photonic
[4] Duki c V. , Khan na G., Gkan tsidis C. , Ka ra gi an nis T., Parm ig ia ni F. , Si ngla
512 × 512 25 ghz ar ra yed wa veguide gr ating ro uter, IEEE J. Sel. Top.
A. , Filer M. , Co x J.L. , Ptas znik A. , Ha rl an d N., Sa unders W., Bela dy C. , Qu an t. Electron. 20 (2014) 310–316, https: //doi. or g/10.1109/JSTQE.2013.
Beyo nd the mega -da ta center: networ ki ng mu lti-da ta center regi ons, 2295879.
SIGCOM M, AC M, 2020, pp. 765–781. [32] Toshiy os hi H. , Fuji ta H. , Electros tatic mi cro tors ion mi rr or s for an optical
[5] Ball an i H. , Co sta P., Behrendt R. , Cl ethero e D., Ha ll er I., Jozwik K. ,
sw itch ma trix , Journa l of Mi croelectro mechan ical sy stems 5 (1996)
Ka ri nou F. , La nge S. , Shi K. , Thom sen B., Will ia ms H. , Si ri us: a flat 231–237.
da tacenter networ k with na nosecond optical sw itching, SIGCOM M, AC M, [33] Tsai J., Wu M. C. , A high port-count wa velength-selectiv e sw itch using a
2020, pp. 782–797. la rg e scan -an gl e, high fill -fa ctor , two-ax is mems scan ner ar ra y, IEEE
[6] Va n Eck N.J., Waltma n L. , Ci tation-ba sed clusteri ng of publications using
Photon. Technol. Lett. 18 (2006) 1439–1441, https: //doi. or g/10.1109/LPT.
Ci tNetExplor er an d VO Sv iewer, Scientom etri cs 111 (2017) 1053–1070. 2006.877235.
[7] Foerster K. -T., Schm id S. , Surv ey of reconfig urab le da ta center networ ks : [34] Fa rr ington N., Porter G., Ra dhak ri shna n S. , Bazzaz H. H. , Subr am an ya V. ,
enab lers , al go ri thms , comp lexi ty, SIGACT News 50 (2019) 62–79, https: // Fa inma n Y. , Papen G., Va hdat A. , Heli os : a hybr id electrical /optical sw itch
doi. or g/10.1145/3351452. 3351464. ar chitecture for mo dula r da ta centers, Proceedings of the AC M SIGCOM M
[8] Celi k A. , Shihad a B., Alouini M. -S. , Op tical wireless da ta center networ ks :
2010 Co nference, 2010, pp. 339–350.
potentia ls , li mi tations, an d pr os pects, Broa dban d Access Co mm unication [35] Baxter G., Fr is ken S. , Abak oumo v D., Zhou H. , Cl ar ke I., Bartos A. , Pool e
Technolo gi es XIII, vo l. 10945, Intern ationa l So ciety for Op tics an d S. , Hi ghly pr og ra mm ab le wa velength selectiv e sw itch ba sed on li quid
Photonics, 2019, p. 109450I. crys tal on si li con sw itching elements, 2006 Op tical Fiber Co mm unication
[9] Lu Y. , Gu H. , Fl exible an d scal ab le optical interconnects for da ta centers:
Co nference an d the National Fiber Op tic Engi neers Co nference, IEEE, 2006,
trends an d chal lenges, IEEE Co mm un. Ma g. 57 (2019) 27–33, https: //doi. p. 3.
or g/10.1109/MC OM .001. 1900326. [36] Wang M. , Zong L. , Ma o L. , Ma rq uez A. , Ye Y. , Zhao H. , Va quero Ca ba ll ero
[10] Will ner A. , Op tical Fiber Telecomm unications, vo l. 11, Acad emic Press,
F. J., Lcos sl m study an d its ap plication in wa velength selectiv e sw itch,
2019. Photonics, vo l. 4, Multidis ciplinar y Digi tal Publis hing Institute, 2017, p.
[11] Butt R. A. , As hraf M. W., Fa heem M. , Idrus S. M. , A surv ey of dyna mi c
22.
ba ndwidth as si gnment schemes for tdm-ba sed pa ss iv e optical networ k, [37] Ya ng H. , Ro bertso n B., Wilk inso n P., Chu D., Sm al l phas e pa ttern 2d beam
J. Op t. Co mm un. 41 (2020) 279–293. steeri ng an d a si ngle lcos desi gn of 40 1 × 12 stacked wa velength selectiv e
[12] Alim i I.A. , Teix eira A. L. , Mo nteiro P.P., Towa rd an efficient c-ra n optical
sw itches, Op t Expr ess 24 (2016) 12240–12253 http://www. opticsexpr ess.
fr onthau l for the future networ ks : a tutori al on technolo gi es, requir ements, or g/ab stra ct.cfm?URI=oe-24-11-12240, https: //doi. or g/10.1364/OE.24.
chal lenges, an d so lutions, IEEE Co mm un. Surv . Tutori al s 20 (2018) 012240.
708–769, https: //doi. or g/10.1109/CO MS T.2017.2773462. [38] Chen H. , Fontai ne N.K. , Ry f R. , Neil so n D.T., Lcos -ba sed photonic
[13] Mcga rr y M. P., Reis sl ein M. , Ma ier M. , Ethern et pa ss iv e optical networ k
cros sconnect, Op tical Fiber Co mm unication Co nference (OFC ) 2019,

17
M. Nance Hall et al. Optical Switching and Networking xxx (xxxx) 100621

Op tical So ciety of Am erica, 2019, p. Th1E.6 http://www. os ap ubli shing. (NSDI 20), 2020, pp. 1–18.
or g/ab stra ct.cfm?URI=OF C-2019-Th1E.6, https: //doi. or g/10.1364/OF C. [64] Ka ss ing S. , Va la da rs k A. , Shahaf G., Schapira M. , Si ngla A. , Beyo nd
2019.Th1E.6. fa ttrees without an tennae, mi rr or s, an d disco-ba ll s, SIGCOM M, 2017.
[39] Zhan g Z., Yo u Z., Chu D., Fundam entals of phas e-only li quid crys tal on [65] Shin J., Si rer E.G., Weathers poon H. , Ki ro vs ki D., On the feas ibil ity of
si li con (lcos) devi ces, Li ght: Science & Applications 3 (2014) e213. comp letely wireless s da tacenters, IEEE/AC M Tran s. Netw. 21 (2013)
[40] Filer M. , Gaudette J., Ghobad i M. , Ma haja n R. , Issenhuth T., Kl inkers B., 1666–1679.
Co x J., Elas tic optical networ ki ng in the mi cros oft cloud, IEEE/OS A J. [66] Wang G., Anders en D.G., Ka mi nsky M. , Ko zuch M. , Ng T.S. E.,
Op tic. Co mm un. Netw. 8 (2016) A45–A54. Papa gi an na ki K. , Glick M. , Mumm ert L. B., Yo ur da ta center is a ro uter: the
[41] Jinno M. , Ko zicki B., Taka ra H. , Watana be A. , So ne Y. , Tana ka T., Hi ra no case for reconfig urab le optical circuit sw itched pa ths, Ho tNets, AC M
A. , Distan ce-ad ap tive sp ectrum reso urce al lo cation in sp ectrum -sl iced SIGCOM M, 2009.
elas tic optical pa th networ k [topics in optical comm unications], IEEE [67] Edmo nds J., Paths, trees an d flowers, Ca n. J. Ma th. 17 (1965) 449–467.
Co mm un. Ma g. 48 (2010) 138–145, https: //doi. or g/10.1109/MC OM .2010. [68] Fa rr ington N., Porter G., Ra dhak ri shna n S. , Bazzaz H. H. , Subr am an ya V. ,
5534599. Fa inma n Y. , Papen G., Va hdat A. , Heli os : a hybr id electrical /optical sw itch
[42] Sa mb o N., Ca stol di P., D’Erri co A. , Ri ccar di E., Paga no A. , Mo reol o M. S. , ar chitecture for mo dula r da ta centers, SIGCOM M, AC M, 2010, pp. 339–350.
Fa br ega J.M. , Ra fique D., Napoli A. , Fr ig erio S. , et al ., Next genera tion [69] Dai W., Foerster K. -T., Fuchss teiner D., Schm id S. , Lo ad -optimi zation in
sl iceable ba ndwidth va ri ab le tran sp onders , IEEE Co mm un. Ma g. 53 (2015) reconfig urab le networ ks : al go ri thms an d comp lexi ty of flow ro uting,
163–171. SIGMETRICS Perf or m. Eval . Rev. 48 (3) (2020) 39–44, https: //doi. or g/10.
[43] Ha tori N., Shim izu T., Ok an o M. , Ishizaka M. , Ya ma mo to T., Urino Y. , 1145/3453953. 3453962.
Mo ri M. , Naka mu ra T., Ar ak aw a Y. , A hybr id integr ated li ght so urce on a [70] Müll er-Ha nneman n M. , Schwar tz A. , Im plementing weig hted b-ma tching
si li con platform using a trident sp ot-si ze converter, J. Li ghtwav e Technol. al go ri thms : insi ghts fr om a comp utationa l study, AC M Journa l of
32 (2014) 1329–1336, https: //doi. or g/10.1109/JLT.2014.2304305. Experi mental Algo ri thmi cs 5 (2000) 8.
[44] Bill ah M. R. , Blai cher M. , Ho os e T., Dietri ch P.-I., Ma ri n-Palo mo P., [71] Si ngla A. , Si ngh A. , Ra ma chan dr an K. , Xu L. , Zhan g Y. , Proteus: a
Li ndenma nn N., Nesi c A. , Ho fm an n A. , Troppenz U., Mo ehrl e M. , Ra ndel S. , topolo gy ma ll eable da ta center networ k, Ho tNets, AC M, 2010.
Fr eude W., Ko os C. , Hy br id integr ation of si li con photonics circuits an d inp [72] Chen K. , Si ngla A. , Si ngh A. , Ra ma chan dr an K. , Xu L. , Zhan g Y. , Wen X. ,
la sers by photonic wire bonding, Op tica 5 (2018) 876–883 http://www. Chen Y. , OS A: an optical sw itching ar chitecture for da ta center networ ks
os ap ubli shing. or g/optica/ab stra ct.cfm?URI=optica5-7-876, https: //doi. with unpr ecedented flexibil ity, IEEE/AC M Tran s. Netw. 22 (2014) 498–511.
or g/10.1364/OP TICA .5. 000876. [73] Bienkows ki M. , Fuchss teiner D., Ma rcinkows ki J., Schm id S. , On li ne
[45] Ma tsum oto T., Kura hashi T., Ko noik e R. , Tanizawa K. , Suzuki K. , Uetake dyna mi c b-ma tching with ap plications to reconfig urab le da tacenter
A. , Taka ba ya shi K. , Ikeda K. , Ka wa shim a H. , Akiy am a S. , Seki guchi S. , In- networ ks , Proc. 38th Intern ationa l Sy mp os ium on Co mp uter Perf or ma nce,
li ne optical am plification for si li con photonics platform by flip-chip bonded Mo deli ng, Meas urements an d Eval ua tion (PERFO RM ANCE), 2020.
inp-so as , 2018 Op tical Fiber Co mm unications Co nference an d Exposi tion [74] Foerster K. -T., Ghobad i M. , Schm id S. , Char acteri zing the al go ri thmi c
(OFC ), 2018, pp. 1–3. comp lexi ty of reconfig urab le da ta center ar chitectures, ANCS , IEEE/AC M,
[46] Will ia ms K. , Li u X. , Ma tters-Ka mm erer M. , Meig han A. , Spiegelberg M. , 2018.
va n der Tol J., Traj kovi c M. , Wale M. , Ya o W., Zhan g X. , Indium phos phide [75] Foerster K. -T., Pacut M. , Schm id S. , On the comp lexi ty of non-segr egated
photonic circuits on si li con electronics, Op tical Fiber Co mm unication ro uting in reconfig urab le da ta center ar chitectures, SIGCOM M Co mp ut.
Co nference (OFC ) 2020, Op tical So ciety of Am erica, 2020, p. M3A. 1 http:// Co mm un. Rev. 49 (2) (2019) 2–8.
www. os ap ubli shing. or g/ab stra ct.cfm?URI=OF C-2020-M3A. 1, https: //doi. [76] Fenz T., Foerster K. -T., Schm id S. , Vill edieu A. , Efficient non-segr egated
or g/10.1364/OF C. 2020.M3A. 1. ro uting for reconfig urab le dema nd-aw ar e networ ks , Co mp ut. Co mm un.
[47] Thom so n D., Zilk ie A. , Bowers J.E., Ko ml jenovi c T., Reed G.T., Vivi en L. , 164 (2020) 138–147.
Ma rr is -Mo ri ni D., Ca ss an E., Viro t L. , F´ed´eli J.-M. , et al ., Ro ad ma p on [77] Teh M. Y. , Hu ng Y. -H. , Mi chelog ia nnak is G., Ya n S. , Glick M. , Shal f J.,
si li con photonics, J. Op t. 18 (2016) 073003. Berg ma n K. , Tago : rethinki ng ro uting desi gn in high perf or ma nce
[48] Ha mz a A. S. , Deog un J.S. , Alexan der D.R. , Cl as si fication fr am ewor k for reconfig urab le networ ks , Proceedings of the Intern ationa l Co nference for
fr ee sp ace optical comm unication li nks an d sy stems, IEEE Co mm un. Surv . Hi gh Perf or ma nce Co mp uting, Networ ki ng, Stor ag e an d Anal ys is , SC ’20,
Tutori al s 21 (2019) 1346–1382. IEEE Press, 2020.
[49] Azim i N.H. , Qa zi Z.A. , Gupta H. , Seka r V. , Das S. R. , Lo ngtin J.P., Shah H. , [78] Va li an t L. G., A scheme for fa st pa ra ll el comm unication, SIAM J. Co mp ut.
Tanwer A. , Firefly: a reconfig urab le wireless da ta center fa br ic using fr ee- 11 (1982) 350–361.
sp ace optics, SIGCOM M, AC M, 2014, pp. 319–330. [79] Chan g C. , Lee D., Jou Y. , Lo ad ba la nced birk hoff-vo n neum an n sw itches,
[50] Ghobad i M. , Ma haja n R. , Phan is hayee A. , Deva nur N.R. , Kulk ar ni J., pa rt I: one-stag e buffering, Co mp ut. Co mm un. 25 (2002) 611–622.
Ra na de G., Blan che P., Ra stegar fa r H. , Glick M. , Ki lper D.C. , Proj ector: [80] Si ngla A. , Ho ng C. , Popa L. , Godfrey P.B., Jell yfis h: networ ki ng da ta
ag il e reconfig urab le da ta center interconnect, SIGCOM M, AC M, 2016. centers ra ndom ly , NSDI, 2012 https: //www. usenix .or g/conference/nsdi12/
[51] Ha mz a A. S. , Ya da v S. , Ketan S. , Deog un J.S. , Alexan der D.R. , Ow cell : technical-sess ions/pr esentation/si ngla .
optical wireless cell ular da ta center networ k ar chitecture, IC C, IEEE, 2017, [81] Xi a Y. , Sun X. S. , Dzinam ar ir a S. , Wu D., Hu an g X. S. , Eugene Ng T.S. ,
pp. 1–6. A tale of two topolo gi es: explor ing convertible da ta center networ k
[52] Cu i Y. , Xi ao S. , Wang X. , Ya ng Z., Ya n S. , Zhu C. , Li X. , Ge N., Diam ond, ar chitectures with flat-tree, SIGCOM M, AC M, 2017.
Nesting the da ta center networ k with wireless ri ngs in 3-d sp ace, IEEE/AC M [82] Porter G., Stro ng R. D., Fa rr ington N., Forencich A. , Sun P., Ro si ng T.,
Tran s. Netw. 26 (2018) 145–160. Fa inma n Y. , Papen G., Va hdat A. , Integr ating mi cros econd circuit sw itching
[53] Bloom S. , Ko reva ar E., Schuster J., Will ebra nd H. , Unders tanding the into the da ta center, SIGCOM M, AC M, 2013, pp. 447–458.
perf or ma nce of fr ee-sp ace optics, J. Op t. Netw. 2 (2003) 178–200. [83] Goel A. , Ka pr al ov M. , Khan na S. , Perf ect ma tchings in o(nlo g n) time in
[54] Trichili A. , Co x M. A. , Oo i B.S. , Alouini M. -S. , Ro ad ma p to fr ee sp ace regula r bipa rtite gr ap hs, SIAM J. Co mp ut. 42 (2013) 1392–1404.
optics, J. Op t. So c. Am . B 37 (2020) A184–A201. [84] Li u H. , Mukerj ee M. K. , Li C. , Feltma n N., Papen G., Sa va ge S. , Seshan S. ,
[55] Terzi C. , Ko rp eogl u I., 60 ghz wireless da ta center networ ks : a surv ey, Voelker G.M. , Anders en D.G., Ka mi nsky M. , Porter G., Snoeren A. C. ,
Co mp ut. Networ k. 185 (2021) 107730. Scheduli ng techniques for hybr id circuit/pa cket networ ks , Co NEXT, AC M,
[56] Ha mz a A. S. , Deog un J.S. , Alexan der D.R. , Wireless comm unication in da ta 2015, p. 41:1-41:13.
centers: a surv ey, IEEE Co mm un. Surv . Tutori al s 18 (2016) 1572–1595. [85] Venkatak ri shna n S. B., Alizad eh M. , Visw an ath P., Co stly circuits,
[57] Fa rr ington N., Porter G., Fa inma n Y. , Papen G., Va hdat A. , Hu nting mi ce subm odular schedules an d ap pr ox im ate cara th’eodor y theorems , Qu eueing
with mi cros econd circuit sw itches, Ho tNets, AC M, 2012, pp. 115–120. Sy st. 88 (2018) 311–347.
[58] Porter G., Stro ng R. , Fa rr ington N., Forencich A. , Chen-Sun P., Ro si ng T., [86] Gupta H. , Cu rr an M. , Zhan C. , Near -optima l mu ltihop scheduli ng in genera l
Fa inma n Y. , Papen G., Va hdat A. , Integr ating mi cros econd circuit sw itching circuit-sw itched networ ks , Co NEXT, AC M, 2020, pp. 31–45.
into the da ta center, SIGCOM M’13, 2013, pp. 447–458. [87] Schwar tz R. , Si ngh M. , Ya zdan bod S. , On li ne an d Of fline Greedy
[59] Fa rr ington N., Forencich A. , Porter G., Sun P., Ford J.E., Fa inma n Y. , Papen Algo ri thms for Ro uting with Sw itching Co sts, 2019 ar Xi v pr epri nt ar Xi v:
G.C. , Va hdat A. , A mu ltipor t mi cros econd optical circuit sw itch for da ta 1905.02800.
center networ ki ng, IEEE Photon. Technol. Lett. 25 (2013) 1589–1592, [88] Avin C. , Mo ndal K. , Schm id S. , Dema nd-aw ar e networ k desi gns of bounded
https: //doi. or g/10.1109/LPT.2013.2270462. degr ee, DISC , 2017.
[60] Mukerj ee M. K. , Ca nel C. , Wang W., Ki m D., Seshan S. , Snoeren A. C. , [89] Avin C. , Mo ndal K. , Schm id S. , Dema nd-aw ar e networ k desi gn with
Adap ting tcp for reconfig urab le da tacenter networ ks , 17th USENIX mi nima l congestion an d ro ute lengths, Proc. IEEE INFO CO M, 2019.
Sy mp os ium on Networ ked Sy stems Desi gn an d Im plementation (NSDI 20), [90] Avin C. , Schm id S. , Towa rd dema nd-aw ar e networ ki ng: a theory for self-
2020, pp. 651–666. ad justing networ ks , AC M SIGCOM M Co mp uter Co mm unication Revi ew
[61] Alis tarh D., Ball an i H. , Co sta P., Funnell A. , Benjam in J., Watts P.M. , (CCR ), 2018.
Thom sen B., A high-ra dix, lo w-la tency optical sw itch for da ta centers, [91] Avin C. , Hercules A. , Lo ukas A. , Schm id S. , rd an : towa rd ro bust dema nd-
Co mp uter Co mm unication Revi ew 45 (2015) 367–368. aw ar e networ k desi gns, Inform ation Processi ng Letters (IPL), 2018.
[62] Mell ette W.M. , Schuster G.M. , Porter G., Papen G., Ford J.E., A scal ab le, [92] Anan d S. , Garg N., Kuma r A. , Reso urce au gm entation for weig hted flow-
pa rtia ll y config urab le optical sw itch for da ta center networ ks , J. Li ghtwav e time explai ned by dual fitting, Proceedings of the Twenty-Thir d Annual
Technol. 35 (2017) 136–144, https: //doi. or g/10.1109/JLT.2016.2636025. AC M-SIAM Sy mp os ium on Discrete Algo ri thms , SIAM, 2012, pp.
[63] Mell ette W.M. , Das R. , Guo Y. , McGuinness R. , Snoeren A. C. , Porter G., 1228–1241.
Expa nding acro ss time to deli ver ba ndwidth efficiency an d lo w la tency, 17th [93] Dinitz M. , Mo seley B., Scheduli ng for Weig hted Fl ow an d Co mp letion Times
USENIX Sy mp os ium on Networ ked Sy stems Desi gn an d Im plementation in Reconfig urab le Networ ks , 2020 ar Xi v pr epri nt ar Xi v: 2001.07784.

18
M. Nance Hall et al. Optical Switching and Networking xxx (xxxx) 100621

[94] Kulk ar ni J., Schm id S. , Schm idt P., Scheduli ng Op portunis tic Li nks in Two- os ap ubli shing. or g/ab stra ct.cfm?URI=OF C-2013-OW 1H .2, https: //doi. or g/
Tiered Reconfig urab le Datacenters, 2020 ar Xi v pr epri nt ar Xi v: 2010.07920. 10.1364/OF C. 2013.OW 1H .2.
[95] Schm id S. , Avin C. , Scheideler C. , Boro khov ich M. , Ha eupler B., Lo tker Z., [122] Mo ura U., Garr ich M. , Ca rv al ho H. , Sv ol ensk i M. , Andr ad e A. , Cesa r A. C. ,
Spla ynet: towa rd s lo call y self-ad justing networ ks , IEEE/AC M Tran s. Netw. Ol iv eira J., Co nfor ti E., Co gnitiv e methodol og y for optical am plifier ga in
24 (2016) 1421–1433. ad justment in dyna mi c dwdm networ ks , J. Li ghtwav e Technol. 34 (2016)
[96] Peres B., Goussevs ka ia O. , Schm id S. , Avin C. , Co ncurrent self-ad justing 1971–1979.
distri buted tree networ ks , Proc. Intern ationa l Sy mp os ium on Distri buted [123] Mo ura U., Garr ich M. , Cesa r A. C. , Ol iv eira J., Co nfor ti E., Execution time
Co mp uting (DISC), 2017. im pr ov ement for optical am plifier cognitiv e methodol og y in dyna mi c wdm
[97] Avin C. , Ha eupler B., Lo tker Z., Scheideler C. , Schm id S. , Lo call y networ ks , 2017 SBMO /IEEE MTT-S Intern ationa l Mi crowav e an d
selfad justing tree networ ks , Proc. 27th IEEE Intern ationa l Para ll el an d Op toelectronics Co nference (IMO C), IEEE, 2017, pp. 1–5.
Distri buted Processi ng Sy mp os ium (IPDPS), 2013. [124] Ca se J., Fedor M. , Schoffstal l M. L. , Davi n J., Rfc1157: Si mp le Networ k
[98] Peres B., de Ol iv eira So uza O. A. , Goussevs ka ia O. , Avin C. , Schm id S. , Ma na gement Protocol (Snm p), 1990.
Distri buted self-ad justing tree networ ks , INFO CO M, IEEE, 2019. [125] Enns R. , Bjor kl und M. , Schoenwa elder J., Bierma n A. , Networ k
[99] Avin C. , Schm id S. , Renets: statical ly -optima l dema nd-aw ar e networ ks , Co nfig uration Protocol (Netconf), 2011.
Proc. SIAM Sy mp os ium on Algo ri thmi c Principles of Co mp uter Sy stems [126] Shai kh A. , Ho fm eister T., Dangui V. , Vusi ri ka la V. , Vendor -neutra l networ k
(APO CS ), 2021. repr esentations for tran sp or t sdn, 2016 Op tical Fiber Co mm unications
[100] Avin C. , Mo ndal K. , Schm id S. , Dyna mi call y optima l self-ad justing si ngle- Co nference an d Exhibition (OFC ), 2016, pp. 1–3.
so urce tree networ ks , Proc. La tin Am erican Theoretical Inform atics [127] Lehm en A. V. , Dovers pike R. , Cl ap p G., Fr eimu th D.M. , Gannett J., Ki m K. ,
Sy mp os ium (LATIN), 2020. Ko br insk i H. , Ma vr og iorg is E., Pastor J., Ra uch M. , Ra ma kr is hnan K. K. ,
[101] Avin C. , Sa lem I., Schm id S. , Work ing set theorems for ro uting in Sk oog R. , Wils on B., Woodwa rd S. L. , Co ro net: testbeds, cloud comp uting,
selfad justing sk ip li st networ ks , Proc. IEEE INFO CO M, 2020. an d less ons lear ned, Op tical Fiber Co mm unication Co nference, Op tical
[102] Sa lm an S. , Streiffer C. , Chen H. , Benson T., Ka da v A. , Deepconf: So ciety of Am erica, 2014, p. W4B.1 http://www. os ap ubli shing. or g/
au toma ting da ta center networ k topolo gi es ma na gement with ma chine ab stra ct.cfm?URI=OF C-2014-W4B.1, https: //doi. or g/10.1364/OF C. 2014.
lear ning, Proceedings of the 2018 Work shop on Networ k Meets AI & ML , W4B.1.
NetAI’18, AC M, New Yo rk , NY, USA, 2018, pp. 8–14. [128] Sk oog R. A. , Neidhard t A. L. , A fa st, ro bust si gnal ing pr otocol for enab li ng
[103] Ka lm ba ch P., Zerw as J., Baba rczi P., Blenk A. , Kell erer W., Schm id S. , highly dyna mi c optical networ ks , 2009 Co nference on Op tical Fiber
Empoweri ng self-dr iv ing networ ks , Proc. AC M SIGCOM M 2018 Work shop Co mm unication-Incudes Post Dead li ne Papers , IEEE, 2009, pp. 1–3.
on Self-Driv ing Networ ks (SDN),, 2018. [129] Zang H. , Jue J.P., Mukherjee B., et al ., A revi ew of ro uting an d wa velength
[104] Truong-Hu u T., Mo han P.M. , Gurusa my M. , Virtua l networ k embedding in as si gnment ap pr oa ches for wa velength-ro uted optical wdm networ ks , Op t.
ri ng optical da ta centers using Ma rk ov chai n pr obab il ity mo del, IEEE Networ k Ma g. 1 (2000) 47–60.
Tran sa ctions on Networ k an d Serv ice Ma na gement 16 (2019) 1724–1738. [130] Dura ir aj an R. , Barf or d P., So mm ers J., Will inger W., Intertubes: a study of
[105] Chen L. , Chen K. , Zhu Z., Yu M. , Porter G., Qi ao C. , Zhong S. , Enab li ng the US lo ng-haul fiber-optic infr as tructure, SIGCOM M, AC M, 2015, pp.
wide-sp read comm unications on optical fa br ic with mega sw itch, 14th 565–578.
USENIX Sy mp os ium on Networ ked Sy stems Desi gn an d Im plementation [131] Ho ng C. -Y. , Ma ndal S. , Al-Fa res M. , Zhu M. , Alim i R. , Bhag at K. N.B. C. ,
(NSDI 17), USENIX As so ciation, 2017, pp. 577–593 Boston, MA https: // Jain S. , Ka im al J., Li an g S. , Mendelev K. , Padgett S. , Ra be F. , Ra y S. ,
www. usenix .or g/conference/nsdi17/technical-sess ions/pr esentation/chen. Tewa ri M. , Tierney M. , Zahn M. , Zoll a J., On g J., Va hdat A. , B4 an d af ter:
[106] Mi sr a J., Gries D., A constructive pr oof of vi zing’s theorem, Inf. Process. ma na gi ng hierar chy, pa rtitioning, an d as ym metry for av ai la bili ty an d
Lett. 41 (1992) 131–133. scal e in go og le’s so ftwa re-defined wa n, Proceedings of the 2018 Co nference
[107] Chen K. , Wen X. , Ma X. , Chen Y. , Xi a Y. , Hu C. , Dong Q. , Li u Y. , Towa rd a of the AC M Specia l Interest Group on Data Co mm unication, SIGCOM M ’18,
scal ab le, fa ult-tolera nt, high-perf or ma nce optical da ta center ar chitecture, As so ciation for Co mp uting Ma chinery, 2018, pp. 74–87 New Yo rk , NY,
IEEE/AC M Tran s. Netw. 25 (2017) 2281–2294. USA.
[108] Wang H. , Xi a Y. , Berg ma n K. , Ng T.S. E., Sa hu S. , Sr ipan idkulchai K. , [132] Tsir il ak is I., Ma s C. , Tomk os I., Co st comp ar is on of ip/wdm vs . ip/otn for
Rethinki ng the phys ical la yer of da ta center networ ks of the next decade: euro pean ba ckbone networ ks , Proceedings of 2005 7th Intern ationa l
using optics to enab le efficient ∗-cast connectivi ty, Co mp uter Co nference Tran sp ar ent Op tical Networ ks , vo l. 2, IEEE, 2005, pp. 46–49
Co mm unication Revi ew 43 (2013) 52–58. 2005.
[109] Hu an g X. S. , Sun X. S. , Ng T.S. E., Sunflo w: efficient optical circuit [133] Gers tel O. , Fils fils C. , Telk am p T., Gunkel M. , Ho rn effer M. , Lo pez V. ,
scheduli ng for coflows, Co NEXT, AC M, 2016, pp. 297–311. Ma yo ra l A. , Multi-la yer capa city plan ning for ip-optical networ ks , IEEE
[110] Wang H. , Yu X. , Xu H. , Fa n J., Qi ao C. , Hu an g L. , Integr ating coflow an d Co mm un. Ma g. 52 (2014) 44–51.
circuit scheduli ng for optical networ ks , IEEE Tran s. Para ll el Distr. Sy st. [134] Gunkel M. , Autenr ieth A. , Neugir g M. , Elbers J.-P., Adva nced mu ltil ay er
(2019), https: //doi. or g/10.1109/TPDS.2018.2889251. resi li ence scheme with optical restor ation for ip-ov er-dwdm core networ ks ,
[111] Bao J., Dong D., Zhao B., Luo Z., Wu C. , Gong Z., Fl ycas t: fr ee-sp ace optics 2012 IV Intern ationa l Co ngress on Ultra Mo dern Telecomm unications an d
accelera ting mu lticas t comm unications in phys ical la yer, Co mp uter Co ntro l Sy stems, IEEE, 2012, pp. 657–662.
Co mm unication Revi ew 45 (2015) 97–98. [135] Papa nikola ou P., Chri stodoulo poulos K. , Va rv ar ig os E., Joint mu ltil ay er
[112] Sun X. S. , Ng T.S. E., When creek meets ri ver: exploiting high-ba ndwidth plan ning of surv iv ab le elas tic optical networ ks , Op tical Fiber
circuit sw itch in scheduli ng mu lticas t da ta, IC NP, IEEE Co mp uter So ciety, Co mm unication Co nference, Op tical So ciety of Am erica, 2016 M2K–3.
2017, pp. 1–6. [136] Ko zdro ws ki S. , Zotkiewicz M. , Sujecki S. , Op timi zation of optical net-˙
[113] Sun X. S. , Xi a Y. , Dzinam ar ir a S. , Hu an g X. S. , Wu D., Ng T.S. E., Republic: work s ba sed on cdc-ro ad m technolo gy , Appl. Sci. 9 (2019) 399.
da ta mu lticas t meets hybr id ra ck-level interconnections in da ta center, [137] Woodwa rd S. L. , Feuer M. D., Ki m I., Pala char la P., Wang X. , Bihon D.,
IC NP, IEEE Co mp uter So ciety, 2018, pp. 77–87. Serv ice velo city: ra pid pr ov is ioning stra tegi es in optical ro ad m networ ks ,
[114] Xi a Y. , Ng T.S. E., Sun X. S. , Blas t: accelera ting high-perf or ma nce da ta J. Op t. Co mm un. Netw. 4 (2012) 92–98.
an al ytics ap plications by optical mu lticas t, INFO CO M, IEEE, 2015, pp. [138] Chen X. , Zhu Z., Proietti R. , Yo o S. J.B., On incentiv e-dr iv en vnf serv ice
1930–1938. chai ning in inter-da tacenter elas tic optical networ ks : a hierar chical ga me-
[115] Luo L. , Foerster K. , Schm id S. , Yu H. , Spli tcas t: optimi zing mu lticas t flows theoretic mechan is m, IEEE Tran sa ctions on Networ k an d Serv ice
in reconfig urab le da tacenter networ ks , INFO CO M, IEEE, 2020, pp. Ma na gement 16 (2019) 1–12.
2559–2568. [139] Zhan g X. J., Ki m S. , Lumetta S. S. , Op portunity cost an al ys is for dyna mi c
[116] Ya ng M. , Ra stegar fa r H. , Djor djevic I.B., Phys ical -la yer ad ap tive reso urce wa velength ro uted mesh networ ks , IEEE/AC M Tran s. Netw. 19 (2011)
al lo cation in so ftwa re-defined da ta center networ ks , IEEE/OS A Journa l of 747–759.
Op tical Co mm unications an d Networ ki ng 10 (2018) 1015–1026. [140] Thya ga turu A. S. , Mercia n A. , McGarr y M. P., Reis sl ein M. , Kell erer W.,
[117] Teh M. Y. , Zhao S. , Ca o P., Berg ma n K. , CO UDER: Ro bust Topolo gy So ftwa re defined optical networ ks (sdons): a comp rehensiv e surv ey, IEEE
Engi neering for Op tical Ci rcuit Sw itched Data Center Networ ks , 2020 Co RR Co mm un. Surv . Tutori al s 18 (2016) 2738–2786.
ab s/2010.00090. [141] Ca sell as R. , Ma rt´ınez R. , Vila lta R. , noz R. M. , Co ntro l, ma na gement, an d
[118] Teh M. Y. , Zhao S. , Berg ma n K. , METTEOR: Ro bust Multi-Traf fic Topolo gy or chestration of optical networ ks : evol ution, trends, an d chal lenges,
Engi neering for Co mm ercial Data Center Networ ks , 2020 Co RR ab s/2002. J. Li ghtwav e Technol. 36 (2018) 1390–1402 http://jl t.os a. or g/ab stra ct.
00473. cfm?URI=jl t-36-7-1390.
[119] Gringeri S. , Bitar N., Xi a T.J., Extending so ftwa re defined networ k [142] Ko dial am M. , La ks hm an T.V. , Integr ated dyna mi c ip an d wa velength
pr inciples to include optical tran sp or t, IEEE Co mm un. Ma g. 51 (2013) ro uting in ip ov er wdm networ ks , IEEE INFO CO M 1 (2001) 358–366.
32–40, https: //doi. or g/10.1109/MC OM .2013.6476863. [143] Brzezinski A. , Mo dian o E., Dyna mi c reconfig uration an d ro uting
[120] Chri stodoulo poulos K. , Ko kk inos P., Va rv ar ig os E.M. , Indirect an d direct al go ri thms for ip-ov er-wdm networ ks with stochastic traf fic, J. Li ghtwav e
mu lticos t al go ri thms for onli ne im pa ir ment-aw ar e rw a, IEEE/AC M Tran s. Technol. 23 (2005) 3188.
Netw. 19 (2011) 1759–1772. [144] Zhong Z., Hu a N., Torn ator e M. , Li J., Li Y. , Zheng X. , Mukherjee B.,
[121] Ol iv eira J., Ca ba ll ero A. , aes E.M. , Mo ura U., Bork owsk i R. , Cu ri el G., Prov is ioning shor t-term traf fic fluctuations in elas tic optical networ ks ,
Hi ra ta A. , Hecker L. , Porto E., Ziba r D., ao J.M. , Mo nr oy I.T., Ol iv eira J., IEEE/AC M Tran s. Netw. 27 (2019) 1460–1473.
Demo nstration of edfa cognitiv e ga in contro l vi a gm pls for mi xed [145] Patri S. K. , Autenr ieth A. , Ra fique D., Elbers J.-P., Ma chuca C. M. , Hecson:
mo dula tion form ats in hetero geneous optical networ ks , Op tical Fiber heuris tic for config uration selection in optical networ k plan ning, Op tical
Co mm unication Co nference/National Fiber Op tic Engi neers Co nference Fiber Co mm unication Co nference (OFC ) 2020, Op tical So ciety of Am erica,
2013, Op tical So ciety of Am erica, 2013, p. OW 1H .2 http://www. 2020, p. Th2A.32 http://www. os ap ubli shing. or g/ab stra ct.cfm?URI=OF C-

19
M. Nance Hall et al. Optical Switching and Networking xxx (xxxx) 100621

2020-Th2A.32, https: //doi. or g/10.1364/OF C. 2020.Th2A.32. [167] Benzao ui N., Gonzal ez M. S. , Estar´an J.M. , Ma rd oy an H. , La utenschlaeger
[146] Xu D., Li G., Ra ma mu rthy B., Chiu A. L. , Wang D., Dovers pike R. D., On W., Gebhar d U., Demb eck L. , Bigo S. , Pointuri er Y. , Determ inis tic dyna mi c
pr ov is ioning divers e circuits in hetero geneous mu lti-la yer optical networ ks , networ ks (ddn), J. Li ghtwav e Technol. 37 (2019) 3465–3474.
Co mp ut. Co mm un. 36 (2013) 689–697. [168] Ho ng C. -Y. , Ka ndula S. , Ma haja n R. , Zhan g M. , Gill V. , Nanduri M. ,
[147] Ma s C. , Tomk os I., Tonguz O. K. , Fa il ure lo cation al go ri thm for tran sp ar ent Wattenhofer R. , Achieving high util ization with so ftwa re-dr iv en wa n,
optical networ ks , IEEE J. Sel. Ar ea. Co mm un. 23 (2005) 1508–1519. Co mp ut. Co mm un. Rev. 43 (2013) 15–26.
[148] Chen J., Wosi nska L. , Ma chuca C. M. , Jaeger M. , Co st vs . reli ab il ity [169] Figueira S. , Naik sa tam S. , Co hen H. , Cu trell D., Dasp it P., Gutierrez D.,
perf or ma nce study of fiber access networ k ar chitectures, IEEE Co mm un. Ho an g D.B., La vi an T., Ma mb retti J., Merr il l S. , et al ., Dwdm -ra m:
Ma g. 48 (2010) 56–65. enab li ng gr id serv ices with dyna mi c optical networ ks , IEEE Intern ationa l
[149] Luo L. , Yu H. , Foerster K. -T., Noor mo hamm ad pour M. , Schm id S. , Sy mp os ium on Cl uster Co mp uting an d the Grid, 2004. CC Grid 2004, IEEE,
Interd atacenter bulk tran sf ers: trends an d chal lenges, IEEE Netw 34 (5) 2004, pp. 707–714.
(2020) 240–246. [170] Bern ier E., Vukovi c M. , Goodwill D., Dasp it P., Wang G., Om ninet: a
[150] Jin X. , Li Y. , Wei D., Li S. , Gao J., Xu L. , Li G., Xu W., Rexf or d J., metropol itan 10 gb/s dwdm photonic sw itched networ k tria l, Op tical Fiber
Op timi zing bulk tran sf ers with so ftwa re-defined optical WAN, SIGCOM M, Co mm unication Co nference, Op tical So ciety of Am erica, 2004, p. WH4.
2016. [171] Iova nna P., Sa bell a R. , Settembr e M. , A traf fic engi neering sy stem for
[151] Jin X. , Li u H. H. , Gandhi R. , Ka ndula S. , Ma haja n R. , Zhan g M. , Rexf or d J., mu ltil ay er networ ks ba sed on the gm pls pa ra digm , IEEE Networ k 17 (2003)
Wattenhofer R. , Dyna mi c scheduli ng of networ k upda tes, SIGCOM M, AC M, 28–37.
2014, pp. 539–550. [172] Cham an ia M. , Ca ri a M. , Juka n A. , Achieving ip ro uting stab il ity with
[152] Luo L. , Foerster K. , Schm id S. , Yu H. , Dead li ne-aw ar e mu lticas t tran sf ers in optical bypa ss , Op t. Sw itch. Netw. 7 (2010) 173–184 http://www.
so ftwa re-defined optical wide-ar ea networ ks , IEEE J. Sel. Ar ea. Co mm un. sciencedir ect.com/science/ar ticle/pii/S1573427710000305, https: //doi.
38 (7) (2020) 1584–1599. or g/10.1016/j. os n.2010.05.005.
[153] Jia S. , Jin X. , Ghas emiesf eh G., Ding J., Gao J., Co mp etitiv e an al ys is for [173] Turk cu O. , Subr am an ia m S. , Perf or ma nce of optical networ ks with li mi ted
onli ne scheduli ng in so ftwa re-defined optical wa n, INFO CO M, 2017. reconfig urab il ity, IEEE/AC M Tran s. Netw. 17 (2009) 2002–2013, https: //
[154] Dinitz M. , Mo seley B., Scheduli ng for weig hted flow an d comp letion times doi. or g/10.1109/TNET.2009.2014158.
in reconfig urab le networ ks , INFO CO M, 2020. [174] Jin X. , Li Y. , Wei D., Li S. , Gao J., Xu L. , Li G., Xu W., Rexf or d J.,
[155] Si ngh R. , Ghobad i M. , Foerster K. , Filer M. , Gill P., Run, wa lk , craw l: Op timi zing bulk tran sf ers with so ftwa re-defined optical wa n, Proceedings
towa rd s dyna mi c li nk capa cities, Ho tNets, AC M, 2017. of the 2016 AC M SIGCOM M Co nference, AC M, 2016, pp. 87–100.
[156] Si ngh R. , Ghobad i M. , Foerster K. -T., Filer M. , Gill P., Ra dwan : ra te [175] Goss els J., Choudhury G., Rexf or d J., Ro bust networ k desi gn for ip/optical
ad ap tive wide ar ea networ k, Proceedings of the 2018 Co nference of the AC M ba ckbones, IEEE/OS A Journa l of Op tical Co mm unications an d Networ ki ng
Specia l Interest Group on Data Co mm unication, AC M, 2018, pp. 547–560. 11 (2019) 478–490.
[157] Foerster K. , Schm id S. , Viss icchio S. , Surv ey of consis tent so ftwa re-defined [176] Ma ni S. K. , Nance Ha ll M. , Dura ir aj an R. , Barf or d P., Char acteri stics of
networ k upda tes, IEEE Co mm unications Surv eys an d Tutori al s 21 (2019) metro fiber deploy ments in the us, Proceedings of the Networ k Traf fic
1435–1461. Meas urement an d Anal ys is Co nference, 2020.
[158] Tseng S. , Pers everan ce-aw ar e traf fic engi neering in ra te-ad ap tive networ ks [177] Li H. , Zhan g H. -y. , Wang L. , Li Y. -b. , La i J.-s. , Tang R. , Zhao W.-y. , Wu B.-
with reconfig uration dela y, IC NP, IEEE, 2019, pp. 1–10. b. , Wang D., Zhao X. , et al ., Field tria l of networ k surv iv ab il ity ba sed on
[159] Foerster K. -T., Luo L. , Ghobad i M. , Op tflo w: a flow-ba sed ab stra ction for otn an d ro ad m hybr id networ ki ng, As ia Co mm unications an d Photonics
pr og ra mm ab le topolo gi es, SO SR , AC M, 2020. Co nference, Op tical So ciety of Am erica, 2017 M3C–2.
[160] Chla mt ac I., Ganz A. , Ka rm i G., Li ghtpath comm unications: an ap pr oa ch to [178] Od a S. , Mi ya be M. , Yo shida S. , Ka tagi ri T., Aoki Y. , Ra sm ussen J.C. , Birk
high ba ndwidth optical wa n’s, IEEE Tran s. Co mm un. 40 (1992) 1171–1182, M. , Tse K. , A lear ning li vi ng networ k for open ro ad m networ ks , ECOC
https: //doi. or g/10.1109/26.153361. 2016; 42nd Euro pean Co nference on Op tical Co mm unication, VDE, 2016,
[161] Guo J., Zhu Z., When deep lear ning meets inter-da tacenter optical networ k pp. 1–3.
ma na gement: ad va ntag es an d vulnerab il ities, J. Li ghtwav e Technol. 36 [179] Berd e P., Gero la M. , Ha rt J., Hi guchi Y. , Ko ba ya shi M. , Ko ide T., La ntz B.,
(2018) 4761–4773 http://jl t.os a. or g/ab stra ct.cfm?URI=jl t-36-20-4761, O’ Co nnor B., Ra dosl av ov P., Snow W., et al ., On os : towa rd s an open,
https: //doi. or g/10.1364/JLT.36.004761. distri buted sdn os , Proceedings of the Thir d Work shop on Ho t Topics in
[162] Wang Y. , McNulty Z., Nguyen H. , Networ k vi rtua li zation in sp ectrum sl iced So ftwa re Defined Networ ki ng, 2014, pp. 1–6.
elas tic optical pa th networ ks , J. Li ghtwav e Technol. 35 (2017) 1962–1970. [180] Nance Ha ll M. , Li u G., Dura ir aj an R. , Seka r V. , Fighting fire with li ght:
[163] So to P., Ma ya P., Botero J.F. , Reso urce al lo cation ov er eon-ba sed tackli ng extreme tera bit ddos using pr og ra mm ab le optics, Proceedings of
infr as tructures in a networ k vi rtua li zation envi ro nm ent, IEEE Tran sa ctions the Work shop on Secure Prog ra mm ab le Networ k Infr as tructure, SPIN ’20,
on Networ k an d Serv ice Ma na gement 16 (2019) 13–26. As so ciation for Co mp uting Ma chinery, 2020, pp. 42–48 New Yo rk , NY,
[164] Oz da gl ar A. E., Bertseka s D.P., Ro uting an d wa velength as si gnment in USA, https: //doi. or g/10.1145/3405669. 3405824.
optical networ ks , IEEE/AC M Tran s. Netw. 11 (2003) 259–272. [181] Avin C. , Ghobad i M. , Griner C. , Schm id S. , On the comp lexi ty of traf fic
[165] Ko kk inos P., So um plis P., Va rv ar ig os E.A. , Pattern-dr iv en reso urce traces an d im plications, Proc. AC M Meas . Anal . Co mp ut. Sy st. 4 (2020) 20:
al lo cation in optical networ ks , IEEE Tran sa ctions on Networ k an d Serv ice 1-20:29.
Ma na gement 16 (2019) 489–504. [182] Anal ytics, Cl ar iv ate Web of science www. webofs cience.com

20